SESSION 33 - Intelligent Neural Interfaces and Sensing Systems.pdf

1、ISSCC 2024SESSION 33Intelligent Neural Interfaces and Sensing Systems33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Co

2、nference1 of 57A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel SelectionJiahao Liu1,Xiao Liu1,Xu Wang1,Ziyi Xie1,Zirui Zhong1,Jiajing Fan1,Hui Qiu1,Yiming Xu1,Huajing Qin1,Yu Long

3、1,Yuhong Zhou2,Zixuan Shen3,Liang Zhou1,Liang Chang1,Shanshan Liu1,Shuisheng Lin1,Chao Wang3*,Jun Zhou1*1University of Electronic Science and Technology of China 2West China Hospital of Sichuan University3Huazhong University of Science and Technology33.1:A High Accuracy and Energy-Efficient Zero-Sho

4、t-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference2 of 57Outline Background and challenges Zero-Shot-Retraining Seizure Detection Processor Hybrid-feature-driven

5、 adaptive processing architecture with on-chip learningLearning-based Adaptive channel selection technique Measurement results Conclusions33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adapt

6、ive Channel Selection 2024 IEEE International Solid-State Circuits Conference3 of 57Outline Background and challenges Zero-Shot-Retraining Seizure Detection Processor Hybrid-feature-driven adaptive processing architecture with on-chip learningLearning-based Adaptive channel selection technique Measu

7、rement results Conclusions33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference4 of 57Background IntroductionFp1F3

8、F7T3T5O1P3C3O2P4CzPzFzF4Fp2C4F8T4T6F7-Fp1T5-T3T4-T6T6-O2ICSurface/Intracranial EEG ElectrodesSeizure Detection ProcessorSeizure OnsetNormalAlert OR Close-Loop Stimulation forSeizure Suppression High Detection Accuracy Low Power Consumption Low Detection LatencyRequirements:33.1:A High Accuracy and E

9、nergy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference5 of 57 Existing designs achieve high accuracy(97%)by using large amount of patients se

10、izure data for training.Patients need undergo time-consuming and costly hospitalization.1 Y.Wang,ISSCC204 M.Zhang,JSSC226 J.Liu,ISSCC21Non-Seizure Data(Easy to Collect)Seizure Data(Difficult to Collect)Labeled EEG Data of the PatientSeizure Detection ProcessorWeeks Months Hospitalization!3 U.Shin,JS

11、SC22Patient-SpecificTrainingSeizure Detection ClassifierChallenge 1:Time-Consuming&Costly Hospitalization33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE I

12、nternational Solid-State Circuits Conference6 of 57Challenge 2:Limited Detection Accuracy C.Tsai,ISSCC23 presented the first zero-shot-retraining(i.e.zero seizure date is needed for training)seizure detection processor,no need for hospitalization.The accuracy is still limited for practical use.1 Y.W

13、ang,ISSCC202 S.Huang,JSSC203 U.Shin,JSSC224 M.Zhang,JSSC226 J.Liu,ISSCC217 A.Chua,JSSC2022Accuracy Comparison with Prior Patient-Specific Training Processors95%99%90%Pros:No Need for HospitalizationOnly 2-Min Non-Seizure DataCalibrateNon-Seizure PattenSeizure PattenCons:Limited Detection Accuracy5 C

14、.Tsai,ISSCC23:The 1stZero-Short Retraining Processor5 C.Tsai,ISSCC2312 34 6Sensitivity 90%Specificity 93%Latency 8.3sTest Patients EEG 7533.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adapti

15、ve Channel Selection 2024 IEEE International Solid-State Circuits Conference7 of 57Challenge 3:Large Energy Consumption The large neural network in 5 result in mW level power consumption,unsuitable for ultra-low power wearable EEG monitoring devices.Power ConsumptionWmWModel Size 120KBNo.of MAC/Clas

16、s.5MMulti-Channel EEGDense LayerMiddle LayerDeep Layer22x1x6416x116x1x256Neural Pattens2 S.Huang,JSSC205 C.Tsai,ISSCC23Estimated SoC Power 5 1.8mW2.8mW1 Y.Wang,ISSCC203 U.Shin,JSSC226 J.Liu,ISSCC217 A.Chua,JSSC202213672 5 Inception-based CNN in 5 C.Tsai,ISSCC2333.1:A High Accuracy and Energy-Efficie

17、nt Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference8 of 57Outline Background and challenges Zero-Shot-Retraining Seizure Detection Processor Hybrid-fea

18、ture-driven adaptive processing architecture with on-chip learningLearning-based Adaptive channel selection technique Measurement results Conclusions33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-

19、Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference9 of 57Proposed Seizure Detection ProcessorFeature Fusion ControllerData Buffer&Data SegmentLCSMNN Feature Extraction ControllerShared ProcessingElementsShared ProcessingElementsShared ProcessingElementsShared Pr

20、ocessingElementsWeight MemoryBias MemoryROLEHAPEData MemoryHybrid-Feature Classifiation ControllerHybrid-Feature Buffer8-bitChannelMultiplexerManual-Feature BufferFreqency-Domain Feature ExtractionTime-Domain Feature ExtractionCh#0Ch#15Feature-Driven Adaptive Processing Controller NN FeaturesSeizure

21、 DetectionResultsChannel SwitchesSurfaceEEG ElectrodesMulti-Channel EEGAIChipMFEEFeatures of Ch#15Features of Ch#0.Ch_Sel 15:0Segmented EEGActivateManual FeaturesShared NN Processing ElementsActivationBufferWeight BufferMulti-FuncMACUnitCh_Sel 15:0Channel WeightsCh_Sel 15:0Data InterfaceNN Instructi

22、ons&WeightsZero-Shot-Retraining Seizure Processor w/High Accuracy&Energy Efficiency8-bit8-bit8/16bit8-bit33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE I

23、nternational Solid-State Circuits Conference10 of 57Proposed Seizure Detection Processor MFEE:Multi-feature extraction engine HAPE:Hybrid-feature-driven adaptive processing engine ROLE:Reconfigurable on-chip learning engine LCSM:Learning-based channel selection module Data Buffer SPI Data InterfaceO

24、verall ArchitectureFeature Fusion ControllerData Buffer&Data SegmentLCSMNN Feature Extraction ControllerShared ProcessingElementsShared ProcessingElementsShared ProcessingElementsShared ProcessingElementsWeight MemoryBias MemoryROLEHAPEData MemoryHybrid-Feature Classifiation ControllerHybrid-Feature

25、 Buffer8-bitChannelMultiplexerManual-Feature BufferFreqency-Domain Feature ExtractionTime-Domain Feature ExtractionCh#0Ch#15Feature-Driven Adaptive Processing Controller NN FeaturesSeizure DetectionResultsMFEEFeatures of Ch#15Features of Ch#0.Ch_Sel 15:0Segmented EEGActivateManual FeaturesShared NN

26、Processing ElementsActivationBufferWeight BufferMulti-FuncMACUnitChannel WeightsCh_Sel 15:0Data InterfaceNN Instructions&Weights8-bit8-bit8/16bit8-bit.Multi-Channel EEGSPI33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Pr

27、ocessing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference11 of 57Proposed Seizure Detection Processor MFEE:Multi-feature extraction engine HAPE:Hybrid-feature-driven adaptive processing engine ROLE:Reconfigurable on-chip learning engine LCSM:Learn

28、ing-based channel selection module Data Buffer SPI Data InterfaceOverall ArchitectureFeature Fusion ControllerData Buffer&Data SegmentLCSMNN Feature Extraction ControllerShared ProcessingElementsShared ProcessingElementsShared ProcessingElementsShared ProcessingElementsWeight MemoryBias MemoryROLEHA

29、PEData MemoryHybrid-Feature Classifiation ControllerHybrid-Feature Buffer8-bitChannelMultiplexerManual-Feature BufferFreqency-Domain Feature ExtractionTime-Domain Feature ExtractionCh#0Ch#15Feature-Driven Adaptive Processing Controller NN FeaturesSeizure DetectionResultsMFEEFeatures of Ch#15Features

30、 of Ch#0.Ch_Sel 15:0Segmented EEGActivateManual FeaturesShared NN Processing ElementsActivationBufferWeight BufferMulti-FuncMACUnitChannel WeightsCh_Sel 15:0Data InterfaceNN Instructions&Weights8-bit8-bit8/16bit8-bit.Multi-Channel EEGSPI33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining

31、Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference12 of 57Proposed Seizure Detection Processor MFEE:Multi-feature extraction engine HAPE:Hybrid-feature-driven adaptive proces

32、sing engine ROLE:Reconfigurable on-chip learning engine LCSM:Learning-based channel selection module Data Buffer SPI Data InterfaceOverall ArchitectureFeature Fusion ControllerData Buffer&Data SegmentLCSMNN Feature Extraction ControllerShared ProcessingElementsShared ProcessingElementsShared Process

33、ingElementsShared ProcessingElementsWeight MemoryBias MemoryROLEHAPEData MemoryHybrid-Feature Classifiation ControllerHybrid-Feature Buffer8-bitChannelMultiplexerManual-Feature BufferFreqency-Domain Feature ExtractionTime-Domain Feature ExtractionCh#0Ch#15Feature-Driven Adaptive Processing Controlle

34、r NN FeaturesSeizure DetectionResultsMFEEFeatures of Ch#15Features of Ch#0.Ch_Sel 15:0Segmented EEGActivateManual FeaturesShared NN Processing ElementsActivationBufferWeight BufferMulti-FuncMACUnitChannel WeightsCh_Sel 15:0Data InterfaceNN Instructions&Weights8-bit8-bit8/16bit8-bit.Multi-Channel EEG

35、SPI33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference13 of 57Proposed Seizure Detection Processor MFEE:Multi-fe

36、ature extraction engine HAPE:Hybrid-feature-driven adaptive processing engine ROLE:Reconfigurable on-chip learning engine LCSM:Learning-based channel selection module Data Buffer SPI Data InterfaceOverall ArchitectureFeature Fusion ControllerData Buffer&Data SegmentLCSMNN Feature Extraction Controll

37、erShared ProcessingElementsShared ProcessingElementsShared ProcessingElementsShared ProcessingElementsWeight MemoryBias MemoryROLEHAPEData MemoryHybrid-Feature Classifiation ControllerHybrid-Feature Buffer8-bitChannelMultiplexerManual-Feature BufferFreqency-Domain Feature ExtractionTime-Domain Featu

38、re ExtractionCh#0Ch#15Feature-Driven Adaptive Processing Controller NN FeaturesSeizure DetectionResultsMFEEFeatures of Ch#15Features of Ch#0.Ch_Sel 15:0Segmented EEGActivateManual FeaturesShared NN Processing ElementsActivationBufferWeight BufferMulti-FuncMACUnitChannel WeightsCh_Sel 15:0Data Interf

