Project DetailsSpinal and back surgeries carry the risk of impairing the patient’s neural function. To reduce this risk, neurosurgeons monitor SSEP signals in real time. These signals are complex medical data that require expert labeling, a resource-intensive process that leaves us with both labeled and many unlabeled examples. In part A of the project, we used only the labeled data to build an SSEP classifier using fully supervised learning. The goal of this follow-up project is to develop a deep learning and signal processing system that also leverages the unlabeled data, aiming to surpass the performance of the initial model. Based on the architecture and preprocessing pipeline of part A, we explored various unsupervised and self-supervised approaches. Attempts to extract meaningful representations and train classifiers on top of them failed, mainly due to the sensitivity of the signals to augmentations and due to the high similarity between classes. In contrast, the pseudo-labeling approach proved effective, yielding a stable and reliable classifier that outperformed the original model.
