Development and validation of a clinical wearable deep learning based continuous inhospital deterioration prediction model

Abstract

Standard episodic patient monitoring of vital signs on the medical-surgical wards can potentially miss changes in health status and delay recognition of risk. To reduce these delays, we develop a clinical wearable-based deep learning model, using 9 inputs, comprised of continuous vital signs and demographics, to identify the onset of deterioration early and accurately. Using validated data from 888 adult non-intensive care unit inpatient visits with 135 outcomes, from two different clinical grade wearables, we train a recurrent neural network to predict clinical alerts and adverse clinical outcomes in the subsequent 24 hours. Our continuous clinical alert model is able to predict both clinical alerts (Area under both the Receiver Operator Characteristic curve 0.89 + /− 0.3, Precision Recall curve 0.58 + /− 0.14) and adverse clinical outcomes (accuracy: 81.8% on 11 events) up to 17 hours in advance. Our wearable based continuous clinical alert system outperforms episodic clinical support tools in detecting deterioration, retains its performance when tested on data from a different clinical wearable than the one it was trained on and can produce alerts ahead of a broad class of adverse clinical outcomes, enabling timely interventions that can avert preventable deteriorations and reduce hospital costs.