39、aceNN Instructions&Weights8-bit8-bit8/16bit8-bit.Multi-Channel EEGSPI33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Co

40、nference14 of 57Proposed Seizure Detection Processor MFEE:Multi-feature extraction engine HAPE:Hybrid-feature-driven adaptive processing engine ROLE:Reconfigurable on-chip learning engine LCSM:Learning-based channel selection module Data Buffer SPI Data InterfaceOverall ArchitectureFeature Fusion Co

41、ntrollerData Buffer&Data SegmentLCSMNN Feature Extraction ControllerShared ProcessingElementsShared ProcessingElementsShared ProcessingElementsShared ProcessingElementsWeight MemoryBias MemoryROLEHAPEData MemoryHybrid-Feature Classifiation ControllerHybrid-Feature Buffer8-bitChannelMultiplexerManual

42、-Feature BufferFreqency-Domain Feature ExtractionTime-Domain Feature ExtractionCh#0Ch#15Feature-Driven Adaptive Processing Controller NN FeaturesSeizure DetectionResultsMFEEFeatures of Ch#15Features of Ch#0.Ch_Sel 15:0Segmented EEGActivateManual FeaturesShared NN Processing ElementsActivationBufferW

43、eight BufferMulti-FuncMACUnitChannel WeightsCh_Sel 15:0Data InterfaceNN Instructions&Weights8-bit8-bit8/16bit8-bit.Multi-Channel EEGSPI33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive

44、 Channel Selection 2024 IEEE International Solid-State Circuits Conference15 of 57Proposed Seizure Detection Processor MFEE:Multi-feature extraction engine HAPE:Hybrid-feature-driven adaptive processing engine ROLE:Reconfigurable on-chip learning engine LCSM:Learning-based channel selection module D

45、ata Buffer SPI Data InterfaceOverall ArchitectureFeature Fusion ControllerData Buffer&Data SegmentLCSMNN Feature Extraction ControllerShared ProcessingElementsShared ProcessingElementsShared ProcessingElementsShared ProcessingElementsWeight MemoryBias MemoryROLEHAPEData MemoryHybrid-Feature Classifi

46、ation ControllerHybrid-Feature Buffer8-bitChannelMultiplexerManual-Feature BufferFreqency-Domain Feature ExtractionTime-Domain Feature ExtractionCh#0Ch#15Feature-Driven Adaptive Processing Controller NN FeaturesSeizure DetectionResultsMFEEFeatures of Ch#15Features of Ch#0.Ch_Sel 15:0Segmented EEGAct

47、ivateManual FeaturesShared NN Processing ElementsActivationBufferWeight BufferMulti-FuncMACUnitChannel WeightsCh_Sel 15:0Data InterfaceNN Instructions&Weights8-bit8-bit8/16bit8-bit.Multi-Channel EEGSPI33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hyb

48、rid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference16 of 57State-of-the-Arts(Manual-Feature based Classification)Overlapped WindowsEEG ChannelsML-based Classifier Time-Domain FeaturesZero Crossing RateLine Lengt

49、hTest Patients EEG Frequency-Domain FeaturesSpectral Band PowerPower Spectral Density Other FeaturesShort-time Fourier TransformHilbert TransformManual Feature ExtractionSupport Vector Machine 124Logistic Regression 8Neural Tree 3Pros:Lower computation complexity More robust performance across patie

50、ntsCons:Limited classification accuracy33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference17 of 57State-of-the-A

51、rts(End-to-End NN-based Classification)Overlapped WindowsEEG ChannelsPublic EEG DatasetTest Patients EEGEnd-to-End Neural Network6-8Pros:Less feature engineering effort High classification accuracyCons:More sensitive to inter-patient variation High computation complexityNN FeatureExtractionSeizure C

52、lassificationSeizureProbabilityOff-Chip TrainingNormalSeizurePatientsEEG/Label in Dataset33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Sol

53、id-State Circuits Conference18 of 57Hybrid-Feature Driven Adaptive Processing w/On-Chip Learning02040608010012000.51 Pre-Processing based on 3 IIR FilteringGainFrequency/HzPassbands|=|=|=|+|=Spectral Band Power(,)&Line Length FeaturesDepthwise Conv1 w/Dialated Conv Pointwise Conv1 LayerMax-Pooling1L

54、ayer 16-32Hz 32-96Hz 8-16Hz Flatten ReLuDepthwise Conv2 w/Dialated Conv Pointwise Conv2LayerMax-Pooling2LayerReLuLightweight Conv Block#1Lightweight Conv Block#24MFx16ch.96x1 NNF.16x1 16x1 FC1_1FC1_2Hybrid-Feature Class.ResultsManual-Feature Class.ResultsFC2_1FC2_2 threshold,output Non-Seizure thres

55、hold,activate the hybrid-feature classification NN threshold,output Non-Seizure threshold,output Seizure 33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE I

56、nternational Solid-State Circuits Conference19 of 57Manual Features Extraction02040608010012000.51 Pre-Processing based on 3 IIR FilteringGainFrequency/HzPassbands|=|=|=|+|=Spectral Band Power(,)&Line Length FeaturesDepthwise Conv1 w/Dialated Conv Pointwise Conv1 LayerMax-Pooling1Layer 16-32Hz 32-96

57、Hz 8-16Hz Flatten ReLuDepthwise Conv2 w/Dialated Conv Pointwise Conv2LayerMax-Pooling2LayerReLuLightweight Conv Block#1Lightweight Conv Block#24MFx16ch.96x1 NNF.16x1 16x1 FC1_1FC1_2Hybrid-Feature Class.ResultsManual-Feature Class.ResultsFC2_1FC2_2 threshold,output Non-Seizure threshold,activate the

58、hybrid-feature classification NN threshold,output Non-Seizure threshold,output Seizure Hybrid-Feature Driven Adaptive Processing w/On-Chip Learning33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Ba

59、sed Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference20 of 5702040608010012000.51 Pre-Processing based on 3 IIR FilteringGainFrequency/HzPassbands|=|=|=|+|=Spectral Band Power(,)&Line Length FeaturesDepthwise Conv1 w/Dialated Conv Pointwise Conv1 LayerMax-Pooling1Lay

60、er 16-32Hz 32-96Hz 8-16Hz Flatten ReLuDepthwise Conv2 w/Dialated Conv Pointwise Conv2LayerMax-Pooling2LayerReLuLightweight Conv Block#1Lightweight Conv Block#24MFx16ch.96x1 NNF.16x1 16x1 FC1_1FC1_2Hybrid-Feature Class.ResultsManual-Feature Class.ResultsFC2_1FC2_2 threshold,output Non-Seizure thresho

61、ld,activate the hybrid-feature classification NN threshold,output Non-Seizure threshold,output Seizure NN Features ExtractionNo.of Weights:0.4K,No.MAC Operations:0.02MHybrid-Feature Driven Adaptive Processing w/On-Chip Learning33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure De

62、tection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference21 of 5702040608010012000.51 Pre-Processing based on 3 IIR FilteringGainFrequency/HzPassbands|=|=|=|+|=Spectral Band Power(,)&Line Len

63、gth FeaturesDepthwise Conv1 w/Dialated Conv Pointwise Conv1 LayerMax-Pooling1Layer 16-32Hz 32-96Hz 8-16Hz Flatten ReLuDepthwise Conv2 w/Dialated Conv Pointwise Conv2LayerMax-Pooling2LayerReLuLightweight Conv Block#1Lightweight Conv Block#24MFx16ch.96x1 NNF.16x1 16x1 FC1_1FC1_2Hybrid-Feature Class.Re

64、sultsManual-Feature Class.ResultsFC2_1FC2_2 threshold,output Non-Seizure threshold,activate the hybrid-feature classification NN threshold,output Non-Seizure threshold,output Seizure Hybrid-Feature based ClassificationHybrid-Feature Driven Adaptive Processing w/On-Chip Learning33.1:A High Accuracy a

65、nd Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference22 of 5702040608010012000.51 Pre-Processing based on 3 IIR FilteringGainFrequency/H

66、zPassbands|=|=|=|+|=Spectral Band Power(,)&Line Length FeaturesDepthwise Conv1 w/Dialated Conv Pointwise Conv1 LayerMax-Pooling1Layer 16-32Hz 32-96Hz 8-16Hz Flatten ReLuDepthwise Conv2 w/Dialated Conv Pointwise Conv2LayerMax-Pooling2LayerReLuLightweight Conv Block#1Lightweight Conv Block#24MFx16ch.9

67、6x1 NNF.16x1 16x1 FC1_1FC1_2Hybrid-Feature Class.ResultsManual-Feature Class.ResultsFC2_1FC2_2 threshold,output Non-Seizure threshold,activate the hybrid-feature classification NN threshold,output Non-Seizure threshold,output Seizure Hybrid-Feature Driven Adaptive ProcessingHybrid-Feature Driven Ada

68、ptive Processing w/On-Chip Learning33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference23 of 5702040608010012000.

