eetop.cn_T8 - Intelligent Neural Interfaces Fundamentals and Future Directions.pdf

1、 2025 IEEE International Solid-State Circuits ConferenceISSCC 2025 Tutorial 8Intelligent Neural Interfaces:Fundamentals and Future DirectionsProf.Mahsa ShoaranEPFL,Switzerlandmahsa.shoaranepfl.chFebruary 16,20251 of 86Mahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future Direct

2、ions 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future DirectionsOutline2 of 86 Introduction and motivationsBasics of open-loop and closed-loop stimulationMajor trend:higher-channel-count systems Intelligent SoCs for sy

3、mptom control:seizure detectionRecent trends:Scalability,adaptability,and data efficiency Intelligent SoCs for brain-machine interfaces Conclusion and future directions Relevant papers in ISSCC 2025 2025 IEEE International Solid-State Circuits ConferencePrevalence of Major Brain DisordersMahsa Shoar

4、anISSCC T8 Intelligent Neural InterfacesSource:WHO3 of 86Over 1 in 3 people are affected by neurological conditions The leading cause of illness and disability worldwide Treatment resistance affects 2060%of psychiatric patients,30%in epilepsyUrgent need for effective treatment strategiesIncreasing i

5、nterest in neural interfaces and neuromodulation therapies 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranIntroduction:Neural InterfacesTypes of neural signals:Non-invasive recordings:Electroencephalography(EEG)0.5100HzInvasive recordings:Action Potential or Spike 0.2510kHzElect

6、rocorticography(ECoG)1500HzLocal Field Potential(LFP)1500Hz4 of 86Depth Leads(LFP)Neural Interface ChipCortical Grids(ECoG)Microelectrodes(Spike)Bidirectional System:Sensing+Stimulation(neuromodulation)ScalpEEGISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future Directions 2025 IEEE Intern

7、ational Solid-State Circuits ConferenceMahsa ShoaranNeuromodulation DevicesDeep-Brain Stimulation(DBS)Medtronic Percept PC(w BrainSense)NeuroPace RNS AspireSR VNS(closed-loop)(w cardiac sense)AlphaDBSDyNeuMoCorTecDevices for Movement DisordersDevices for Epilepsy(investigational,closed-loop)5 of 86I

8、SSCC T8 Intelligent Neural InterfacesSource:Medtronic,NeuroPace,Livanova,Newronika,Amber,CorTec NeuroElectrodes 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces.Source:D.B.Rubin et al.,Neurology23,BrainGateBrain-Machine Interfaces(BMIs)Ultima

9、te Goal:Restoring Movement and Communication after Paralysis(A)The Utah microelectrode array(Blackrock Neurotech):1010 array on a 44-mm platform(B)The array is connected through a bundle of wires to a percutaneous pedestal(C)Signals from the pedestal are transmitted through a cable or wirelessly to

10、external devices for processing and decoding6 of 86 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Source:https:/www.vis.caltech.edu/Chen_Center,YoutubeCurrent Systems are Effective but DBS for Tremor SuppressionBMI for Prosthetic Control 7

11、 of 86 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future DirectionsLimitations of Current SystemsExisting neuromodulation devices:Low channel count(48)Limited embedded processingFew/simple biomarkers,simple detection al

12、gorithmsLeading to false detections,missed events,high latencyLimited application(epilepsy,movement disorders)Existing BMIs:Bulky,tetheredLow channel count(100)Limited performance(i.e.,suitable for simple tasks only)8 of 86 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8

13、 Intelligent Neural InterfacesSource:MedtronicOpen-Loop StimulationStimulationPrimarily programmed and updated by physiciansContinuous pulse delivery regardless of the ongoing brain activityDrawbacks:High energy consumption and potential side effects from stimulationPhysician-Programmed ParametersEx

14、ample:DBS9 of 86 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:NeuroPace,J.Yoo and M.Shoaran,Curr.Op.Biotech21Closed-Loop StimulationStimulationAmplifiers,ADCsFeedback from neural signals to control stimulationOn-demand pulse deliver

15、y responsive to ongoing brain activityAdvantages:Low energy consumption,less side effects,(often)higher efficacySymptom Detection Detection AlgorithmBrain Activity MonitoringBiomarkerExtractionExample:RNS(Responsive Neurostimulator)10 of 86Model Training 2025 IEEE International Solid-State Circuits

16、ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future DirectionsClosed-Loop Systems:Off-vs.On-Chip DetectionExternal Detection11 of 86StimulationAmplifiers,ADCs 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:

17、Fundamentals and Future DirectionsClosed-Loop Systems:Off-vs.On-Chip DetectionOn-Chip Detection:Low latencyLow transmission power Better securityRequires efficient on-chip processing12 of 86Applications:Neurological symptom detection&control:Epilepsy,Parkinsons disease,depression,Sensory feedback,re

18、storing movement,and rehabilitation StimulationAmplifiers,ADCsDetection AlgorithmBiomarkerExtraction 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:N.A.Steinmetz et al.,Science20,Precision Neuroscience,Paradromics,NeuralinkTrend:Highe

