Optimizing the trade-off between utility and performance in interpretable sleep classification
Primary Area: applications to neuroscience & cognitive science
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: sleep staging, interpretability, representation learning, embedding, cnn, lstm, eeg
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: NormIntSleep combines clinical guidelines & neural embeddings for interpretable sleep state annotation, outperforming prior techniques and providing a safe, generalizable option for healthcare.
Abstract: Deep learning has made significant strides in numerous fields, yet its adoption in healthcare has been slow due to the considerable risks associated with clinical applications. Explainable models are essential to foster trust and accountability. This work examines the trade-off between interpretable techniques for automating sleep state annotation, a critical step in diagnosing sleep disorders. We introduce an interpretable approach, NormIntSleep, that produces explanations grounded in clinical guidelines by combining meaningful features with deep neural network embeddings. Furthermore, we propose the metric $Alignment_{DT}$ to quantify domain-grounded interpretability and the resulting utility of explanations. Crucially, NormIntSleep outperforms prior interpretable techniques with 0.814--0.847 accuracy, 0.787--0.793 F1-score, 0.759--0.788 $\kappa$, and the hightest $Alignment_{DT}$ score. NormIntSleep represents a potentially generalizable interpretable machine learning approach where domain knowledge is essential for safe and efficient implementation in healthcare.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5949
Loading