69、51 Pre-Processing based on 3 IIR FilteringGainFrequency/HzPassbands|=|=|=|+|=Spectral Band Power(,)&Line Length FeaturesDepthwise Conv1 w/Dialated Conv Pointwise Conv1 LayerMax-Pooling1Layer 16-32Hz 32-96Hz 8-16Hz Flatten ReLuDepthwise Conv2 w/Dialated Conv Pointwise Conv2LayerMax-Pooling2LayerReLuL

70、ightweight Conv Block#1Lightweight Conv Block#24MFx16ch.96x1 NNF.16x1 16x1 FC1_1FC1_2Hybrid-Feature Class.ResultsManual-Feature Class.ResultsFC2_1FC2_2 threshold,output Non-Seizure threshold,activate the hybrid-feature classification NN threshold,output Non-Seizure threshold,output Seizure Hybrid-Fe

71、ature Driven Adaptive ProcessingNN feature extraction&hybrid-feature classification are adaptively activated according to the manual-feature classification results.Manual feature classification are biased trained towards classify the input as seizure.Hybrid-Feature Driven Adaptive Processing w/On-Ch

72、ip Learning33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference24 of 57Hardware Architecture of MFEE&HAPEPNNF0 PN

73、NF15Spectral Band Power CalculatorReg|.|2Bandpass IIR FilterReg|.|RegRegRing BufferZero/Pole CoefAcc CtrollerBand Power UnitLine Length CalculatorChannel Multiplexer.Manual Feature BufferActs MEM0Acts MEM1Hybrid-Feature BufferPMF0 PMF15MFEEDEMUXMUXWeight MEMBias MEMInstruction MEMNNInput BufferPing-

74、Pong MEMProjected FeaturesActivationsManual FeaturesShiftAdderMUX4/8 Bit FlagZero-Value Flag08b WgtsLow 4b ActsHigh 4b ActsAdderTreeRegBiasAccumulationControllerAccumulatorWeightBufferActivation BufferInstruction DecoderHybrid-Feature Classification ControllerNN Feature ExtractionControllerFeature F

75、usion ControllerHybrid-Feature-Driven Adaptive Processing ControllerMulti-Precision MAC UnitDEMUXHybrid FeaturesWgtsHAPEReLu UnitMax-Pooling UnitOutput Acts(Manual-Feature Class Results)MUXManual-Feature Class ResultsW/R Zero-Skipping ControllerNN_CtrlNN_CtrlOutput ActsNN_CtrlSkip Y/NWake-upWake-upI

76、nstructionBiasWake-upWake-up ModuleDepthwise Conv&Dilated Conv ModePointwise Conv&Max-Pooling ModeFC ModeLayer Computation ControllerNN Instruction&Weights&BiasSegmented EEG3 Calculators Reused by 16-ChCh_SelWgtsInput Acts33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detecti

77、on Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference25 of 57PNNF0 PNNF15Spectral Band Power CalculatorReg|.|2Bandpass IIR FilterReg|.|RegRegRing BufferZero/Pole CoefAcc CtrollerBand Power Uni

78、tLine Length CalculatorChannel Multiplexer.Manual Feature BufferActs MEM0Acts MEM1Hybrid-Feature BufferPMF0 PMF15DEMUXMUXWeight MEMBias MEMInstruction MEMNNInput BufferPing-Pong MEMProjected FeaturesActivationsShiftAdderMUX4/8 Bit FlagZero-Value Flag08b WgtsLow 4b ActsHigh 4b ActsAdderTreeRegBiasAcc

79、umulationControllerAccumulatorWeightBufferActivation BufferInstruction DecoderHybrid-Feature Classification ControllerNN Feature ExtractionControllerFeature Fusion ControllerHybrid-Feature-Driven Adaptive Processing ControllerMulti-Precision MAC UnitDEMUXHybrid FeaturesWgtsHAPEReLu UnitMax-Pooling U

80、nitOutput Acts(Manual-Feature Class Results)MUXManual-Feature Class ResultsW/R Zero-Skipping ControllerNN_CtrlNN_CtrlOutput ActsNN_CtrlSkip Y/NWake-upInstructionBiasWake-upWake-up ModuleDepthwise Conv&Dilated Conv ModePointwise Conv&Max-Pooling ModeFC ModeLayer Computation ControllerNN Instruction&W

81、eights&BiasSegmented EEG3 Calculators Reused by 16-ChCh_SelWgtsManual FeaturesHardware Architecture of MFEE&HAPE 3 Spectral Band(,)Power CalculatorsBandpass IIR FilterApproximate Band Power Unit 1 Line Length CalculatorMFEEReused by 16-channel EEG through time multiplexing33.1:A High Accuracy and En

82、ergy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference26 of 57Hardware Architecture of MFEE&HAPE Reconfigurable Neural Network Archtecture1-D

83、depth-wise/point-wise/dilated ConvMaxPooling/FC Mode 16 Multi-Precision MAC UnitsDynamic 4b/8bmultiplication Activations Write/Read Zero-SkippingHAPEPNNF0 PNNF15Spectral Band Power CalculatorReg|.|2Bandpass IIR FilterReg|.|RegRegRing BufferZero/Pole CoefAcc CtrollerBand Power UnitLine Length Calcula

84、torChannel Multiplexer.Manual Feature BufferActs MEM0Acts MEM1Hybrid-Feature BufferPMF0 PMF15MFEEDEMUXMUXWeight MEMBias MEMInstruction MEMNNInput BufferPing-Pong MEMProjected FeaturesActivationsShiftAdderMUX4/8 Bit FlagZero-Value Flag08b WgtsLow 4b ActsHigh 4b ActsAdderTreeRegBiasAccumulationControl

85、lerAccumulatorWeightBufferActivation BufferInstruction DecoderHybrid-Feature Classification ControllerNN Feature ExtractionControllerFeature Fusion ControllerHybrid-Feature-Driven Adaptive Processing ControllerMulti-Precision MAC UnitDEMUXHybrid FeaturesWgtsHAPEReLu UnitMax-Pooling UnitOutput Acts(M

86、anual-Feature Class Results)MUXManual-Feature Class ResultsW/R Zero-Skipping ControllerNN_CtrlNN_CtrlOutput ActsNN_CtrlSkip Y/NWake-upInstructionBiasWake-upWake-up ModuleDepthwise Conv&Dilated Conv ModePointwise Conv&Max-Pooling ModeFC ModeLayer Computation ControllerNN Instruction&Weights&Bias3 Cal

87、culators Reused by 16-ChCh_SelWgtsSegmented EEGManual FeaturesWake-upInput Acts33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State C

88、ircuits Conference27 of 57PNNF0 PNNF15Spectral Band Power CalculatorReg|.|2Bandpass IIR FilterReg|.|RegRegRing BufferZero/Pole CoefAcc CtrollerBand Power UnitLine Length CalculatorChannel Multiplexer.Manual Feature BufferActs MEM0Acts MEM1Hybrid-Feature BufferPMF0 PMF15MFEEDEMUXMUXWeight MEMBias MEM

89、Instruction MEMNNInput BufferPing-Pong MEMProjected FeaturesActivationsShiftAdderMUX4/8 Bit FlagZero-Value Flag08b WgtsLow 4b ActsHigh 4b ActsAdderTreeRegBiasAccumulationControllerAccumulatorWeightBufferActivation BufferInstruction DecoderHybrid-Feature Classification ControllerNN Feature Extraction

90、ControllerFeature Fusion ControllerHybrid-Feature-Driven Adaptive Processing ControllerMulti-Precision MAC UnitDEMUXHybrid FeaturesWgtsHAPEReLu UnitMax-Pooling UnitOutput Acts(Manual-Feature Class Results)MUXManual-Feature Class ResultsW/R Zero-Skipping ControllerNN_CtrlNN_CtrlOutput ActsNN_CtrlSkip

91、 Y/NWake-upInstructionBiasWake-upWake-up ModuleDepthwise Conv&Dilated Conv ModePointwise Conv&Max-Pooling ModeFC ModeLayer Computation ControllerNN Instruction&Weights&Bias3 Calculators Reused by 16-ChCh_SelWgtsSegmented EEGManual FeaturesWake-upInput ActsHardware Architecture of MFEE&HAPEHybrid-Fea

92、ture-Driven Adaptive Processing FlowHAPEWaitEEG SampleManualFeatureExtractionManualFeatureProjectionManualFeatureClassNNFeatureExtractionNNFeatureProjectionHybrid Feature ClassIf Non-SeizureIf Seizure,ActivateFeature Fusion33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detect

93、ion Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference28 of 57PNNF0 PNNF15Spectral Band Power CalculatorReg|.|2Bandpass IIR FilterReg|.|RegRegRing BufferZero/Pole CoefAcc CtrollerBand Power Un

94、itLine Length CalculatorChannel Multiplexer.Manual Feature BufferActs MEM0Acts MEM1Hybrid-Feature BufferPMF0 PMF15MFEEDEMUXMUXWeight MEMBias MEMInstruction MEMNNInput BufferPing-Pong MEMProjected FeaturesActivationsShiftAdderMUX4/8 Bit FlagZero-Value Flag08b WgtsLow 4b ActsHigh 4b ActsAdderTreeRegBi

95、asAccumulationControllerAccumulatorWeightBufferActivation BufferInstruction DecoderHybrid-Feature Classification ControllerNN Feature ExtractionControllerFeature Fusion ControllerHybrid-Feature-Driven Adaptive Processing ControllerMulti-Precision MAC UnitDEMUXHybrid FeaturesWgtsHAPEReLu UnitMax-Pool

96、ing UnitOutput Acts(Manual-Feature Class Results)MUXManual-Feature Class ResultsW/R Zero-Skipping ControllerNN_CtrlNN_CtrlOutput ActsNN_CtrlSkip Y/NWake-upInstructionBiasWake-upWake-up ModuleDepthwise Conv&Dilated Conv ModePointwise Conv&Max-Pooling ModeFC ModeLayer Computation ControllerNN Instruct

97、ion&Weights&Bias3 Calculators Reused by 16-ChCh_SelWgtsSegmented EEGManual FeaturesWake-upInput ActsHardware Architecture of MFEE&HAPEHybrid-Feature-Driven Adaptive Processing FlowHAPE Manual-Feature Classification(account for 68%)WaitEEG SampleManualFeatureExtractionManualFeatureProjectionManualFea

98、tureClassNNFeatureExtractionNNFeatureProjectionHybrid Feature ClassIf Non-SeizureIf Seizure,ActivateFeature Fusion33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 20

99、24 IEEE International Solid-State Circuits Conference29 of 57WaitEEG SampleManualFeatureExtractionManualFeatureProjectionManualFeatureClassNNFeatureExtractionNNFeatureProjectionHybrid Feature ClassIf Non-SeizureIf Seizure,ActivateFeature FusionPNNF0 PNNF15Spectral Band Power CalculatorReg|.|2Bandpas

100、s IIR FilterReg|.|RegRegRing BufferZero/Pole CoefAcc CtrollerBand Power UnitLine Length CalculatorChannel Multiplexer.Manual Feature BufferActs MEM0Acts MEM1Hybrid-Feature BufferPMF0 PMF15MFEEDEMUXMUXWeight MEMBias MEMInstruction MEMNNInput BufferPing-Pong MEMProjected FeaturesActivationsShiftAdderM

101、UX4/8 Bit FlagZero-Value Flag08b WgtsLow 4b ActsHigh 4b ActsAdderTreeRegBiasAccumulationControllerAccumulatorWeightBufferActivation BufferInstruction DecoderHybrid-Feature Classification ControllerNN Feature ExtractionControllerFeature Fusion ControllerHybrid-Feature-Driven Adaptive Processing Contr

102、ollerMulti-Precision MAC UnitDEMUXHybrid FeaturesWgtsHAPEReLu UnitMax-Pooling UnitOutput Acts(Manual-Feature Class Results)MUXManual-Feature Class ResultsW/R Zero-Skipping ControllerNN_CtrlNN_CtrlOutput ActsNN_CtrlSkip Y/NWake-upInstructionBiasWake-upWake-up ModuleDepthwise Conv&Dilated Conv ModePoi

