Predicting Post-Operative Complications with Wearables

Abstract

Post-operative complications and hospital readmission are of great concern to surgical patients and health care providers. Wearable devices such as Fitbit wristbands enable long-term and non-intrusive monitoring of patients outside clinical environments. To build accurate predictive models based on wearable data, however, requires effective feature engineering to extract high-level features from time series data collected by the wearable sensors. This paper presents a pipeline for developing clinical predictive models based on wearable sensors. The core of the pipeline is a multi-level feature engineering framework for extracting high-level features from fine-grained time series data. The framework integrates a set of techniques tailored for noisy and incomplete wearable data collected in real-world clinical studies: (1) singular spectrum analysis for extracting high-level features from daily features over the course of the study; (2) a set of daily features that are resilient to missing data in wearable time series data; (3) a K-Nearest Neighbors (KNN) method for imputing short missing heart rate segments; (4) the integration of patients' clinical characteristics and wearable features. We evaluated the feature engineering approach and machine learning models in a clinical study involving 61 patients undergoing pancreatic surgery. Linear support vector machine (SVM) with integrated feature engineering achieved an AUROC of 0.8802 for predicting post-operative readmission or severe complications, which significantly outperformed the existing rule-based model used in clinical practice and other state-of-the-art feature engineering approaches.