Surgical Prediction with Interpretable Latent Representation

Bing Xue, York Jiao, Thomas Kannampallil, Joanna Abraham, Christopher Ryan King, Bradley A Fritz, Michael Avidan, Chenyang Lu

Keywords: machine learning, healthcare applications, latent encoding, surgical predictions

Abstract: Given the risks and cost of surgeries, there has been significant interest in exploiting predictive models to improve perioperative care. However, due to the high dimensionality and noisiness of perioperative data, it is challenging to develop accurate, robust and interpretable encoding for surgical applications. We propose surgical VAE (sVAE), a representation learning framework for perioperative data based on variational autoencoder (VAE). sVAE provides a holistic approach combining two salient features tailored for surgical applications. To overcome performance limitations of traditional VAE, it is prediction-guided with explicit expression of predicted outcome in the latent representation. Furthermore, it disentangles the latent space so that it can be interpreted in a clinically meaningful fashion. We apply sVAE to two real-world perioperative datasets and the open MIMIC-III dataset to evaluate its efficacy and performance in predicting diverse outcomes including surgery duration, postoperative complication, ICU duration, and mortality. Our results show that the latent representation provided by sVAE leads to superior performance in classification, regression and multi-task predictions. We further demonstrate the interpretability of the disentangled representation and its capability to capture intrinsic characteristics of surgical patients.

One-sentence Summary: Application of a tailored VAE model to surgical datasets that achieves high-performance prediction and clinically meaningful phenotyping effects.