103、ntwise Conv&Max-Pooling ModeFC ModeLayer Computation ControllerNN Instruction&Weights&Bias3 Calculators Reused by 16-ChCh_SelWgtsSegmented EEGManual FeaturesWake-upInput ActsHardware Architecture of MFEE&HAPEHybrid-Feature-Driven Adaptive Processing FlowHAPE Manual-Feature Classification(account for

104、 68%)Hybrid-Feature Classification(account for 32%)33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference30 of 57No

105、n-Seizure ClassTest Patients EEGPublic Dataset EEGSeizure ClassNon-Seizure ClassPublic Dataset EEGSeizure Class Non-seizure data of test patient directly used for retraining Cons:Specificity increases quickly(to 100%),but sensitivity degrades(to 0%).Trained on public EEG dataset,high accuracy Tested

106、 on unseen new patients,due to the inter-patient variations,the accuracy degrades significantly.Hybrid-Feature Driven Adaptive Processing w/On-Chip Learning33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Le

107、arning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference31 of 57RetrainHybrid-Feature Classification CNN Pretrain1-Min Non-Seizure Dataof the Test Patient1-Min Seizure Data From the Public Dataset23 Patients EEG in CHB-MIT Non-seizure data from the patient&seiz

108、ure data from the public dataset are mixed as the on-chip learning dataset.Only the hybrid-feature classification layer&the projection layers are re-trained.Hybrid-Feature Driven Adaptive Processing w/On-Chip Learning33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Pr

109、ocessor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference32 of 574MFx16ch.96x1 NNF.16x1 16x1 FC1_1FC1_2Hybrid-Feature Class.ResultsManual-Feature Class.ResultsFC2_1FC2_2RetrainHybrid-Feature Classifica

110、tion CNN Pretrain1-Min Non-Seizure Dataof the Test Patient1-Min Seizure Data From the Public Dataset23 Patients EEG in CHB-MIT Only the hybrid-feature classification layer&the projection layers are re-trained.Hybrid-Feature Driven Adaptive Processing w/On-Chip Learning33.1:A High Accuracy and Energy

111、-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference33 of 57Retraining Stage#1Retraining Stage#2Inference of the HFNNCalculate HF ClassOutput No

112、de Loss Retraining of HF Class.Layer FC2_2Retraining HF Layer FC1_1&FC1_2Calculate MF ClassOutput Node Loss Retraining MF Class.Layer FC2_1Inference of the MFNNTwo-stage on-chip learning processing flowRetrainHybrid-Feature Classification CNN Pretrain1-Min Non-Seizure Dataof the Test Patient1-Min Se

113、izure Data From the Public Dataset23 Patients EEG in CHB-MITHybrid-Feature Driven Adaptive Processing w/On-Chip Learning33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Select

114、ion 2024 IEEE International Solid-State Circuits Conference34 of 570%20%40%60%80%100%123456789101112131415161718192021222324Before Zero-Shot RetrainingAfter Zero-Shot RetrainingPatient IDAccuracyInter-patient test in surface EEG CHB-MIT dataset.Average classification accuracy improved from 76.7%to 9

115、3.1%.Only 0.3%degradationin Patient 12 Hybrid-Feature Driven Adaptive Processing w/On-Chip Learning33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE Interna

116、tional Solid-State Circuits Conference35 of 57Hardware Architecture of ROLEMUXMUXWeight&Bias Updating CircuitwoldboldLUT-BasedSigmoid ModuleMUXMUXActswiFC2 dbdyiFC1 or dwdyiFC1 Learning Rate(lr)Label(0/1)Pred_Err1 Gradient&Loss Back Propagation Calculation 2 Weight&Bias Updatingwnew=wold+dwbnew=bold

117、+dbPred_Err=Sigmoid(Out_ActsFC2)-LabeldbFC2=-lrFC2*Pred_Errfor i=1:32 dwiFC2=-lrFC2*Pred_Err*In_ActsFC2i for i=1:32 dyiFC1=wiFC2*Pred_Err dbiFC1=-lrFC1*dyiFC1 for j=1:96+64:dwi,jFC1=-lrFC1*dyiiFC1*In_ActsFC1jReuseReuseOutFC2Pred_ErrTrunctTrunctRegReg16b8b16b16b16b16b8b8b16b16bdwdbTrunctMUX24b24b16b1

118、6b16b8bwnew or bnew 8/16b16bGradient&Loss Back Propagation Calculation Circuit-lrFC2*Pred_Err-lrFC1*dyiFC1 32b24b 5%2%10 J.Liu,DAC2233.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Ch

119、annel Selection 2024 IEEE International Solid-State Circuits Conference39 of 57.Ch#0.Ch#15FC1_1FC2_1OutFC2_1From FC1_2Learning-based Channel Selection TechniqueStep 1:On-chip learning of the hybrid-feature CNN.Step 2:Inference using 8-s non-seizure data from the patient&8-s seizure data from the pub

120、lic dataset.Step 3:Calculate activation rate of toStep 4:Calculate channel importance scores,each channel contains 4 sub-channel corresponding to 4 manual features.Step 5:Channel score sorting&remove the K EEG channels with the lowest scores.Activation Rate 33.1:A High Accuracy and Energy-Efficient

121、Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference40 of 57Computation Reduction in Manual-Feature Extraction&ClassificationEEG Ch#0EEG Ch#1EEG Ch#14EEG C

122、h#15.MF ExtractionMF ExtractionMF ExtractionMF Extraction.SkippedSkipped.Computation Skipping in Depthwise Conv&Pointwise Conv LayerSkippedSkippedComputation Reduction in NN-Feature Extraction Computation reduction of disable EEG channels Manual feature extraction and classification NN feature extra

123、ction computation in Depthwise Conv&Pointwise Conv layersProposed Learning-based Channel Selection Technique33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEE

124、E International Solid-State Circuits Conference41 of 57Hardware Architecture of LCSMReg0Reg1.Reg14Reg15DEMUXMUXMUX01CMP1b0OutFC1_1DEMUXReg1st Sample Inference63rd SampleInference Activation Rate Calculation CircuitAR Calculation ControllerChannel Score Calculation CircuitReg0Reg1Reg14Reg15DEMUXMUXMI

125、N RegScore CalculationControllerRegRegSub-ChannelCntChannelCntAccumulatorAccumulatorSortingController.AR CtrlCtrlCS CtrlCS CtrlAR Ctrl5CMPChannel Score Sorting CircuitWeights of FC1_1&FC2_1KCh_Sel0:15.LCSM accounts for only 0.1%hardware overhead of the entire design.33.1:A High Accuracy and Energy-E

126、fficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference42 of 57Outline Background and challenges Zero-Shot-Retraining Seizure Detection Processor Hyb

127、rid-feature-driven adaptive processing architecture with on-chip learningLearning-based Adaptive channel selection technique Measurement results Conclusions33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Le

128、arning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference43 of 57Measurement ResultsSpecificationsTechnology55 nmCore Area0.98 mm2Data PrecisionINT8/INT16SRAM Size8 KBSupply Voltage0.68 VFreqency2.5 MHzPower Consumption25.8 WFabricated in 55nm CMOS processTest S

129、etup33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference44 of 57Measurement Results23 Patients EEG in CHB-MIT1 Te

130、st Patients EEG in CHB-MITPretrainTestHybrid-Feature Classification CNN Inter-patient test methodTesting without any patient data for training Frame-based VS Event-basedsensitivity4 M.Zhang,JSSC2233.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-F

131、eature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference45 of 5774.7%67.9%74.9%70.3%78.5%74.6%90.2%94.8%65%70%75%80%85%90%95%100%SensitivitySpecificityPercentageManual Feature+NNEnd-to-End NNHF-CNNHF-CNN w/On-Chip Learnin

132、g11.7%20.2%3.8%4.3%Measurement Results Hybrid-feature NN achieves higher accuracy than both manual features based and end-to-end NN based classification On-chip learning using only 1-minute non-seizure data,frame-based sensitivity is increased to 90.2%,specificity is increased to 94.8%.Event-based s

133、ensitivity:100%33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference46 of 5790.2%94.8%98.8%99.3%85.0%90.0%95.0%100

134、.0%SensitivitySpecificityPercentageOn-Chip Learning using 1-Min Non-Seisure DataOn-Chip Learning using 30%Seisure&Non-Seisure Data8.6%4.5%30%of the patients seizure&non-seizure data used for retraining as existing work 2,4 Sensitivity and specificity is dramatically increased to 98.8%&99.3%.Measurem

135、ent ResultsPatient-Specific Training33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference47 of 570.2uJ0.177uJ0.159

136、uJ0.144uJ0.128uJ0.111uJ0.098uJ0.9030.9080.9050.9040.9080.8940.8960.050.090.130.170.210.250.290.330.370.880.8850.890.8950.90.9050.910.915024681012Energy Consumption/JF1-scoreNo.of disabled channelsEnergy ConsumptionF1 score(calculated fromsensitivity&precision)Reduced by 36%Proposed learning-based ad

137、aptive channel selection reduces the energy consumption by 36%and with slight F1 score improvement.Measurement Results33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selectio

138、n 2024 IEEE International Solid-State Circuits Conference48 of 57 The hybrid-feature-driven adaptive processing reduce the energy consumption reduces by 43.8%with only 0.5%sensitivity loss.Measurement Results93.8%94.0%91.8%91.3%0.128uJ0.072uJ00.0250.050.0750.10.1250.150%20%40%60%80%100%w/o hybrid-fe

139、ature-drivenadaptive processingw/hybrid-feature-drivenadaptive processingEnergy Consumption/JPercentageSpecificitySensitivityEnergy ConsumptionReducedby 43.8%33.1:A High Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor with Hybrid-Feature-Driven Adaptive Processing and

140、Learning-Based Adaptive Channel Selection 2024 IEEE International Solid-State Circuits Conference49 of 57ISSCC2020 1JSSC2020 2JSSC2022 3JSSC2022 4ISSCC2023 5This Work18040654040551.50.581.20.70.90.6840M130K/65K128K1MNA2.5M3.512.553.481.53#1.56#0.986435.83.171341208SVMNL-SVMNeuralTreeGTCA-SVMSciCNNHF

141、-CNNNoYesNoYesYesYes16241210NA8/16CHB-MITCHB-MITCHB-MITCHB-MITCHB-MITCHB-MIT814256162216NA35%80%35%0.1%(2-Min)0.05%(1-Min)Frame-based97.8%96.6%NA97.8%NA98.8%*NANA95.6%100.0%90.3%100%*99.7%(%of patients data fortraining is unkown)NA96.8%99.5%(35%of paitents datafor traning)NA99.30%*(30%of paitents da