19、r-Channel-Count SystemsNeuropixels Probe,960-1280Ch.Precision Neuroscience,1024Ch.Neuralink N1,1024Ch.Paradromics Connexus,1684Ch.Recent systems integrate 1000+channels for recording neural signals from the cortical surface or deeper brain layersLow-power,area-efficient ASICs are crucial to support

20、such high densitiesBottleneck:The data rates exceed the capacity of low-power wireless links Efficient on-site processing13 of 86Translational Systems for Potential Human Use Research Tool 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSourc

21、e:T.Kaiju et al.,Front.Neural Circuits17,Medtronic Why More Channels?EpilepsyAccurate detection of seizure onsetPrecise localization of seizure fociMovement Disorders Reduced stimulation side effectsImproved symptom detectionBrain-Machine Interfaces(BMIs)Precise motor decodingMore natural prosthetic

22、 controlTime(ms)Channel CountDecoding AccuracyHigh-Density ECoGMulti-Lead Stereo-EEG14 of 86Deep Brain StimulationFinger Movement DecodingSeizure Detection 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future DirectionsOut

23、line15 of 86 Introduction and motivationsBasics of open-loop and closed-loop stimulationMajor trend:higher-channel-count systems Intelligent SoCs for symptom control:seizure detectionRecent trends:Scalability,adaptability,and data efficiency Intelligent SoCs for brain-machine interfaces Conclusion a

24、nd future directions Relevant papers in ISSCC 2025 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future DirectionsEpilepsy$15.5 BILLION50 MILLION1/3 DRUG-RESISTANT4thMOST COMMON ONE IN CHILDRENSeizure-free with medications

25、Partially controlled by surgeryNeed alternative therapies TP Not suitable for imbalanced datasetsF1 Score:Recommended metrics:Sensitivity(both window-based and event-based),False Alarm Rate,and F1 Score We should avoid metrics directly dependent on True Negatives,such as accuracyOther Performance Me

26、trics for Seizure DetectionMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:J.Dan et al.,Epilepsia2420 of 86 2025 IEEE International Solid-State Circuits ConferenceDistinct from system latency!Detection latency is clinically more relevant than system latencyLow detection latency is critical

27、 for effectively closing the loopDetection LatencyMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future DirectionsLate detection(positive latency)Expert-marked onsetEarly detection(negative latency)21 of 86InterictalIctal 2025 IEEE International Solid-State Circuits ConferenceM

28、ahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:NeuroPaceSeizure Detection in EpilepsyPrimary goalAchieve high sensitivity in detecting seizure onsetMinimize the false alarm rateReduce detection latencyRecent systems are aiming to enhanceEnergy efficiencyScalability Data efficiencyAdaptabil

29、ity and versatility22 of 86 2025 IEEE International Solid-State Circuits ConferenceSeizure Biomarkers(Features)Mahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:U.Shin et al.,ISSCC22,JSSC22 Line-Length23 of 86 2025 IEEE International Solid-State Circuits ConferenceSeizure Biomarkers(Features

30、)Mahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:U.Shin et al.,ISSCC22,JSSC22 Line-LengthSpectral Energy in non-overlapping bandsDelta 14HzTheta 48HzAlpha 813HzBeta 1330HzGamma(Low,High)30150HzHigh-Frequency Oscillations(HFO)150500HzEEG24 of 86bandpass-filtered signal ECoG or intracranial

31、EEG(iEEG)2025 IEEE International Solid-State Circuits ConferenceSeizure Biomarkers(Features)Mahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:U.Shin et al.,ISSCC22,JSSC22 Line-LengthSpectral Energy in non-overlapping bandsDelta 14HzTheta 48HzAlpha 813HzBeta 1330HzGamma(Low,High)30150HzHigh-F

32、requency Oscillations(HFO)150500HzPhase-Domain FeaturesMeasures of cross-regional synchrony between neuronal oscillationsEEG25 of 86bandpass-filtered signal ECoG or intracranial EEG(iEEG)2025 IEEE International Solid-State Circuits ConferenceThresholdResponsive Neurostimulator(RNS)for EpilepsyMahsa

33、ShoaranISSCC T8 Intelligent Neural InterfacesSource:NeuroPace26 of 86Seizure detection by thresholding selected features:Line-Length,Half-Wave,AreaMeasures of frequency and amplitude change,and oscillations in selected frequency bandsPros and Cons:Simple detection algorithmsInability to integrate in

34、formation from multiple channels and features Low sensitivity and high false alarm rate,high latencyMachine learning(ML)can address these issues and enhance performanceLine-Length 2025 IEEE International Solid-State Circuits ConferenceVariations across and within Individual PatientsMahsa ShoaranISSC

35、C T8 Intelligent Neural InterfacesSource:M.A.B.Altaf et al.,TBioCAS16Seizure patterns vary across patients,even in a single patient!Seizure patterns vary for different seizure types(e.g.,focal vs.generalized seizures)Seizure patterns in an adult patient(left)and an infant(right)27 of 86 2025 IEEE In

