One-shot Learning for Seizure Detection and Identification of Epileptogenic Brain Regions from Long-time Human iEEG Recording with End-to-end Binary Operations

This thesis presents an efficient algorithm by combining symbolic dynamics and brain-inspired hyperdimensional (HD) computing for both seizure onset detection and identification of ictogenic (= seizure generating) brain regions from intracranial electroencephalography (iEEG). Moreover, the simplicity of the algorithm eases its implementation on an embedded AI computing device platform (e.g. the NVIDIA Jetson TX2 Module) for long term operation. The proposed algorithm provides: (1) a unified method for both learning and classification tasks with end-to-end binary operations; (2) one-shot learning from seizure examples; (3) linear computational scalability to any number of electrodes; (4) generation of transparent codes with interpretable features; (5) a simple embedded implementation which is fast and energy efficient. The algorithm first transforms iEEG time series from each electrode into symbolic local binary pattern codes from which a distributed representation of the brain state of interest is constructed across all the electrodes and over time in a hyperdimensional space. Such holographic representation is used to quickly learn from seizures, detect their onset, and identify the spatial brain regions that generated them. Moreover, HD computing is characterized by one-shot or anyway fast learning, making it a prime candidate for utilization in such a domain with a typical low quantity of training data. I assess the performance of the proposed algorithm on two different dataset: (1) the first contains 99 short-time iEEG recordings from 16 drug-resistant epilepsy patients being implanted with 36 to 100 electrodes; (2) the second is composed by 18 long-time recordings from 18 drug-resistant epilepsy patients: a total of 2656 interictal hours and 120 seizures are contained in the recordings. All the patients come from the epilepsy surgery program of the Inselspital Bern. On the first dataset, for the majority of the patients (10 out of 16), our algorithm quickly learns from one or two seizures and perfectly (100%) generalizes on novel seizures using k-fold cross-validation. For the remaining six patients, the algorithm requires three to six seizures for learning. Our algorithm surpasses the state-of-the-art including deep learning algorithms by achieving higher specificity (94.84% vs. 94.77%) and macroaveraging accuracy (95.42% vs. 94.96%), and 74× lower memory footprint, but slightly higher average latency in detection (15.9 s vs. 14.7 s). On the second dataset,the algorithm learns from one or two seizures and achieves 0.0 false detection rate for all the patients. The state-of-the-art achieves again lower latency in detection (12.8 s vs. 17.3 s), but higher false detection rate (0.31 f/h) Moreover, the algorithm can reliably identify (with a p-value < 0.01) the relevant electrodes covering an ictogenic brain region at two levels of granularity: cerebral hemispheres and lobes. Finally, the algorithm shows 15× gain in execution time and 18× gain in energy consumption with respect to the state-of-the-art competitors, when implemented on the NVIDIA TX2 platform.