142、ta fortraning)NANANANA93.6%94%*0.30.716435.83.171341208SVMNL-SVMNeuralTreeGTCA-SVMSciCNNHF-CNNNoYesNoYesYesYes16241210NA8/16CHB-MITCHB-MITCHB-MITCHB-MITCHB-MITCHB-MIT814256162216NA35%80%35%0.1%(2-Min)0.05%(1-Min)Frame-based97.8%96.6%NA97.8%NA98.8%*NANA95.6%100.0%90.3%100%*99.7%(%of patients data for

143、training is unkown)NA96.8%99.5%(35%of paitents datafor traning)NA99.30%*(30%of paitents data fortraning)NANANANA93.6%94%*0.30.716435.83.171341208SVMNL-SVMNeuralTreeGTCA-SVMSciCNNHF-CNNNoYesNoYesYesYes16241210NA8/16CHB-MITCHB-MITCHB-MITCHB-MITCHB-MITCHB-MIT814256162216NA35%80%35%0.1%(2-Min)0.05%(1-Mi

144、n)Frame-based97.8%96.6%NA97.8%NA98.8%*NANA95.6%100.0%90.3%100%*99.7%(%of patients data fortraining is unkown)NA96.8%99.5%(35%of paitents datafor traning)NA99.30%*(30%of paitents data fortraning)NANANANA93.6%94%*0.30.716435.83.171341208SVMNL-SVMNeuralTreeGTCA-SVMSciCNNHF-CNNNoYesNoYesYesYes16241210NA

145、8/16CHB-MITCHB-MITCHB-MITCHB-MITCHB-MITCHB-MIT814256162216NA35%80%35%0.1%(2-Min)0.05%(1-Min)Frame-based97.8%96.6%NA97.8%NA98.8%*NANA95.6%100.0%90.3%100%*99.7%(%of patients data fortraining is unkown)NA96.8%99.5%(35%of paitents datafor traning)NA99.30%*(30%of paitents data fortraning)NANANANA93.6%94%

146、*0.30.71 90%-96.8%/96.9%-99.7%Sensitivity98.5%-90%96.7%95.6%/94%-97.8%Comparative Analysis:System33.4 A Multi-Loop Neuromodulation Chipset Network with Frequency-Interleaving Front-End and Explainable AI for Memory Studies in Freely Behaving Monkeys 2024 IEEE International Solid-State Circuits Confe

147、rence28 of 31Comparative Analysis:CircuitsPublicationThis workISSCC 23 8 ISSCC 23 9VLSI 23 10ISSCC 22 6 ISSCC 20 4 ISSCC 20 11Neural InterfaceRec ch#8 ch(x3)64 ch22 ch16 ch256 ch8 ch8 chAFE Noise 3.81 V2.2 V-4.7 V3.2 V-1.7 VAFE NEF2.122.55-3.65-2.07ADC ENOB16.89 b14 b11.3 b-AFE BW250/7 kHz2.5 kHz-1

148、kHz500 Hz0.01-250 Hz200/1 kHzStim ch#8 ch(x3)64 ch-10 ch16 ch8 ch8 chStim res8 bN/A-8 bN/AN/AN/ACom.Volt10 V3.3V-5 V8 V*N/AN/AQ-balancingAdaptiveAdaptive-AdaptiveActiveActiveAdaptiveWirelessData rate67 Mbps20 Mbps-35 Mbps-Wireless PwrYesYes-Yes-NoNoTX Power4.11 pJ/b1.38 pJ/b-9.7 pJ/b-Rx Power87 pJ/b

149、-131 pJ/b-Based on mode 1*Estimated or deduced from the paper33.4 A Multi-Loop Neuromodulation Chipset Network with Frequency-Interleaving Front-End and Explainable AI for Memory Studies in Freely Behaving Monkeys 2024 IEEE International Solid-State Circuits Conference29 of 31Highlights of Innovatio

150、ns Flexible star-connected network topology supports multi-loopneuromodulation Tri-mode neural front-end for flexible,energy-efficient recording,featuring frequency interleaving by reusing circuits from adjacent channel ADCs Low-power feature extraction and explainable machine learning for high sens

151、itivity(98.5%),short latency(0.2 ms),and low power(0.19 J/class)closed-loop operation Successful experiments in freely behaving monkeys lead toscientific discoveries33.4 A Multi-Loop Neuromodulation Chipset Network with Frequency-Interleaving Front-End and Explainable AI for Memory Studies in Freely

152、 Behaving Monkeys 2024 IEEE International Solid-State Circuits Conference30 of 31References1 H.Hampel,et al.,“Designing the Next-Generation Clinical Care Pathway for Alzheimers Disease,”Nat.Aging,vol.2,pp.692-703,2022.2 A.S.Rios,et al.,“Optimal Deep Brain Stimulation Sites and Networks for Stimulati

153、on of the Fornix in Alzheimers Disease,”Nat.Commun.,vol.13,no.7707,2022.3 H.Chandrakumar,et al.,“A 15.2-ENOB Continuous-Time ADC for a 7.3W 200mVpp-Linear-Input-Range Neural Recording Front-End,”ISSCC,pp.232-233,2018.4 G.OLeary,et al.,“A Neuromorphic Multiplier-Less Bit-Serial Weight-Memory-Optimize

154、d 1024-Tree Brain-State Classifier and Neuromodulation SoC with an 8-Channel Noise-Shaping SAR ADC Array,”ISSCC,pp.402-403,2020.5 A.G.Richardson,et al.,“Hippocampal Gamma-Slow Oscillation Coupling in Macaques During Sedation and Sleep,”Hippocampus,vol.27,no.11,pp.1125-1139,2017.6 U.Shin,et al.,“A 25

155、6-Channel 0.227J/class Versatile Brain Activity Classification and Closed-Loop Neuromodulation SoC with 0.004mm2-1.51W/channel Fast-Settling Highly Multiplexed Mixed-Signal Front-End,”ISSCC,pp.338-339,2022.33.4 A Multi-Loop Neuromodulation Chipset Network with Frequency-Interleaving Front-End and Ex

156、plainable AI for Memory Studies in Freely Behaving Monkeys 2024 IEEE International Solid-State Circuits Conference31 of 31References7 H.Nori,et al.,“InterpretML:A Unified Framework for Machine Learning Interpretability.”arXiv,Sept.19,2019.8 J.Xu,et al.,“Fascicle-Selective Bidirectional Peripheral Ne

157、rve Interface IC with 173dB FOM Noise-Shaping SAR ADCs and 1.38pJ/b Frequency-Multiplying Current-Ripple Radio Transmitter,”ISSCC,pp.490-491,2023.9 C.-W.Tsai,et al.,“SciCNN:A 0-Shot-Retraining Patient-Independent Epilepsy-Tracking SoC,”ISSCC,pp.488-489,2023.10 Y.Zhu,et al.,“A Wireless Sensor-Brain I

158、nterface System for Tracking and Guiding Animal Behaviors Through Goal-Directed Closed-loop Neuromodulation,”VLSI Circuits Symp.,2023.11 Y.Wang,et al.,“A Closed-Loop Neuromodulation Chipset with 2-Level Classification Achieving 1.5Vpp CM Interference Tolerance,35dB Stimulation Artifact Rejection in

159、0.5ms and 97.8%Sensitivity Seizure Detection,”ISSCC,pp.404-405,2020.33.4 A Multi-Loop Neuromodulation Chipset Network with Frequency-Interleaving Front-End and Explainable AI for Memory Studies in Freely Behaving Monkeys 2024 IEEE International Solid-State Circuits Conference32 of 31Please Scan to R

160、ate Please Scan to Rate This PaperThis Paper33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference1 of 78Closed

161、-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally-Distributed Optogenetic Stimulation Tayebeh Yousefi,Georg Zoidl,Hossein KassiriIntegrated Circuits and Systems LabDepartment of Electrical Engineering and Comp

162、uter Science,York University,Toronto,Ontario,Canada33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference2 of 7

163、8Outline Introduction and Motivation Poisson-Coded Temporally-Distributed Optogenetic Stimulation Analog ED-Based Adaptive-Threshold Spike Detection Measurement Results Comparison and Conclusions33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshol

164、d Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference3 of 78Outline Introduction and Motivation Poisson-Coded Temporally-Distributed Optogenetic Stimulation Analog ED-Based Adaptive-Threshold Spike Detection Measure

165、ment Results Comparison and Conclusions33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference4 of 78Retina Stru

166、cture33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference5 of 78Retina Structure33.5:Closed-Loop 100-Channel

167、Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference6 of 78Retina Structure33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant wi

168、th 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference7 of 78Retina Structure33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-

169、Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference8 of 78Retina Structure33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Pois

170、son-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference9 of 78Retina Structure33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed O

171、ptogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference10 of 78Retina Structure33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE

172、International Solid-State Circuits Conference11 of 78Retina Structure33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuit

173、s Conference12 of 78Retina Signal ProcessingON Pathway:Responsive to light increase33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid

174、-State Circuits Conference13 of 78Retina Signal ProcessingOFF Pathway:Responsive to light decrease33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE Inte

175、rnational Solid-State Circuits Conference14 of 78Retina Signal ProcessingStimulusReconstructed Perception33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IE

176、EE International Solid-State Circuits Conference15 of 78Retinal DegenerationA progressive neurologic disorder:loss of photoreceptorsTwo leading causes:Age-related macular degeneration(AMD)Retinitis pigmentosa(RP)AMD affects 196 million worldwideExpected to double by 2050.33.5:Closed-Loop 100-Channel

177、 Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference16 of 78Retinal DegenerationA progressive neurologic disorder:loss of photoreceptorsT

178、wo leading causes:Age-related macular degeneration(AMD)Retinitis pigmentosa(RP)AMD affects 196 million worldwideExpected to double by 2050.33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distribut

179、ed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference17 of 78Retinal DegenerationA progressive neurologic disorder:loss of photoreceptorsTwo leading causes:Age-related macular degeneration(AMD)Retinitis pigmentosa(RP)AMD affects 196 million worldwideExpected to double by

180、 2050.Rest of retina network Remain functional33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference18 of 78Vis

181、ion Restoration Electrical Stimulation Sub-retinal stimulatorsUse the remaining retinal networkMore invasive33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024

182、 IEEE International Solid-State Circuits Conference19 of 78Vision Restoration Electrical Stimulation Sub-retinal stimulatorsUse the remaining retinal networkMore invasive Epi-retinal stimulatorsLess invasiveNo use of the remaining retinal network33.5:Closed-Loop 100-Channel Highly-Scalable Retinal I