36、ternational Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:https:/ et al.,ICML24The Importance of Reliable Benchmarking Seizures are rare:Challenge in training ML algorithms on highly imbalanced dataData split and validation procedures are critical:Data leak

37、age and shuffling are not allowedEssential to report:Exact data used and excludedASIC designers should use large,public EEG and intracranial EEG datasets widely benchmarked by the ML community:Extensive validation across diverse datasetsCHB-MIT dataset:24 children only,few seizures,relatively easy t

38、o classifyTUH dataset:622 adults,more seizure types,also publicly accessibleEU Epilepsia:Large,but not publicly accessible;iEEG.org:from diverse hospitalsAlgorithm code sharing:Essential to enable reproducibility of results and fair benchmarking28 of 86ISSCC24 2025 IEEE International Solid-State Cir

39、cuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:IEEE XploreIntelligent SoCs for EpilepsyISSCC15 ISSCC20ISSCC23ISSCC22VLSI Symp21ISSCC24ISSCC22ISSCC20ISSCC21ISSCC1829 of 86 2025 IEEE International Solid-State Circuits ConferenceEnergy efficiencyPower consumption divided by s

40、ampling rate(energy/class.)Reporting both total power consumption and power per channel for multi-channel systems is also essential for evaluating design efficiency and scalabilityArea:Total area and area per channelSystem(SoC)vs.digital back-end(DBE):distinction between the two is necessaryHardware

41、 Evaluation MetricsMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:C.Ding et al.,ESSERC2430 of 86TXADCUplink DSPCH1CHMFeature Extra./DecodingSTIMAFELNAMUXCH1CHN(T/FDMA in UWB)ECoG GridUtah ArrayNeuropixels ProbeNearby Base-Station/ReceiverTransmission Range:0.6mClosed-Loop Neural ImplantClosed-L

42、oop Neural ImplantElectrodesElectronicsRXDownlink(FDD with Uplink)Feature Extraction&ClassificationDBEUplinkTXSystem(SoC)2025 IEEE International Solid-State Circuits ConferenceEarly ML SoCs:Support Vector Machines(SVMs)Mahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:https:/scikit-learn.org

43、/Linear SVM:Finds the hyperplane that maximizes the margin between classesKernel Trick:Maps non-linear data to a higher-dimensional space using a feature map(),making it linearly separable 31 of 86 2025 IEEE International Solid-State Circuits ConferenceSeizure Detection with Linear SVMMahsa ShoaranI

44、SSCC T8 Intelligent Neural InterfacesSource:J.Yoo et al.,JSSC12 Linear SVM:Simplicity,ease of modellingLow accuracy for non-linear decision boundaries32 of 86X:The incoming feature vectorW and:Model parameters(patient-specific)2025 IEEE International Solid-State Circuits ConferenceSVM learns based o

45、n similarity of input data points using the kernel trickRadial Basis Function(RBF)kernel:(,)=exp(|)Non-linear SVM:Higher accuracyNon-linear kernels are complex:power-and area-inefficientMahsa ShoaranSeizure Detection with Non-Linear SVMRadial Basis Function(RBF)ISSCC T8 Intelligent Neural Interfaces

46、Source:M.A.B.Altaf et al.,ISSCC13,TBioCAS16,https:/ of 86 2025 IEEE International Solid-State Circuits ConferenceEnergy-Efficient Two-Level Seizure DetectionMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:Y.Wang et al.,ISSCC20,TBioCAS21 Two-level detection:A coarse classifier triggers a fi

47、ne classifier when a seizure is suspectedCoarse Classifier:Multi-level feature thresholding with high sensitivity&low specificity Fine Classifier:Discrete wavelet transform(DWT)and SVM Off for more than 80%of the time(CHB-MIT dataset)34 of 86SVMDWT feature extraction 2025 IEEE International Solid-St

48、ate Circuits ConferenceSVM-based Seizure DetectorsMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:IEEE XploreISSCC15 ISSCC20ISSCC18ISSCC12ISSCC13 Pros:Ease of modelling,reproducibility of results,and robustnessThe number of support vectors(SVs)depends on separability of the featuresHardwar

49、e complexity increases for greater numbers of SVsSVM requires all features from all channels:not scalable(typically 816 channels)Memory-demanding64kB96kB64kB96kB64kB35 of 86 2025 IEEE International Solid-State Circuits ConferenceInterim Q&A Session Mahsa ShoaranISSCC T8 Intelligent Neural Interfaces

50、:Fundamentals and Future Directions36 of 86 2025 IEEE International Solid-State Circuits ConferenceYesNoYesNoHigher Efficiency:Decision Trees(DTs)Mahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:M.Shoaran et al.,JETCAS18DTs:Computationally simpleMain processing unit:ComparatorsEnable energy

51、-efficient classificationDTs are non-linear modelsHigh accuracy can be achieved through ensemble techniquesCompatible with single-path sequential processingOnly one tree path is traversed per decisionReduce computational complexity37 of 86YesNoYesNoNoYesDecisionXi:Input featureTi:Threshold value 202

