Two Algorithms for Real-Time Seizure Prediction and Detection, for an Implanted, Closed-Loop, Epilepsy Prosthesis In Vivo P. Rajdev1, S. Raghunathan2, P. Irazoqui3 1, 2 Graduate Student, Purdue University, 3 Asst. Professor, Purdue University Outline • • • Motivation Experimental Setup Prediction Algorithm – – – – • • Algorithm Digital Signal Processor Thresholds Results Hardware Constraints in Implantable Applications Event Based Seizure Detection – Algorithm – Hardware Implementation – Results • Applications and Future Directions Motivation • Real time Implementation – Algorithm with low computational complexity. – Implementable on Digital Signal Processor or Application Specific Integrated Circuit. • Local Field Potential (LFP) Vs Electroencephalogram (EEG) – Disadvantages of EEG – Disadvantages of LFPs Experimental Setup • Animal Model of epilepsy: Kainate model – Commonly used model for temporal lobe epilepsy. – Primary site of action of kainic acid is the CA3 cells of the hippocampus. – Pathological, clinical, and electrographic characteristics of the seizures caused by kainic acid treatment strongly resemble those seen in human temporal lobe epilepsy Digital Signal Processor • The prediction algorithm was first developed in Matlab. • A real-time implementation was then realized on a floating point TMS320C6713T digital signal processor (DSP). – DSP chip operating at 225MHz, along with 16Mbytes of SDRAM, 415Kbytes of Flash memory and a JTAG emulator. Weiner Prediction Based Algorithm • Four-step process – – – – signal enhancement, adaptive auto-regressive modeling and prediction, envelope detection, and a binomial decision rule. Algorithm (contd…) • • Implemented a real-time wiener-prediction based algorithm on a digital signal processor. – Quasi-stationary signal – Adaptive nature of the algorithm ensures that the prediction coefficients provide effective prediction of baseline activity. – Lower Computational complexity In an autoregressive (AR) model, the future value is modeled as a linear combination of the p past values of the signal. p s[n 1] a[k]s[n k] k 0 Algorithm (contd…) 40 sec prior to seizure onset 1400 1200 Mean =124.36 1000 Mean = 22.45 800 600 400 200 Mean = 4.93 0 Baseline Pre-ictal Ictal Thresholds Sensitivity T False positive rate (FPR) Latency Latency (sec) 2 50 0 1 1 1.2 1.4 1.6 Lambda (λ) Lambda (λ) 1.8 2 0 False positives (/hr) Sensitivity 100 1 N 2 e [n] N n1 Results Results ID # of Seizures Sensitivity FP/Min (%) Median Mean latency Std of latency Latency (sec) (sec) (sec) 1 32 96.87 0.0064 19.96 26.02 20.84 2 27 96.29 0.0095 34.79 33.29 18.12 3 25 88 0.0063 29.18 31.34 19.61 4 25 96 0.0143 13.82 15.51 10.50 Median Latency Mean latency Std of latency (sec) (sec) (sec) ID # of Seizures Sensitivity (%) FP/Min 1 14 92.85 0.115 6.33 5.63 4.29 2 14 92.85 0.077 7.68 8.35 7.63 3 24 91.67 0.074 5.18 6.35 4.87 4 18 88.89 0.080 6.91 7.24 5.34 Hardware constraints in an implantable application 1. 2. 3. 4. Power consumed / Battery life Total area /size Programmability/Communication link to an external monitor Integration capabilities Goals for feedback algorithm: 1. 2. 3. 4. Simplicity in implementation Good sensitivity Adaptability, allowing for patient to patient variations Integration capabilities Temporal evolution and spread 1. Radial spread at speeds up to 60cm/sec (Jung , 2003) 2. 2-70 seconds from hippocampal focus to neo-cortex (Spencer ,1987) 3. Animal studies indicate a delay of ~ 20 s before spreading away from the temporal lobe focus (Litt, 2003) Event based seizure detection 1. Amplitude of recorded signal (Kamp ) 2. Measure of frequency content obtained from inter-event interval (IEIth ) 3. Measure of rhythmicity obtained from sustained levels of increasing amplitude, high frequency content in recorded signal (NStage) Distributing the event threshold (Kamp ) Distributing the IEI threshold (IEIth) Hardware implementation Hardware implementationcircuit timing results Where do the hardware trade-offs figure? Optimizing threshold selection Results Power consumption 350 nW Supply voltage 250 mV Area 370 x 130 um (per channel) Sensitivity 0.94 Selectivity 0.90 Detection delay 6.4 s *references available on request ASIC Design/ Where are we going? Microchip Reid Harrison Lab, Univ. of Utah Center for Wireless Integrated Microsystems, Univ. of Michigan Applications of device Integration with multi-channel neural recording devices Integration with implantable neural stimulators Seizure focus identification and tracking Acknowledgments • Research partners • • • • • • BCI Lab members – – – – Funding • • • Robert Worth, M.D., Ph.D. Thomas Sutula, Ph.D. Jenna Rickus, Ph.D. Edward Bartlett, Ph.D. Kaushik Roy, Ph.D. • Cyberonics, Inc. Wallace H. Coulter Early Career Award Additional industry collaborators • Texas Instruments – https://engineering.purdue.edu/BCILab Professor: Pedro P. Irazoqui Research Associate: Casey Ellison Post Doctorate: Kate Musick PhD Students • Travis Hassell, BME • Eric Chow, ECE • Pooja Rajdev, BME • Shriram Raghunathan, BME • Matt Ward, BME • Brooke Beier, BME • Art Chlebowski, BME • Bhupendra Manola, BME Masters Students • Matthew Graves, BME • Adam Kahn, BME • Gabriel Albors, BME