183、mplant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference20 of 78Vision Restoration Electrical Stimulation Sub-retinal stimulatorsUse the remaining retinal networkMore

184、invasive Epi-retinal stimulatorsLess invasiveNo use of the remaining retinal networkProblem with both:Indiscriminative stimulation of ON and OFF pathways33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Tempor

185、ally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference21 of 78Vision Restoration Electrical StimulationDifferentCell Types33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-C

186、oded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference22 of 78Vision Restoration Electrical Stimulation33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Tem

187、porally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference23 of 78Healthy RetinaONParasolOFFParasolONMidgetOFFMidgetElectrical StimulationReconstructedPerceptionVision Restoration Electrical Stimulation33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Impl

188、ant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference24 of 78Electrical vs Optogenetic Stimulation33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.0

189、2W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference25 of 78Vision Restoration Optogenetic StimulationPathway specific stimulationEpiretinal less invasiveUses the remaining retin

190、al network33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference26 of 78Optogenetic Retinal Implant33.5:Closed-

191、Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference27 of 78Challenge 1:Scalability and Energy Efficiency Simultaneous St

192、imulation individual LED addressingTwo wires per LED non-scalableMemory and wireless throughput requirementVery high instantaneous power33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed

193、Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference28 of 78Challenge 2:Non-Uniform Light Sensitivity Genetic modification distribution is not perfectNon-uniform opsin distributionVarying light sensitivity in different regionsQuality and consistency of visual perceptionHig

194、h OpsinConcentrationLow OpsinConcentration33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference29 of 78Outline

195、 Introduction and Motivation Poisson-Coded Temporally-Distributed Optogenetic Stimulation Analog ED-Based Adaptive-Threshold Spike Detection Measurement Results Comparison and Conclusions33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike

196、Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference30 of 78Photon Integration-ConceptIStim33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Pois

197、son-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference31 of 78Photon Integration-Concept2IStim50%duty cycle33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-

198、Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference32 of 78Photon Integration In Vitro Validation33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Tempo

199、rally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference33 of 78Photon Integration In Vitro Validation33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distr

200、ibuted Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference34 of 78Sequential Stimulation Individual LEDsOne LED at a time Instantaneous power For an NN array Sweeping across N2LEDs in one frame Very short time for each LED to deliver its power33.5:Closed-Loop 100-Channel

201、Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference35 of 78Sequential Stimulation Row by Row Raster scanning Row and column addressing Sc

202、an the image row by row33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference36 of 78Sequential Stimulation Row

203、 by Row20A10A30A Challenge LEDs on the same row could have different currents33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State

204、 Circuits Conference37 of 78Sequential Stimulation Temporal Distribution Challenge LEDs on the same row could have different currents Solution:Current-to-Time Conversion(k1T)IUnit(k2T)IUnit(k1+k2)T IUnit33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-

205、Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference38 of 78 Column selection data received wirelesslySequential StimulationColumn Control33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.

206、02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference39 of 783Iunit A unary I-DAC controls the total current of each row.Sequential StimulationColumn Control33.5:Closed-Loop 100-

207、Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference40 of 78 Rows are scanned using a shift register with a marching 1.Sequential

208、StimulationRow Control33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference41 of 78Sequential Stimulation Temp

209、oral DistributionConventionalTemporal Coding33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference42 of 78Seque

210、ntial Stimulation Poisson DistributionConventionalTemporal CodingPoisson Temporal Coding33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International

211、Solid-State Circuits Conference43 of 78Outline Introduction and Motivation Poisson-Coded Temporally-Distributed Optogenetic Stimulation Analog ED-Based Adaptive-Threshold Spike Detection Measurement Results Comparison and Conclusions33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1

212、.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference44 of 78Challenge 2:Non-Uniform Light SensitivityHigh OpsinConcentrationLow OpsinConcentrationNon-uniform opsin distribution

213、Varying light sensitivity in different regions33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference45 of 78Opt

214、rode Design10 umLEDs33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference46 of 78Optrode Design10 umelectrode3

215、3.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference47 of 78Closed-Loop Calibration33.5:Closed-Loop 100-Channe

216、l Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference48 of 78Closed-Loop Calibration Challenge:Spike detection threshold should adapt to

217、Background noise variations Across the array Over time33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference49

218、of 78Adaptive Spike Detection33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference50 of 78Adaptive Spike Detec

219、tion33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference51 of 78SNR BoostingSensitivity:98.9%Sensitivity:44%L

220、ow SNRHigh SNR33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference52 of 78SNR BoostingSensitivity:98.9%Sensit

221、ivity:44%Low SNRHigh SNRLow SNRHigh SNRSensitivity:99.5%Sensitivity:81%33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circu

222、its Conference53 of 78Adaptive Spike Detection With SNR Boosting(dv/dt)233.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circ

223、uits Conference54 of 78Need for an Analog Implementation To avoid input signal(BW:8kHz)digitization To avoid memory and computation required for rolling average.33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Code

224、d Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference55 of 78Recording Front-End Bandwidth:300 Hz-8 kHz-33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temp

225、orally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference56 of 78SC Differentiators-33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic

226、 Stimulation 2024 IEEE International Solid-State Circuits Conference57 of 78SC Differentiators-33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE Interna

227、tional Solid-State Circuits Conference58 of 78SC Differentiators-33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Co

228、nference59 of 78SC Differentiators-33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference60 of 78SC Differentia

229、tors-33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference61 of 78SC Rolling Averager-33.5:Closed-Loop 100-Cha

230、nnel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference62 of 78Gilbert Cell Squarer V1V2EDout=(V1-V2)(V1-V2)+C-33.5:Closed-Loop 100-Chan

231、nel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference63 of 78Threshold Generation Block Diagram-33.5:Closed-Loop 100-Channel Highly-Sca

232、lable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference64 of 78Threshold Generation Block Diagram-33.5:Closed-Loop 100-Channel Highly-Scalable Retinal

233、Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference65 of 78Threshold Generation-Circuit Implementation33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implan

234、t with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference66 of 78Outline Introduction and Motivation Poisson-Coded Temporally-Distributed Optogenetic Stimulation Analog ED-B

235、ased Adaptive-Threshold Spike Detection Measurement Results Comparison and Conclusions33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International So

236、lid-State Circuits Conference67 of 78System-level Implantation33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Confe

237、rence68 of 78Inductive Links Setup and Results33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference69 of 78Mea

238、surement Results Front-End33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference70 of 78Measurement Results Spi

239、ke Detection 33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference71 of 78Measurement Results Spike Detection

240、33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference72 of 78Measurement Results Closed-Loop Adaptation33.5:Cl

241、osed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference73 of 78Measurement Results Closed-Loop Adaptation33.5:Closed-Lo

242、op 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference74 of 78Measurement Results Closed-Loop AdaptationLightSensitivityDecre

243、ased33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference75 of 78Outline Introduction and Motivation Poisson-C

244、oded Temporally-Distributed Optogenetic Stimulation Analog ED-Based Adaptive-Threshold Spike Detection Measurement Results Comparison and Conclusions33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally

245、 Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference76 of 78Comparison with the SotA33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic

246、Stimulation 2024 IEEE International Solid-State Circuits Conference77 of 78Comparison with the SotA33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Analog ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE Int

247、ernational Solid-State Circuits Conference78 of 78Conclusions A highly-scalable optogenetic stimulator IC was presented 100-channel custom-designed optrode array A Poisson-coded,temporally-distributed stimulation technique was employedFlattened the ICs instantaneous power profileReducing number of i

248、nterconnects.Channel-specific closed-loop stimulation intensity calibration was employedUsing a 1.02uW fully-analog adaptive spike detection circuit.Enables stimulation efficacy optimizationEnhanced channel-count scalability33.5:Closed-Loop 100-Channel Highly-Scalable Retinal Implant with 1.02W Anal

249、og ED-Based Adaptive-Threshold Spike Detection and Poisson-Coded Temporally Distributed Optogenetic Stimulation 2024 IEEE International Solid-State Circuits Conference79 of 78Please Scan to Rate Please Scan to Rate This PaperThis Paper33.6:A Millimetric Batteryless Biosensing and Stimulating Implant

250、with Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference1 of 46A Millimetric Batteryless Biosensing and Stimulating Implant with Magnetoelectric Power Transfer and 0.9pJ/b PWM BackscatterZhanghao Yu*1,Huan-Cheng Liao*,Fatima Alrashdan,Zi

251、yuan Wen,Yiwei Zou,Joshua Woods,Wei Wang,Jacob T.Robinson,Kaiyuan YangRice University,Houston TX 1now at Intel,Santa Clara,CA*Equally Credited Authors33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE Internation

252、al Solid-State Circuits Conference2 of 46Implantable Bioelectronic MedicineDeep Brain Stimulation&ECoG RecordingTreat movement disordersElectromyography(EMG)Muscle StimulationEMG RecordingVagus Nerve StimulationTreat epilepsy and depressionElectrocardiogram(ECG)Electrocorticography(ECoG)Spinal Cord

253、StimulationPain reliefCardiac PacingTreat heart failuresGastrointestinal StimulationModulate stomach mobilityRef:Zou,Fundamental Research21.33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-S

254、tate Circuits Conference3 of 46Millimetric Wireless Batteryless Bio-ImplantsStimulationJ.C.Chen,Nat.BME22D.K.Piech,Nat.BME20Y.Jia,ISSCC18Z.Yu,JSSC22Sensing/RecordingM.M.Ghanbari,JSSC21A.Burton,PNAS20S.Sonmezoglu,Nat.Biotech21C.Shi,Sci.Adv.21A.Singer,Neuron20J.Lee,Nat.Elec.2133.6:A Millimetric Batter

255、yless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference4 of 46Challenges in Wireless Bio-Implant DevelopmentKey ChallengeSafe,robust,and efficient wireless power transferHighly efficient bidirecti

256、onal telemetry with sufficient bandwidth-Downlink:instructions and configuration parameters-Uplink:monitoring and sensing dataUplink DataWireless Power&Downlink Data33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 I

257、EEE International Solid-State Circuits Conference5 of 46Magnetoelectric(ME)Power for mm-Sized ImplantsGood efficiency with mm-scale RXLower sensitivity to misalignmentHigh delivered power under safety limitRef:Z.Yu,TBioCAS20;Z.Yu,JSSC21.Joule EffectLow-Frequency Magnetic FieldStressEpoxyPiezoelectri

258、c:PZT-5Piezoelectric EffectMagnetostrictive:Metglas ME Power&DownlinkVAC33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference6 of 46Bidirectional ME EffectsJoule EffectV

