EE M255, BME M260, NS M206:

Transcription

1 EE M255, BME M260, NS M206: NeuroEngineering Lecture Set 6: Neural Recording Prof. Dejan Markovic Agenda Neural Recording EE Model System Components Wireless Tx 6.2

2 Neural Recording Electrodes sense action potentials (spikes) Peak membrane voltage ~ mv Recorded voltage ~ mv Firing i rates of neurons ~ Hz Courtesy: Z. Nadasdy (Caltech) A single electrode can detect many (1 8) neurons 6.3 Frequency Components of Neural Signals < 300 Hz: Local Field Potentials (LFP) < 3 Hz:Delta(slow wave sleep) 4 7 Hz: Theta (wake and REM) 8 12 Hz: Alpha (drowsiness) Hz: Beta (decision-making, motor planning) Hz: Gamma (sensory, attentive) Hz: Single-Unit Activity ( spikes ) LFPs Spikes 6.4

3 Goal: Recording and Stimulation [1] M.A.L. Nicolelis, Actions from thoughts, Nature 409 (2001), pp Neural Recording: EE Model 6.6

4 Front-End Design Analog vs. Digital Detection? S. Gibson et al., An Efficiency Comparison of Analog and Digital Spike Detection, NER, May 2009, pp Front-End Components: Low-Noise Amplifier (LNA) Noise due to circuits β: proportionality constant, k: tech parameter Input-referred noise (assumes ideal amplifier) Goal: Input-referred noise below few µv R. Chandler et al., A System-Level View of Optimizing High-Channel-Count Wireless Biosignal Telemetry, EMBC, Sep. 2009, pp

5 LNA Input-Referred Noise Goal: Input-referred noise below few µv γ: MOSFET noise coefficient, K AMP : amp architecture, IC: tech parameter, (for a given C 1 ) R. Chandler et al., A System-Level View of Optimizing High-Channel-Count Wireless Biosignal Telemetry, EMBC, Sep. 2009, pp Front-End Components: Analog-to-Digital Converter (ADC) Simple ADC Models Parameters Speed (Fs) Resolution (B) Figure-of-Merit (FoM) [fj / conv-step] Power Area P ADC = FoM Fs 2 B A ADC = A 0 2 2(B B0) = (B 8) (A 0 is the unit area) S. Gibson et al., An Efficiency Comparison of Analog and Digital Spike Detection, NER, May 2009, pp

6 Performance of Recent ADCs and FoM Contours State-of-the-art ADCs: ~10 fj / conv-step S. Gibson et al., An Efficiency Comparison of Analog and Digital Spike Detection, NER, May 2009, pp ADC Analysis As technology and circuits improve, FoM also improves Power increases Linearly with speed (Fs) Exponentially with resolution (B) Area increases Exponentially with resolution Example: 8-bit, 24kS/s, 100fJ/conv-step Power ~ 600nW Area ~ 0.05mm 2 (90nm technology) S. Gibson et al., An Efficiency Comparison of Analog and Digital Spike Detection, NER, May 2009, pp

7 Parameters: Effective ADC Power k: ratio of standby/switching power E.g. Biasing and reference circuits r D : detection rate [Samples/s] r D = max{r N L P D + (Fs r N L) P FA, Fs} r N : firing rate of the neurons [Spikes/s] L: length of a spike [Samples/s] P buf : power of readout circuitry for buffered samples (analog memory) Effective power: P eff = ((1 k) FoM 2 B + P buf /Fs) r D + k FoM 2 B Fs S. Gibson et al., An Efficiency Comparison of Analog and Digital Spike Detection, NER, May 2009, pp Detection: Pulsed / Spike Pulsed: a pulse at the output each time the waveform crosses the threshold Spike: spike samples at the output S. Gibson et al., An Efficiency Comparison of Analog and Digital Spike Detection, NER, May 2009, pp

8 Analog-Digital ~8 bits S. Gibson et al., An Efficiency Comparison of Analog and Digital Spike Detection, NER, May 2009, pp What is Neural Spike Sorting? Electrophysiology is the technique of recording electrical signals from the brain using microelectrodes Single-unit activity (signals recorded from individual neurons) is needed for: Basic neuroscience experiments Medical applications, e.g.: Epilepsy treatment Brain-machine interfaces (BMIs) Neural signals recorded by microelectrodes are frequently composed of activity from multiple neurons surrounding the electrode Spike sorting is the process of assigning each action potential to its originating neuron so that information can be decoded Courtesy: S. Gibson 6.16