52、5 IEEE International Solid-State Circuits ConferenceTree Ensembles:Bagging and BoostingMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future DirectionsTree ensemble:A learning algorithm composed of a set of base learners(i.e.,DTs)Bagging:Trains multiple trees in parallel on ran

53、dom samples of the data and averages their predictions to reduce variance and improve accuracy;Example:Random Forest Boosting:Sequentially builds decision trees,each correcting errors of previous ones,to reduce bias and improve accuracy;Examples:Gradient Boosting,AdaBoost38 of 86 2025 IEEE Internati

54、onal Solid-State Circuits ConferenceTree Ensemble with Gradient BoostingMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Source:M.Shoaran et al.,JETCAS18,https:/xgboost.readthedocs.io/Goal:Minimize the total loss,i.e.,the error between true values(yi)and predictions(H(xi)By successively training

55、new DTs to approximate gradients of the loss function and correct the errors of the current model39 of 86x:Input feature vector ht(x):Output of the t-th treet:Weight of the t-th treeObjective function Loss function 2025 IEEE International Solid-State Circuits ConferenceTree Ensemble with Gradient Bo

56、ostingMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Source:M.Shoaran et al.,JETCAS18Each DT:Single-path inferenceSequential processing with a single feature extraction engine(FEE)Enhanced efficiency in seizure detection:DBE energy efficiency of 41.2 nJ/class.,1mm2area,32-Channel,1kB of memoryT

57、he tree count increases for more complex tasks40 of 861kBDT1DTkk10CH1CHNh1hkDecisionTrainingdepth 4next featureA/DMUXInput Control Feature Ext.&Comp.1mm2 2025 IEEE International Solid-State Circuits ConferenceSingle NeuralTree ClassifierMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:U.Shi

58、n et al.,ISSCC22,JSSC22,B.Zhu et al.,TBioCAS20A probabilistic oblique tree,combining the benefits of neural networks and DTsOblique decision boundary defined by weighted combinations of input featuresSigmoid function()enables soft,probabilistic routing scheme during training41 of 86fi:The i-th input

59、 featurewi:Weight of feature fiTH.:Threshold valueConventionalObliqueA Tree-structured Neural Network 2025 IEEE International Solid-State Circuits ConferenceNeuralTree Classifier:Model CompressionMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:U.Shin et al.,ISSCC22,JSSC22Soft decision boun

60、daries are compatible with gradient-based optimizationEnabling model compression techniques such as weight pruning and quantizationTH.2025 IEEE International Solid-State Circuits ConferenceEfficient Feature Extraction TechniquesMahsa ShoaranISSCC T8 Intelligent Neural Interfaces Source:U.Shin et al.

61、,ISSCC22,JSSC22,CICC22 -Norm AmplitudeZ +01PAC/PLVLUT01PAC/PLV01Z +32Z +32 -Norm AmplitudesincosTDM Phase Feature ExtractorLAA PhasePACPLVReIm|01Z +64Z +64+|Z +64Z +64+|Z +64Leading-Zero DetectorReciprocal LUTRatio CalculatorDenom.Numer.FTYPEADC/BPF OutSELMPLLHFO RatioHjorthLMP/SETDM Spectral&Tempor

62、al Feature Extractor32-TAP BUFFER32-TAP BUFFER32-Tap BPF FIFO6432-TAP BUFFER32-TAP BUFFER31-Tap HT FIFO64Data MUXBandpass&Hilbert TDM FIR Filter+C0+C1+C14+C15+DnDn Dn Dn Dn Dn Dn Dn BPF OutADC OutFIR CoefficientsHardware sharing among featuresEnd-to-end time-division multiplexing(TDM)Efficient featu

63、re approximations 44 of 86Use the absolute value of the filtered output to avoid square operation 2025 IEEE International Solid-State Circuits ConferenceNeuralTree:Energy-Aware Learning Mahsa ShoaranISSCC T8 Intelligent Neural Interfaces Source:U.Shin et al.,ISSCC22,JSSC22Introducing a power-depende

64、nt regularization term in the objective functionGoal:To penalize power-demanding features Energy-aware model:64%power saving in feature extractionEnergy-AwareLowHighTemporalSpectralPhase16.6%28.4%55.0%81.8%16.6%1.6%ConventionalFeature ComplexityFeature Distribution 45 of 86energy:Power consumed by f

65、eature extractionC:A hyperparameter controlling energy-accuracy trade-offEnergy-aware objective function:(Higher Power)2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces Source:U.Shin et al.,ISSCC22,JSSC22State-of-the-art SoC energy efficiency

66、of 0.227J/class.46 of 86NeuralTree:Chip Micrograph,SoC Output256-Channel sensing16-Channel stimulationActive area:3.48mm2Total power:453W StimulationSensingFront-EndFEE&ML4mm2mmiEEG.orgCHB-MIT EEG Annotated Seizure 2025 IEEE International Solid-State Circuits ConferenceDecision Tree-based SoCsMahsa