259、illari EffectBackscattered Magnetic FieldLow-Frequency Magnetic FieldStressEpoxyPiezoelectric:PZT-5Piezoelectric EffectMagnetostrictive:Metglas ME Power&DownlinkME UplinkVACRef:Z.Yu,MobiCom22.33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9

260、pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference7 of 46ME Powered and Programmed Wireless ImplantsEndovascular Stimulation of Peripheral Nerves in PigsBrain Stimulation of Freely Moving RatsMultisite StimulationRef:J.C.Chen,Nat.BME22;A.Singer,Neuron20;Z.Yu,JSSC22.33.6:A M

261、illimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference8 of 46Existing Uplink Technologies for ME Implants Hybrid Scheme:ME Power+Inductive BackscatterRef:Z.Yu,RFIC22Low-power bac

262、kscatter uplinkLarger implant size,harder device integrationInterferenceME TXInductiveTRXME PowerInductive BackscatterCoil SoCME Transducer33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-St

263、ate Circuits Conference9 of 46Active ME Uplink:Driving ME TransducerRef:S.Hosur,ISCAS23Uplink leveraging converse ME effect Low impedance of ME transducer Higher power consumptionExisting Uplink Technologies for ME Implants Applied AC VoltageME TransducerME Uplink Signal 2VPPk-Level Impedance33.6:A

264、Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference10 of 46ME Backscatter with Load-Shift-Keying(LSK)Ref:Z.Yu,MobiCom22.Existing Uplink Technologies for ME Implants ME Transd

265、ucerAmp.f1f23kHzBit 0:f1Bit 1:f2fCapacitive LSK-Induced FSK33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference11 of 46ME Backscatter with Load-Shift-Keying(LSK)Ref:Z.Y

266、u,MobiCom22Single ME transducer,lower powerTradeoff:frequency change vs.SNR Data rate limited by excitation and ringdown timeExisting Uplink Technologies for ME Implants ME TransducerAmp.f1f23kHzBit 0:f1Bit 1:f2fCapacitive LSK-Induced FSK33.6:A Millimetric Batteryless Biosensing and Stimulating Impl

267、antwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference12 of 46 Motivation System overview Implant SoC design Measurement results Conclusion Outline33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric

268、 Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference13 of 46This Work:Pulse-Width-Modulated ME BackscatterME ImplantBackscattered Signalt1Applied Magnetic Fieldt3ExcitationRingdownExternalTRXEnergy Extractiont2Goal:achieving high SNR,high data rate ME b

269、ackscatter uplinkLarger amplitude change Better SNREncoding multiple bits in a ringdown Higher data rate33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference14 of 46This

270、 Work:Pulse-Width-Modulated ME BackscatterChallenge:fast signal amplitude change with the high-Q ME transducerSolution:Switched-Capacitor Energy Extraction(SCEE)MagnetostrictiveResponsePiezoelectricResponseZmCPHACME Backscattered SignalEnergyDissipation ME Transducer VoltageHigh QME Transducer Equiv

271、alent ModelEnergy Switched-CapacitorEnergy Extraction 50%reduction in backscattered signals amplitude in 2 ME cycles33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference

272、15 of 46A Wireless Implantable ME Bio-System External TRXME Power TXMEBackscatter RXSupply VoltagesElectrodesME Power&DownlinkPWM ME BackscatterPower ManagementDownlink RXControlSensorsTemp.&VoltageUplink DataDownlink DataControlRecording DataSensing DataControlCoilsBackscatter Uplink TXRecordingME

273、TransducerStimulator33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference16 of 46 External TRXME Power TXMEBackscatter RXSupply VoltagesElectrodesME Power&DownlinkPWM ME

274、 BackscatterPower ManagementDownlink RXControlSensorsTemp.&VoltageUplink DataDownlink DataControlRecording DataSensing DataControlCoilsBackscatter Uplink TXRecordingME TransducerStimulatorA Wireless Implantable ME Bio-SystemWireless ME power,downlink,and PWM backscatter uplink A 6.7mm3implant with a

275、 single ME transducer5x2mm2 ME Transducer6.7mm3SoC33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference17 of 46 Motivation System overview Implant SoC design Measurement

276、 results Conclusion Outline33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference18 of 46Implant SoCs PWM ME Backscatter TXCore:Switched-Capacitor Energy Extractor(SCEE)+

277、Peak Detector(PD)-Extract and dissipate the energy stored in the ME transducerRingdown Detector+Baseband+FIFO:timing&controlMETransducerSwitched-Capacitor Energy Extractor(SCEE)ENSCEEBasebandAsync FIFODATAImplant SoC s TXENPDPeak Detector(PD)Fast Amplitude ReductionSensing DataRingdown DetectorVME33

278、.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference19 of 46Switched-Capacitor Energy Extraction(SCEE)Extract energy from the ME piezoelectric capacitor CpSequentially ac

279、tivated switched-capacitor array-No reverse current sensing-Less sensitive to switch on-resistance-On-chip integrationControl1919.464655129METransducerVoltage.Peak Point DetectionMax.Stored Energy in ME MaterialsENSCEE33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelect

280、ric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference20 of 46Peak Detection for PWM ME BackscatterRequirementAccurate-Maximized energy dissipation for fast amplitude reduction Low power-High-efficiency uplink for implants with a limited power budgetCh

281、allengeWorking in ringdown-ME transducer voltage decaysWoking at a much higher frequency-331kHz in ME vs VRECT Not suitable for ME backscatterRef:D.A.Sanchez,JSSC16;Z.Chen,JSSC17.VINVRECTPeak Detection33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfe

282、r and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference22 of 46Prior-Art Peak DetectorActive DiodeTrack&HoldRequire VIN VRECT Not suitable for ME backscatterPeakDetectionVINVTRACKRequire a wide bandwidth amplifier Large powerRef:D.A.Sanchez,JSSC16;Z.Chen,JSSC17.Ref:M.Di

283、ni,TPEL16.VINVRECTPeak Detection33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference23 of 46Prior-Art Peak DetectorActive DiodeTrack&HoldV-to-I ConversionRequire VIN VR

284、ECT Not suitable for ME backscatterPeakDetectionVINVTRACKRequire a wide bandwidth amplifier Large powerI-ComparatorVINIREFIINPeakDetectionFixed IREFSmall(fF)CIN Sensitive to parasiticRef:D.A.Sanchez,JSSC16;Z.Chen,JSSC17.Ref:M.Dini,TPEL16.Ref:T.Hehn,JSSC12.VINVRECTPeak Detection33.6:A Millimetric Bat

285、teryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference24 of 46This Work:Zero-Crossing Detection+Phase ShiftZero-crossing detection is easier and more accurate than peak detection Elimination o

286、f amplifiers,heavily duty-cycled operation Low power consumption,total average power of 15nWZero-Crossing Detection90 180 METransducer180 360 90 33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International So

287、lid-State Circuits Conference25 of 46Circuit Diagram of Peak Detector(PD)Ringdown DetectorME Cycle MonitorZero-Crossing Detector(ZCD)PZCME CLK RecoveryCLKMEFolded Delay ChainPulseGeneratorRegisterY,YBCounterSTGCYCTDC ResultTDCDelay-LockedPulse GeneratorDigital Phase Shifter(DPS)ENDPSControllerCLKRDT

288、o SCEEPDPPP,PNP,ENSCEEUplink DataControlENZCD33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference26 of 46Ringdown DetectorME Cycle MonitorZero-Crossing Detector(ZCD)PZC

289、ME CLK RecoveryCLKMEFolded Delay ChainPulseGeneratorRegisterY,YBCounterSTGCYCTDC ResultTDCDelay-LockedPulse GeneratorDigital Phase Shifter(DPS)ENDPSControllerCLKRDTo SCEEPDPPP,PNP,ENSCEEUplink DataControlENZCDOperation Scheme of Peak Detector(PD)ME Transducer VoltageCalibrationPhase shifter works as

290、 TDCData TransmissionPhase shifter works as delay-locked pulse generatorTMEME Transducer VoltageZero-Crossing Pulse,PZCTME/4TME/2TMEME Transducer VoltageZero-Crossing Pulse,PZCENSCEETDC ResultSTG,CYC33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer

291、and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference27 of 46Ringdown DetectorME Cycle MonitorZero-Crossing Detector(ZCD)PZCME CLK RecoveryCLKMEFolded Delay ChainPulseGeneratorRegisterY,YBCounterSTGCYCTDC ResultTDCDelay-LockedPulse GeneratorDigital Phase Shifter(DPS)END

292、PSControllerCLKRDTo SCEEPDPPP,PNP,ENSCEEUplink DataControlENZCDZero-Crossing Detector with Current CompensationAdd offset current I1and I2to minimize delay without large biasingI2increases with ME cycles in the ringdown Smaller and more consistent delayWith I1 and I2 04080Simulated Delay(ns)Only Wit

293、h I1 Original1st Cycle 30th CycleVME2VME1nAI1I2ME CycleEnvelop ExtractorENZCDOUTI2I1VME1DL DataArrayArray12033.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference28 of 46

294、Digital Phase Shifter(DPS)Circuit DesignDDDQVDDSTGSTGYBYDQSTGSTGYBYDDQSTGSTGYBYDQSTGYBYCycle CounterYYBSTARTCYCSTARTBYBRegisterRSTRSTRSTRSTCYC,STG ENCALIDCYC,DSTGCYC,STGPulse GeneratorPPP,PNP,ENSCEE33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer a

295、nd 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference29 of 46DDDQVDDSTGSTGYBYDQSTGSTGYBYDDQSTGSTGYBYDQSTGYBYCycle CounterYYBSTARTCYCSTARTBYBRegisterRSTRSTRSTRSTCYC,STG ENCALIDCYC,DSTGCYC,STGPulse GeneratorPPP,PNP,ENSCEEDigital Phase Shifter(DPS)Circuit DesignCore:Differe

296、ntial 4-Stage Folded Delay Chain15ns delay step(i.e.,0.5%of TME)33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference30 of 46DDDQVDDSTGSTGYBYDQSTGSTGYBYDDQSTGSTGYBYDQSTG

297、YBYCycle CounterYYBSTARTCYCSTARTBYBRegisterRSTRSTRSTRSTCYC,STG ENCALIDCYC,DSTGCYC,STGPulse GeneratorPPP,PNP,ENSCEEPulse Generation with Shift-Only ArithmeticLow computing overhead,low powerMax.error 0.75tdelaytdelay:delay of a delay chain stagePositive Peak Point=Zero Point+TME/4Negative Peak Point=