9 The Spike-Sorting Process 1. Spike Detection: Separating spikes from noise 2. Alignment: Aligning detected spikes to a common reference 3. Feature Extraction: Transforming spikes into a certain set of features (e.g. principal components) 4. Dimensionality Reduction: Choosing which features to use in clustering 5. Clustering: Classifying spikes into different groups (neurons) based on extracted features Courtesy: S. Gibson 6.17 Spike Sorting: Simplified View Spike sorting: The process of classifying action potentials according to the source neurons Detection (D) & Alignment (A) Feature Extraction (FE) Clustering (C) 6.18

10 Need for On-Chip Spike Sorting Traditional neural recording system Transmission of raw data using wires Spike sorting offline in software Disadvantages of traditional approach Not real time Limited number of channels Limited movement of subject Risk of infection by wires On-chip spike sorting Faster processing Data-rate reduction Wireless transmission 6.19 Example: Epilepsy Monitoring Technology Example: 100-channel recording Courtesy: R. Staba (UCLA) 100 channels x 25 ks/s x 10 bits = 25 Mbps (raw data) Technical challenges 2 TB / week!!! Power density: < 800 µw/mm 2 (~size of a filer) Low data rate for wireless transmission Our research Real-time on-chip data compression (a.k.a. spike sorting ) Rapid processing of exiting massive data records 6.20

11 Spike Sorting Algorithms A large number of algorithms Steps: Detection, Alignment, Feature Extraction, Coefficient Selection, Clustering Table 2: Summary of algorithms for spike sorting. Last row indicates our current progress in algorithm studies. Detection Alignment Feature Extraction Coeff. Selection Clustering * Abs. value * Threshold crossing * PCA * k-means * K-S test * Nonlinear energy * Maximum * Discrete wavelet * Fuzzy c-means * Lilliefors test operator * Center of mass transform * Valley-seeking * Uniform sampling * Stationary wavelet * Max. derivative * Integral transform * SPC/WaveClus * Smart sampling transform product * Max. NEO * Discrete derivatives * Osort Preliminary Near-term Preliminary Near-term Longer-term Focus of next lecture We will assume numbers here to look into a complete system 6.21 Wireless Telemetry System Multi-channel recording and wireless Tx R. Chandler et al., A System-Level View of Optimizing High-Channel-Count Wireless Biosignal Telemetry, EMBC, Sep. 2009, pp

12 Radio Transmitter (Tx) Required Tx Power (from Tx out to demodulated data) Parameters: 2 k T Rs: noise generated by a 50-Ohm Rs (antenna impedance) NF: ratio of the total noise at the Rx output to the total noise contribution due to Rs SNR: signal-to-noise noise ratio required for decoding the data with < e.g P bit error rate SNR db = 10log signal 10 P noise BW: bandwidth of the communication channel PL: path loss (function of distance) G Tx(Rx) : Tx(Rx) antenna gain (antenna design) 6.23 Typical Link Budget Data rate: 10 Mbps Bandwidth: 10 MHz Distance: 10 m Frequency: 2.4 GHz Total link power: ~ 3 mw R. Chandler et al., A System-Level View of Optimizing High-Channel-Count Wireless Biosignal Telemetry, EMBC, Sep. 2009, pp

13 Total System Power Several System Design Options (100-ch target) (a) Raw data, (b) DSP on chip, (c) Analog Det + DSP R. Chandler et al., A System-Level View of Optimizing High-Channel-Count Wireless Biosignal Telemetry, EMBC, Sep. 2009, pp (a) Raw Streaming Mode 10 Mbps link: we can support 52 channels mw It takes T start for frequency synthesizer output to stabilize Burst mode: synthesizer and Tx turned on every T cycle Energy overhead can be reduced by faster T start or longer latency Courtesy: R. Chandler 6.26

14 (b) DSP Detection Mode 100 channels possible with 3.2 Mbps Data compression: 1/5 100 Hz detection rate 54% power reduction Courtesy: R. Chandler 6.27 Total System Power (mw) Different levels of processing Digital processing relaxes Tx requirements Bold: feasible low-power configurations R. Chandler et al., A System-Level View of Optimizing High-Channel-Count Wireless Biosignal Telemetry, EMBC, Sep. 2009, pp

15 Wireless Recording Neural Systems Courtesy: R. Chandler 6.29 Takeaway Points Analog detection is suitable for large ADC resolutions (~ 8 or 9 bits) Spike-sorting (compression) allows relaxes Tx requirements Lowers overall system power Increases number of channels All system components have to be in balance for minimum overall power 6.30