67、ShoaranISSCC T8 Intelligent Neural InterfacesSource:IEEE XploreDecision trees are compact,low-power,and memory-efficientScalable:Current designs integrate 8256 channelsISSCC20ISSCC22ISSCC24JETCAS181kB3.2kB47 of 86120kB8kB54 of 86 2025 IEEE International Solid-State Circuits Conference Introduction a

68、nd motivationsBasics of open-loop and closed-loop stimulationMajor trend:higher-channel-count systems Intelligent SoCs for symptom control:seizure detectionRecent trends:Scalability,adaptability,and data efficiency Intelligent SoCs for brain-machine interfaces Conclusion and future directions Releva

69、nt papers in ISSCC 2025Mahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future DirectionsOutline55 of 86 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:U.Shin et al.,ISSCC22,JSSC22 Training:256-ChannelHigh-density

70、recording with high spatial resolution for model training:256-channelRelevant channels can be learned through model training56 of 86464-Ch.TDM Front-EndChannel ImportanceChannel IndexChannel importance in seizure detectionScalability with Channel Count 2025 IEEE International Solid-State Circuits Co

71、nferenceHigh-Density Training,Channel-Selective InferenceMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:U.Shin et al.,ISSCC22,JSSC22 Class Labels Channel-selective inference(64 channels per node)Channels dynamically activated along NeuralTrees decision pathAdvantage:Lower power during inf

72、erence57 of 86Inference:64-Channel164-Ch.TDM Front-EndTrained Sequence 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:IEEE XploreVersatility and ReconfigurabilityBioAIP:Programmability in FIR tap count,window size,activation functions

73、Seizure detection,arrythmia,hand gestureNeuralTree and NURIP SoCs:Versatile biomarker extraction NeuralTree:Seizure and tremor detection,finger movement classification Versatility:Area and memory overhead,may impact energy efficiency73kBISSCC21(BioAIP)ISSCC22(NeuralTree)3.2kBISSCC18(NURIP)96kB 58 of

74、 86 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces Source:M.Zhang et al.,VLSI Symp21,C.Tsai et al.,ISSCC23,TBioCAS23,B.Zhu et al.,IEEE NER21 Enhancing Data EfficiencyConventional patient-specific training requires extensive labelleddata fro

75、m each patientFails to detect symptoms in new patients with limited recordingsOne-shot learning:Trained with limited seizure data,can incorporate new data via online tuning for effective seizure trackingZero-shot-retraining:Fully eliminate data collection and analysis from target patientInitially tr

76、ained offline with a pre-existing EEG databaseTuned online for fine calibration in an unseen patient59 of 86ISSCC23VLSI Symp21 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces Source:A.Chua et al.,VLSI Symp21,JSSC22Online AdaptabilitySeizure

77、patterns can vary between patients and may also change over timeLogistic regression with unsupervised stochastic gradient descent-based weight update Utilizing the classifiers output probability as the label:rounded off to 0 or 1A series of high-confidence predictions trigger the online weight updat

78、eSimple&memory-efficient:96%success rate in 1-and 2-D finger movement control68 of 86 2025 IEEE International Solid-State Circuits ConferenceConventional BMIs:Point-and-Click ControlMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:C.Pandarinath et al.,elife1769 of 86Conventional BMIs forcom

79、munication recovery:Point-and-click control of a cursorAverage typing rate:24.4 3.3 correct characters per minute 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:F.R.Willett et al.,Nat21,S.L.Metzger et al.,Nat23BMIs for Rapid Communica

80、tion RecoveryHandwriting296-Ch.UtahRNN model90 char./min94.1%(31 char.)Nature21Speech253-Ch.ECoGRNN+language model78 words/min91.8%(119 words)74.5%(1024 words)Nature23Decoding intricate motor behaviours,rather than point-by-point movementsAdvantage:Faster communication speed than traditional BMIs70

81、of 86Handwriting decoder Nat21:Recurrent neural network(RNN)with 7M parameters 2025 IEEE International Solid-State Circuits ConferenceOn-Chip MiBMI with brain-inspired Distinctive Neural Codes(DNCs)Significant dimensionality reduction,320 reduction in#of MACs Low-complexity decoder:Linear discrimina

82、nt analysis(LDA),1750 parametersMahsa ShoaranISSCC T8 Intelligent Neural Interfaces Source:M.Shaeri et al.,ISSCC24,JSSC24MiBMI:Miniaturized Brain-to-Text BMITBioCAS2215571 of 86DNC-based 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces Source

83、:M.Shaeri et al.,JSSC24,Nat BME25(in review)Distinctive Neural Code(DNC):Class Saliency MetricNeural Activity MapsClass Saliency MapsDNC SpaceNeural activity maps:Spike rates72 of 86 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces Source:M.S

84、haeri et al.,JSSC24,Nat BME25(in review)Distinctive Neural Code(DNC):Class Saliency MetricNeural Activity MapsClass Saliency MapsDNC SpaceClass saliency maps:For class discriminationNeural activity maps:Spike rates73 of 86 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8

