Opening The Black Box: Enhancing Machine Learning Interpretability To Optimize Clinical Response To Sudden Deterioration In COVID-19 Patients

项目总结/文摘

Project Summary/Abstract Advanced machine learning (ML) has consistently been shown to outperform expert opinion and more simple analytics for predicting clinical outcomes. However, there has been a paucity of successful prospective clinical implementations of such tools. The unique barriers to advanced ML implementation and adoption in healthcare are (1) the technological challenges of running and displaying these models in real-time within existing workflows and (2) a general distrust for black box algorithms among highly skilled providers. As a result, the promise of these tools is largely lost in healthcare. This is particularly problematic in COVID-19, where patients can deteriorate rapidly, from appearing stable to suddenly being in respiratory failure or shock with little obvious warning. Early recognition of this deterioration is vital to proactive interventions, which can improve outcomes. eCART is a predictive analytic that has been developed iteratively at the University of Chicago over the past decade to identify hospitalized patients at risk for acute clinical deterioration. A simple (logistic regression based) ML model (eCARTv2) is commercially available within electronic health records on AgileMD’s clinical decision support platform. eCARTv2 was developed in a retrospective multicenter dataset and its use in clinical practice was associated with a 29% relative risk reduction in mortality in a multicenter trial. Our team recently completed development and validation of a gradient boosted machine (GBM) version of the model (eCARTv4), using nearly 100 variables, including trends and interactions. The advanced ML model was significantly more accurate than the simple ML and other models for predicting acute clinical deterioration across all hospital settings, in both septic and non-septic patients as well as in COVID-19 patients. The next challenge is clinically implementing it. The goals of this project are to a) upgrade the existing AgileMD platform to support the previously derived and validated eCARTv4 model and overhaul the human-machine interface for an advanced user experience (UX) that provides, for the first time, interpretable, graphical insight into the contribution of individual variables to a real-time EHR-embedded advanced ML analytic, and b) measure the impact of the new tool on HCP effectiveness, efficiency and satisfaction. We hypothesize that the combination of high accuracy and interpretability afforded by the advanced ML and UX will result in earlier recognition of acute deterioration as well as increased System Usability Scores (SUS) and usefulness scores in the treatment of deteriorating COVID-19 patients over standard care.