Hello, my name is Michael Martini and today I'll be presenting our work from our team at Mount Sinai on Deep Self-Supervised Learning for Patient-Specific Anomaly Detection in Stereoencephalography, a method inspired from the Mars Rover Curiosity. As some quick background, epilepsy is a chronic disorder of the central nervous system that predisposes individuals to recurrent seizures, affecting approximately 3 million people in the United States. Of note, a subset of approximately 1.2 million patients have medically refractory epilepsy and experience frequent, unpredictable seizures that do not respond to anti-epileptic medications. This necessitates neurosurgical intervention to ablate the responsible region to treat the seizures. To better monitor these patients, they receive intracranial stereo EEG leads, and specific EEG waveforms indicative of abnormal activity must be recognized by expert clinicians to inform subsequent neurosurgical intervention. These patients are monitored continually for days to weeks, and abnormal events are rare and often difficult to distinguish from background signal, which makes detection quite difficult, as well as a time- and resource-intensive process. Switching gears, in 2018, NASA's Jet Propulsion Laboratory proposed a Modified Long Short-Term Memory Network, or LSTM, utilizing nonparametric dynamic thresholding as a novel pruning procedure to enhance anomaly detection and mitigate false positives in time-series data. Using a sliding-window approach, the algorithm finds a threshold such that if all values above it are removed, it would cause the greatest percent decrease in the mean and standard deviation of the smoothed error. The function also penalizes for having large numbers of anomalies to prevent overly greedy behavior. Of note, this technique was deployed to effectively detect anomalous events in real-time within data acquired from the Mars Rover Curiosity. The methods for our study are as follows. Continuous in-hospital recordings were acquired using the Natus Neuroworks platform at our hospital. This study included 14 medically refractory epilepsy patients who received continuous stereo EEG recordings in diverse brain regions for their treatment. Pre-processing pipelines using MNE and Scikit-learn packages in Python were used to transform recordings to scaled standardized formats that were easily ingested into a neural network. Our LSTM network and experimental parameters were adapted from the NASA Jet Propulsion Laboratory's GitHub account, and predicted anomalies were compared to those recognized by fellowship-trained epileptologists at our institution. When we applied these methods to our dataset, we found a significant difference in performance. This figure demonstrates the process of signal reconstruction and shows why dynamic thresholding successfully mitigates false positives. Shown above is the reconstructed or predicted signal overlaid on the actual signal. There are a few areas highlighted in purple where the reconstruction does not adequately match the actual signal. This produces spikes in the error signal shown below. Dynamic thresholding along a sliding window captures the true positives and prunes the false positives, while the static thresholding method captures both, leading to a higher false positive rate. Crossover experiments were also performed to assess the patient specificity of the individual models. In the ordinary paradigm of this, patient A and B have recordings that are split evenly for training and testing. In the crossover paradigm, the same testing recordings are kept, however, the models are trained on train recordings derived from different patients in the study population. This is just a quick representative example showing that these models are in fact patient specific. On the bottom right, you can see that the errors are dramatically increased compared to the bottom left when training was performed on the same patient. This suggests that the self-learned latent representations of the stereo EEG recordings during training are patient specific, which is not surprising given the diversity of waveforms and signal properties between patients and brain regions. However, this is also not to pigeonhole the utility of this method, as non-crossover models worked for most patients across most brain regions, such as hippocampal, premotor, etc. To conclude, this is the first study to apply novel dynamic thresholding and error pruning procedures from NASA's Jet Propulsion Laboratory to improve the deep anomaly detection in stereo EEG data. High reconstruction errors in known anomalous regions suggested that the modified LSTM network clearly detected seizures in most of the patients tested, despite diverse epileptic waveforms in varied brain regions. The dynamic thresholding reduced the number of false positives compared to traditional static thresholding approaches, evidenced by an increased PPV, and conversely, crossover experiments produced dramatic increases in the false positive detection, suggesting that these trained models were tuned to be patient-specific, evidenced by their much lower positive predictive value. Finally, I want to thank the people at Mount Sinai who made this project possible, including our team at AI Sinai, led by Drs. Ohrman and Costa, as well as our neurosurgical leadership, led by Drs. Betterson and Dr. Mako, and finally, the epilepsy care team, including Dr. Panov and Dr. Ghattan. Thank you for your attention.

deep self-supervised learning

patient-specific anomaly detection

stereoencephalography

Modified Long Short-Term Memory Network

dynamic thresholding