85、Intelligent Neural Interfaces Source:M.Shaeri et al.,JSSC24,Nat BME25(in review)Distinctive Neural Code(DNC):Class Saliency MetricNeural Activity MapsClass Saliency MapsDNC SpaceDNCs:Class-specific low-dimensional spaceClass saliency maps:For class discriminationNeural activity maps:Spike rates74 of

86、 86 A few DNCs are sufficient for accurate class discrimination 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranClass Saliency Metric75 of 86i:The class of interest dij:The discrimination(distance)between a pair of classes GMjdij:Geometric mean of class i discrimination from all

87、other classesAMjdij:Arithmetic mean of class i discrimination Homogeneity of distances Class saliency for each feature is formulated as:Class saliency:The specific discrimination value for a class,calculated for each featureIt quantifies the homogeneous discriminative power of each feature for a cla

88、ss of interestOverall class discriminationISSCC T8 Intelligent Neural Interfaces Source:M.Shaeri et al.,JSSC24 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranNeural Activity Separation in DNC Sub-Space 76 of 86Three spatiotemporal features with the highest class saliency for eac

89、h character:DNC dimensionsA higher-dimensional DNC space may be required to identify a particular class,as in the case of class a”ISSCC T8 Intelligent Neural Interfaces Source:M.Shaeri et al.,JSSC24 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Inter

90、faces Source:M.Shaeri et al.,ISSCC24,JSSC24MiBMI Architecture and Timing DiagramBroadband AFEDC10kHz Small area(0.009mm2/Ch.)3.44W/Ch.77 of 86DAFE 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural Interfaces Source:M.Shaeri et al.,ISSCC24,JSSC24Broadband

91、 AFEDC10kHz Small area(0.009mm2/Ch.)3.44W/Ch.DNC-based decoderInput:512-Ch.spike ratesActivity onset detectionDNC selection31-class decoding78 of 86DAFEPredicted ClassSpike rateMiBMI Decoder ChipMiBMI Architecture and Timing DiagramTiming Diagram 2025 IEEE International Solid-State Circuits Conferen

92、ceMahsa ShoaranPower Breakdown and Performance DNC decoder chip:65nm TSMC,total power:223W(512-Ch.),activity-drivenDecoder area:0.75mm2Measured accuracy:91.3%Speed:60 Characters per minute79 of 86System power breakdownDecoder Chip OutputISSCC T8 Intelligent Neural Interfaces Source:M.Shaeri et al.,I

93、SSCC24,JSSC24 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranOutline80 of 86 Introduction and motivationsBasics of open-loop and closed-loop stimulationMajor trend:higher-channel-count systems Intelligent SoCs for symptom control:seizure detectionRecent trends:Scalability,adapta

94、bility,and data efficiency Intelligent SoCs for brain-machine interfaces Conclusion and future directions Relevant papers in ISSCC 2025ISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future Directions 2025 IEEE International Solid-State Circuits ConferenceIntelligent SoCs for closed-loop sym

95、ptom control and brain-machine interfacingKey focus areas in todays neural interface SoCsIntelligence,efficiency,miniaturization,and scalabilityFuture intelligent SoCs:new algorithms+new circuit techniquesIncreasing need for hardware-software co-design,cross-disciplinary innovationsEmerging trends a

96、nd future directionsNeuromorphic processing,in-memory computingSo far:ML models designed for tabular dataTransition to advanced time-series and sequence modellingConclusion and Future DirectionsMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future Directions81 of 86 2025 IEEE I

97、nternational Solid-State Circuits ConferenceSession 15:Neural Interfaces and Edge Intelligence for Medical Devices15.1 A 3.9mW 200words/min Neural Signal Processor in Speech Decoding for Brain-Machine Interface15.2 A 1024-Channel 0.00029mm2/ch 74nW/ch Online Spatial Spike-Sorting Chip with Event-Dri

98、ven Spike Detection and Self-Organizing Map Clustering 15.3 A 65nm Uncertainty-Quantifiable Ventricular Arrhythmia Detection Engine with 1.75J per Inference 15.4 A Neuroprosthetic SoC with Sensory Feedback Featuring Frequency-Splitting-Based Wireless Power Transfer with 200Mb/s 0.67pJ/b Backscatter

99、Data Uplink and Unsupervised Multi-Class Spike Sorting Relevant Papers in ISSCC 2025Mahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future Directions82 of 86 2025 IEEE International Solid-State Circuits ConferenceReferencesMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fund

100、amentals and Future DirectionsU.Shin et al.,A 256-Channel 0.227 J/class Versatile Brain Activity Classification and Closed-Loop Neuromodulation SoC with 0.004 mm2-1.51 W/channel Fast-Settling Highly Multiplexed Mixed-Signal Front-End”,ISSCC,2022.U.Shin,C.Ding,B.Zhu,Y.Vyza,A.Trouillet,E.C.Revol,S.P.L

101、acour,M.Shoaran,“Neuraltree:A 256-channel 0.227-j/class versatile neural activity classification and closed-loop neuromodulation soc,”IEEE Journal of Solid-State Circuits,vol.57,no.11,pp.32433257,2022.M.Shoaran,B.A.Haghi,M.Taghavi,M.Farivar,and A.Emami-Neyestanak,“Energy-efficient classification for