298、Zero Point+TME/4+TME/2For example:TME/4=8decDCYC+decDCYC,DCYC,DSTG tdelay33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference31 of 46Implants Neural Recording FrontendA

299、C-coupled LNA for a high input dynamic range A 3rd-order NS SAR ADC with 1kHz bandwidth and an OSR of 8.8b 2kSa/s continuous data streamingRef:T.H.Wang,JSSC21.SAR Logic3-Stage CIC FilterDelta Modulation14b8bCDACRPSEUDOVinp(10+1)bRPSEUDOVinnVCMVAMP,NVAMP,PVCMVAMP,PVAMP,NVCMLNADigital ProcessVCMVAMP,P

300、VAMP,NVREFPVREFNVCMVREFPVREFNNoise-Shaping SAR ADCTo UplinkTX33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference32 of 46 Motivation System overview Implant SoC design

301、Measurement results Conclusion OutlineRecordingEnergy Extraction Switched CapacitorTesting PadsUplinkFunctional PadsControlPower Management,Downlink,&Capacitor2mm1.3mmImplant SoC Micrograph33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/

302、b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference33 of 46Measured Waveform of Implants Wireless OperationImplant is magnetoelectrically powered and programmed to perform different tasksTested ex-vivo with a 2cm porcine tissuePPPME Transducer Output Time-Domain Downlink(92 bit

303、)Stimulation2.5V,1.5msUplink2VME Backscatter RXImplantME Power TXTRX Coils(under Porcine Tissue)33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference34 of 46Measured Wav

304、eform of ME Energy ExtractionSCEE extracts the ME transducers energy at different time points 18ns(i.e.,0.6%of TME)detection error in the first cycleME Transducer Output CLKRDEnergy Extraction 3.021s3.039sPPPME Transducer Output PPP33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwit

305、h Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference35 of 46Portable External ME TRX DesignBattery-powered ME Backscatter RX:AFE+MCUDifferential RX coil for common-mode interference cancellationVGAMCULPFLDOsADCIASuppliesBatteriesME Powe

306、r TXME Backscatter RXControlDifferentialRX CoilTX CoilUplinkData33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference36 of 46Measured Waveform of PWM ME Backscatter Upli

307、nkDistinguishable pulse widths in ringdown,modulated by 3bit uplink data 17.73kb/s data rate with a 331kHz carrier and 0.9 pJ/bit efficiencyImplant Uplink Data(output from implant)ME Transducer Output 101011111101001011000010000ME Backscattered Signal(amplified)Data Rate:17.73kbps33.6:A Millimetric

308、Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference37 of 46Measured Waveform of PWM ME Backscatter Uplink50%amplitude reduction within 2 ME cycles at distinct time pointsME Backscattered

309、 Signal(amplified)No amplitude change for data 0001V35s33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference38 of 46Measured PWM ME Backscatter SNRSNR=VBACKSCATTER,DROP/

310、NoiseSNR 10.9dB at up to 5cm distance10203040SNR(dB)0010100111001011101111 cm2 cm3 cm4 cm5 cm 3-Bit Uplink Data33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference39 of

311、 46Demodulation Schemes for PWM ME BackscatterScheme 1:Drop DetectionScheme 2:Multilayer Perceptron(MLP)VDROP threshold300 100258Decoded OutputBackscattered Signal.Input,Activation,and Weight:5 bit,Number of Parameter:3260033.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magneto

312、electric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference40 of 46Measured PWM ME Backscatter BER8.5E-5 BER at 5cm with MLP data demodulation123451E 1E 1E 1E 1E 1E BERExternal TRX-Implant Distance(cm)Demodulation:Drop DetectionDemodulation:MLP33.6:A M

313、illimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference41 of 46103050 0100Amplitude(V)Time(sec)02040Wirelessly Recorded LFPPre-recorded LFP(input to implant)333435360100-100Over-t

314、he-Air CommunicationSystems Wireless Neural Recording MeasurementPre-recorded local field potentials(LFP)from rats Wirelessly received signal via ME backscatter closely matches the ground truth33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.

315、9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference42 of 46In-Vitro Test of Wireless Neural RecordingElectrodesME TransducerWireless OperationPlace an encapsulated untethered implant in PBSPower,data,and LFP signals are wirelessly transmitted through PBS00.511.50Amplitude(V

316、)Time(sec)500-500233.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference43 of 46Comparison with State-of-the-Art Wireless ImplantsThis WorkZ.YuMobiCom22J.LimVLSI21J.LeeNa

317、t.Elec.21Y.JiaISSCC20S.SonmezogluISSCC20M.GhanbariJSSC19Stimulation,RecordingStimulationRecordingRecordingStimulation,RecordingO2 SensingRecording180180180653506565SourceMEMEPhotodiodeRFInductiveUltrasoundUltrasoundTransfer Efficiency0.37%(2cm)0.2%(2cm)N/A0.08%(0.8cm)N/AN/A0.06%(1.8cm)ModalityMEMEPh

318、otodiodeRFInductiveModulationTimeTimePWMPWMOOKModalityMEMELEDRFRFUltrasoundUltrasound fcarrier(kHz)331335N/A90000043300020001780ModulationPWM BackscatterLSK BackscatterPGMLSK BackscatterOOKLSK BackscatterAM BackscatterData Rate(kbps)17.731.60.31000067806035Data Rate/fcarrier 0.0540.0047N/A0.0110.015

319、60.030.0197Efficiency(pJ/Bit)0.9N/A253N/A1342.255603BER8.5E-5(5cm)9E-4(1.5cm)N/A5E-3(0.8cm)N/A9E-5(5 cm)N/A52N/A0.8N/A55Downlink DataN/AN/AUplinkDataMax Distance(cm)BiomedicalFunctionSoC Technology(nm)Power33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Tr

320、ansfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference44 of 46Comparison with State-of-the-Art Wireless ImplantsThis WorkZ.YuMobiCom22J.LimVLSI21J.LeeNat.Elec.21Y.JiaISSCC20S.SonmezogluISSCC20M.GhanbariJSSC19Stimulation,RecordingStimulationRecordingRecordingStimul

321、ation,RecordingO2 SensingRecording180180180653506565SourceMEMEPhotodiodeRFInductiveUltrasoundUltrasoundTransfer Efficiency0.37%(2cm)0.2%(2cm)N/A0.08%(0.8cm)N/AN/A0.06%(1.8cm)ModalityMEMEPhotodiodeRFInductiveModulationTimeTimePWMPWMOOKModalityMEMELEDRFRFUltrasoundUltrasound fcarrier(kHz)331335N/A9000

322、0043300020001780ModulationPWM BackscatterLSK BackscatterPGMLSK BackscatterOOKLSK BackscatterAM BackscatterData Rate(kbps)17.731.60.31000067806035Data Rate/fcarrier 0.0540.0047N/A0.0110.01560.030.0197Efficiency(pJ/Bit)0.9N/A253N/A1342.255603BER8.5E-5(5cm)9E-4(1.5cm)N/A5E-3(0.8cm)N/A9E-5(5 cm)N/A52N/A

323、0.8N/A55Downlink DataN/AN/AUplinkDataMax Distance(cm)BiomedicalFunctionSoC Technology(nm)Power33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference45 of 46ConclusionA wi

324、reless bio-system including a 6.7mm3ME implant and a custom ME TRX for biosensing and stimulation-Wireless power,downlink,and PWM backscatter uplink through ME effectsPWM ME backscatter uplink telemetry-Switched-capacitor energy extraction reduces the backscattered signal amplitude by 50%within 2 ME

325、 cycles-17.73kb/s data rate and 0.9pJ/b efficiency at 331kHz carrier-Reliable wireless operation at up to 5cm distance,with 8.5E-5 BER using MLP demodulation33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE Inte

326、rnational Solid-State Circuits Conference46 of 46AcknowledgementThis work is supported in part by the National Science Foundation(NSF)ASCENT and CAREER programs(2023849 and 2146476).The authors would like to thank Yonghee Chang,Yan He,Zhiyu Chen,Wonjune Kim,Matthew Parker,and Yumin Su,for technical

327、discussions and support.33.6:A Millimetric Batteryless Biosensing and Stimulating Implantwith Magnetoelectric Power Transfer and 0.9pJ/b PWM Backscatter 2024 IEEE International Solid-State Circuits Conference47 of 46Please Scan to Rate Please Scan to Rate This PaperThis Paper33.7:An Adhesive Interpo

328、ser-Based Reconfigurable Multi-Sensor Patch Interface with On-Chip Application-Tunable Time-Domain Feature Extraction 2024 IEEE International Solid-State Circuits Conference1 of 61An Adhesive Interposer-Based Reconfigurable Multi-Sensor Patch Interface with On-Chip Application-Tunable Time-Domain Fe

329、ature ExtractionUlsan National Institute of Science and Technology,Ulsan,KoreaJeonghoon Cho*,You Jang Pyeon*,Junyeong Yeom*,Hyunjoong Kim*,Sanghyeon Cho,Yonggi Kim,Taejung Kim,Jong-Hyun Kwak,Geonjun Choi,Yoonsik Lee,Heungjoo Shin,Hoon Eui Jeong,Jae Joon Kim*Equally Credited Authors(ECAs)33.7:An Adhe

330、sive Interposer-Based Reconfigurable Multi-Sensor Patch Interface with On-Chip Application-Tunable Time-Domain Feature Extraction 2024 IEEE International Solid-State Circuits Conference2 of 61Outline Introduction and Motivation Application-Tunable Analog AI Processor Multi-Sensor Analog Front-Ends R

331、econfigurable Patch Interface Implementation Comparison and Summary33.7:An Adhesive Interposer-Based Reconfigurable Multi-Sensor Patch Interface with On-Chip Application-Tunable Time-Domain Feature Extraction 2024 IEEE International Solid-State Circuits Conference3 of 61Environmental Impacts on Huma

332、n Body Poor air quality causes cardiovascular/pulmonary diseasesWHO20Gas LeaksFireAccidentsAir Pollutions13 million deaths/yearworldwide33.7:An Adhesive Interposer-Based Reconfigurable Multi-Sensor Patch Interface with On-Chip Application-Tunable Time-Domain Feature Extraction 2024 IEEE Internationa

333、l Solid-State Circuits Conference4 of 61Heterogeneous Classification System Need for long-term to short-term env./bio.detection Alert/diagnostic system beyond simple signal recordingGas LeaksFireAccidentsAir PollutionsSlow ReactionsHazardous Situation alert+DiseasePrediction+Edge&Real time DevicePPGBioZECGFast Reactions33.7:An Adhesive Interposer-Based Reconfigurable Multi-Sensor Patch Interface w