Using deep networks for scientific discovery in physiological signals

Tom Beer, Bar Eini-Porat, Sebastian Goodfellow, Danny Eytan, Uri Shalit
Proceedings of the 5th Machine Learning for Healthcare Conference, PMLR 126:685-709, 2020.

Abstract

Deep neural networks (DNN) have shown remarkable success in the classification of physiological signals. In this study we propose a method for examining to what extent does a DNN’s performance rely on rediscovering existing features of the signals, as opposed to discovering genuinely new features. Moreover, we offer a novel method of "removing" a hand-engineered feature from the network’s hypothesis space, thus forcing it to try and learn representations which are different from known ones, as a method of scientific exploration. We then build on existing work in the field of interpretability, specifically class activation maps, to try and infer what new features the network has learned. We demonstrate this approach using ECG and EEG signals. With respect to ECG signals we show that for the specific task of classifying atrial fibrillation, DNNs are likely rediscovering known features. We also show how our method could be used to discover new features, by selectively removing some ECG features and "rediscovering" them. We further examine how could our method be used as a tool for examining scientific hypotheses. We simulate this scenario by looking into the importance of eye movements in classifying sleep from EEG. We show that our tool can successfully focus a researcher’s attention by bringing to light patterns in the data that would be hidden otherwise.

Cite this Paper

BibTeX
@InProceedings{pmlr-v126-beer20a, title = {Using deep networks for scientific discovery in physiological signals}, author = {Beer, Tom and Eini-Porat, Bar and Goodfellow, Sebastian and Eytan, Danny and Shalit, Uri}, booktitle = {Proceedings of the 5th Machine Learning for Healthcare Conference}, pages = {685--709}, year = {2020}, editor = {Doshi-Velez, Finale and Fackler, Jim and Jung, Ken and Kale, David and Ranganath, Rajesh and Wallace, Byron and Wiens, Jenna}, volume = {126}, series = {Proceedings of Machine Learning Research}, month = {07--08 Aug}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v126/beer20a/beer20a.pdf}, url = {https://proceedings.mlr.press/v126/beer20a.html}, abstract = {Deep neural networks (DNN) have shown remarkable success in the classification of physiological signals. In this study we propose a method for examining to what extent does a DNN’s performance rely on rediscovering existing features of the signals, as opposed to discovering genuinely new features. Moreover, we offer a novel method of "removing" a hand-engineered feature from the network’s hypothesis space, thus forcing it to try and learn representations which are different from known ones, as a method of scientific exploration. We then build on existing work in the field of interpretability, specifically class activation maps, to try and infer what new features the network has learned. We demonstrate this approach using ECG and EEG signals. With respect to ECG signals we show that for the specific task of classifying atrial fibrillation, DNNs are likely rediscovering known features. We also show how our method could be used to discover new features, by selectively removing some ECG features and "rediscovering" them. We further examine how could our method be used as a tool for examining scientific hypotheses. We simulate this scenario by looking into the importance of eye movements in classifying sleep from EEG. We show that our tool can successfully focus a researcher’s attention by bringing to light patterns in the data that would be hidden otherwise.} }
Endnote
%0 Conference Paper %T Using deep networks for scientific discovery in physiological signals %A Tom Beer %A Bar Eini-Porat %A Sebastian Goodfellow %A Danny Eytan %A Uri Shalit %B Proceedings of the 5th Machine Learning for Healthcare Conference %C Proceedings of Machine Learning Research %D 2020 %E Finale Doshi-Velez %E Jim Fackler %E Ken Jung %E David Kale %E Rajesh Ranganath %E Byron Wallace %E Jenna Wiens %F pmlr-v126-beer20a %I PMLR %P 685--709 %U https://proceedings.mlr.press/v126/beer20a.html %V 126 %X Deep neural networks (DNN) have shown remarkable success in the classification of physiological signals. In this study we propose a method for examining to what extent does a DNN’s performance rely on rediscovering existing features of the signals, as opposed to discovering genuinely new features. Moreover, we offer a novel method of "removing" a hand-engineered feature from the network’s hypothesis space, thus forcing it to try and learn representations which are different from known ones, as a method of scientific exploration. We then build on existing work in the field of interpretability, specifically class activation maps, to try and infer what new features the network has learned. We demonstrate this approach using ECG and EEG signals. With respect to ECG signals we show that for the specific task of classifying atrial fibrillation, DNNs are likely rediscovering known features. We also show how our method could be used to discover new features, by selectively removing some ECG features and "rediscovering" them. We further examine how could our method be used as a tool for examining scientific hypotheses. We simulate this scenario by looking into the importance of eye movements in classifying sleep from EEG. We show that our tool can successfully focus a researcher’s attention by bringing to light patterns in the data that would be hidden otherwise.
APA
Beer, T., Eini-Porat, B., Goodfellow, S., Eytan, D. & Shalit, U.. (2020). Using deep networks for scientific discovery in physiological signals. Proceedings of the 5th Machine Learning for Healthcare Conference, in Proceedings of Machine Learning Research 126:685-709 Available from https://proceedings.mlr.press/v126/beer20a.html.

Related Material