102、 resource-constrained biomedical applications,”IEEE Journal on Emerging and Selected Topics in Circuits and Systems,vol.8,no.4,pp.693707,2018.M.A.Shaeri,U.Shin,A.Yadav,R.Caramellino,G.Rainer,and M.Shoaran,“Mibmi:A 192/512-channel 2.46mm2 miniaturized brain-machine interface chipset enabling 31-class

103、 brain-to-text conversion through distinctive neural codes”,ISSCC,2024.M.Shaeri,U.Shin,A.Yadav,R.Caramellino,G.Rainer and M.Shoaran,A 2.46-mm2 Miniaturized Brain-Machine Interface(MiBMI)Enabling 31-Class Brain-to-Text Decoding,in IEEE Journal of Solid-State Circuits,vol.59,no.11,pp.3566-3579,Nov.202

104、4.T.Aflalo et al.Decoding motor imagery from the posterior parietal cortex of a tetraplegic human,Science,2015.F.R.Willett et al.High-performance brain-to-text communication via handwriting.Nature,2021.H.Lorach,et al.Walking naturally after spinal cord injury using a brainspine interface.Nature,2023

105、.S.L.Metzger,et al.A high-performance neuroprosthesis for speech decoding and avatar control.Nature,2023.M.Shaeri,A.Afzal and M.Shoaran,Challenges and opportunities of edge AI for next-generation implantable BMIs,AICAS,2022.C.Ding,M.Gao,A.K.Skrivervik and M.Shoaran,A 3 mm2,Energy-Efficient,Multi-Dat

106、a-Rate,FDMA Transmitter with On-Chip Antenna for Next-Generation Neural Implants,2024 IEEE European Solid-State Electronics Research Conference(ESSERC),Bruges,Belgium,2024.J.T.Robinson et al.,“An application-based taxonomy for braincomputer interfaces”Nat.Biomed.Eng(2024).A.Afzal et al.,Rest:Efficie

107、nt and accelerated eeg seizure analysis through residual state updates”,ICML,2024.J.Dan et al.,SzCORE:Seizure Community OpenSource Research Evaluation framework for the validation of electroencephalographybased automated seizure detection algorithms.Epilepsia(2024).83 of 86 2025 IEEE International S

108、olid-State Circuits ConferenceReferencesMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future Directions84 of 86Z.Taufique,B.Zhu,G.Coppola,M.Shoaran,and M.A.B.Altaf,“A low power multi-class migraine detection processor based on somatosensory evoked potentials,”IEEE Trans.Circui

109、ts Syst.II,Exp.Briefs,vol.68,no.5,pp.17201724,May 2021.G.OLeary et al.,A recursive-memory brain-state classifier with 32-channel track-and-zoom 2 ADCs and Charge-Balanced Programmable Waveform Neurostimulators,ISSCC,2018,pp.296-298.G.OLeary et al.,A Neuromorphic Multiplier-Less Bit-Serial Weight-Mem

110、ory-Optimized 1024-Tree Brain-State Classifier and Neuromodulation SoC with an 8-Channel Noise-Shaping SAR ADC Array,ISSCC,2020,pp.402-404.Y.Wang,Q.Sun,H.Luo,X.Chen,X.Wang and H.Zhang,A Closed-Loop Neuromodulation Chipset with 2-Level Classification Achieving 1.5Vpp CM Interference Tolerance,35dB St

111、imulation Artifact Rejection in 0.5ms and 97.8%Sensitivity Seizure Detection,ISSCC,2020.M.A.B.Altaf and J.Yoo,A 1.83 J/Classification,8-Channel,Patient-Specific Epileptic Seizure Classification SoC Using a Non-Linear Support Vector Machine,in IEEE Transactions on Biomedical Circuits and Systems,vol.

112、10,no.1,pp.49-60,Feb.2016.B.Zhu and M.Shoaran,Unsupervised Domain Adaptation for Cross-Subject Few-Shot Neurological Symptom Detection,2021 10th International IEEE/EMBS Conference on Neural Engineering(NER),Italy,2021,pp.181-184.M.Zhang,L.Zhang,C.-W.Tsai,and J.Yoo,“A patient-specific closed-loop epi

113、lepsy management SoC with one-shot learning and online tuning,”IEEE J.Solid-State Circuits,vol.57,no.4,pp.10491060,Apr.2022.B.Zhu,M.Farivar,and M.Shoaran,“ResOT:Resource-efficient oblique trees for neural signal classification,”IEEE Trans.Biomed.Circuits Syst.,vol.14,no.4,pp.692704,Aug.2020.U.Shin,C

114、.Ding,L.Somappa,V.Woods,A.S.Widge,and M.Shoaran,1071“A 16-channel 60 W neural synchrony processor for multi-mode phase-locked neurostimulation,”in Proc.IEEE Custom Integr.Circuits Conf(CICC),Apr.2022,pp.12.M.Nicolelis,“Actions from thoughts”,Nature 409,403407(2001).D.B.Rubin et al.,“Interim Safety P

