eduzhai > Applied Sciences > Engineering >

Interpretable Factorization for Neural Network ECG Models

  • Save

... pages left unread,continue reading

Document pages: 16 pages

Abstract: The ability of deep learning (DL) to improve the practice of medicine and itsclinical outcomes faces a looming obstacle: model interpretation. Withoutdescription of how outputs are generated, a collaborating physician can neitherresolve when the model s conclusions are in conflict with his or her own, norlearn to anticipate model behavior. Current research aims to interpret networksthat diagnose ECG recordings, which has great potential impact as recordingsbecome more personalized and widely deployed. A generalizable impact beyondECGs lies in the ability to provide a rich test-bed for the development ofinterpretive techniques in medicine. Interpretive techniques for Deep NeuralNetworks (DNNs), however, tend to be heuristic and observational in nature,lacking the mathematical rigor one might expect in the analysis of mathequations. The motivation of this paper is to offer a third option, ascientific approach. We treat the model output itself as a phenomenon to beexplained through component parts and equations governing their behavior. Weargue that these component parts should also be "black boxes " --additionaltargets to interpret heuristically with clear functional connection to theoriginal. We show how to rigorously factor a DNN into a hierarchical equationconsisting of black box variables. This is not a subdivision into physicalparts, like an organism into its cells; it is but one choice of an equationinto a collection of abstract functions. Yet, for DNNs trained to identifynormal ECG waveforms on PhysioNet 2017 Challenge data, we demonstrate thischoice yields interpretable component models identified with visual compositesketches of ECG samples in corresponding input regions. Moreover, the recursiondistills this interpretation: additional factorization of component black boxescorresponds to ECG partitions that are more morphologically pure.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×