115、rofile From the Feasibility Study of the BrainGate Neural Interface System”,Neurology,100(11),2023.A.Chua,M.I.Jordan,and R.Muller,“SOUL:An energy-efficient unsupervised online learning seizure detection classifier,”IEEE J.Solid-State Circuits,vol.57,no.8,pp.25322544,Aug.2022.2025 IEEE International

116、Solid-State Circuits ConferenceReferencesMahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future Directions85 of 86J.Yoo and M.Shoaran,“Neural interface systems with on-device 1075 computing:Machine learning and neuromorphic architectures,”Current Opinion Biotechnol.,vol.72,pp.95

117、101,Dec.2021.C.-W.Tsai,R.Jiang,L.Zhang,M.Zhang,L.Wu,J.Guo,Z.Yan,and J.Yoo,“Scicnn:A 0-shot-retraining patient-independent epilepsy-tracking soc,”ISSCC,2023,pp.488490.Y.Chen,E.Yao,and A.Basu,“A 128-channel extreme learning machine-based neural decoder for brain machine interfaces,”IEEE transactions o

118、n biomedical circuits and systems,vol.10,no.3,pp.679 692,2015.H.An et al.,“A power-efficient brain-machine interface system with a sub-mw feature extraction and decoding asic demonstrated in nonhuman primates,”IEEE transactions on biomedical circuits and systems,vol.16,no.3,pp.395 408,2022.Z.Zhong,Y

119、.Wei,L.C.Go,and J.Gu,“A sub-1j/class headset-integrated mind imagery and control soc for vr/mr applications with teacher-student cnn and general-purpose instruction set architecture”,ISSCC,vol.67,2024,pp.544546.J.Liu et al.,A High Accuracy and Ultra-Energy-Efficient Zero-Shot-Retraining Seizure Dete

120、ction Processor,in IEEE Journal of Solid-State Circuits,vol.59,no.11,pp.3549-3565,Nov.2024.J.Liu et al.,“BioAIP:A reconfigurable biomedical AI processor with adaptive learning for versatile intelligent health monitoring,”ISSCC,2021.J.Yoo,L.Yan,D.El-Damak,M.A.B.Altaf,A.H.Shoeb,and A.P.Chandrakasan,“A

121、n 8-channel scalable EEG acquisition SoC with patient-specific seizure classification and recording processor,”IEEE J.Solid-State Circuits,vol.48,no.1,pp.214228,Jan.2013.N.A.Steinmetz et al.Neuropixels 2.0:A miniaturized high-density probe for stable,long-term brain recordings.Science 372.6539,2021.

122、T.Kaiju,et al.,High Spatiotemporal Resolution ECoG Recording of Somatosensory Evoked Potentials With Flexible Micro-Electrode Arrays,Frontiers in Neural Circuits,2017.C.C.Jouny et al.,Improving early seizure detection.Epilepsy&Behavior 22(2011):S44-S48.C.Pandarinath et al.,High performance communica

123、tion by people with paralysis using an intracortical brain-computer interface”,eLife,2017.2025 IEEE International Solid-State Circuits ConferenceAcknowledgments Mahsa ShoaranISSCC T8 Intelligent Neural Interfaces:Fundamentals and Future Directions86 of 86Special thanks to Prof.Jerald Yoo Dr.Fei Tan

124、Arshia Afzal Huanshihong Deng 2025 IEEE International Solid-State Circuits ConferenceData Throughput BottleneckMahsa ShoaranISSCC T8 Intelligent Neural Interfaces Source:C.Ding et al.,CICC24,ESSERC24Neural SignalECoGLFPAction PotentialBandwidth1500Hz1500Hz0.2510kHzChannel Count100010,000100010,00010

125、0010,000Sampling Frequency 2kS/s2kS/s20kS/sADC Bit#101010Raw Data Rate20200Mb/s20200Mb/s2002000Mb/sData Rate=NchFsampleADCbitTXADCUplink AFELNAMUXCH1CHNAssuming an energy efficiency of 0.11 nJ/bit for the transmitterA data rate of 200 Mb/s results in a transmission power of 20200 mW!This power consu

126、mption exceeds the safe power levels for brain implantsPotential solution:Compressing neural signals prior to transmission87 of 86 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranISSCC T8 Intelligent Neural InterfacesSource:M.Shoaran et al.,TBioCAS14,VLSI Symp16Compressive Sensin

127、g of Neural SignalsNeural Interface Moduleincluding Electrodes and an ASICLow complexity,universal encoder,1020 compression(lossy)Complex reconstruction,high latency,not suitable for real-time applications88 of 86Compressive SensingSpatial-Domain Compressive Sensing 2025 IEEE International Solid-State Circuits ConferenceMahsa ShoaranMiBMI Performance Comparison4 channel#13 area10 power Increased task complexity:First BMI for communication recovery89 of 86ISSCC T8 Intelligent Neural Interfaces Source:M.Shaeri et al.,ISSCC24,JSSC24