Risk sharing arrangements between hospitals and payers together with penalties imposed by the Centers for Medicare and Medicaid (CMS) are driving an interest in decreasing early readmissions. There are a number of published risk models predicting 30day readmissions for particular patient populations, however they often exhibit poor predictive performance and would be unsuitable for use in a clinical setting. In this work we describe and compare several predictive models, some of which have never been applied to this task and which outperform the regression methods that are typically applied in the healthcare literature. In addition, we apply methods from deep learning to the five conditions CMS is using to penalize hospitals, and offer a simple framework for determining which conditions are most cost effective to target.

Summary An emphasis on overly broad notions of generalisability as it pertains to applications of machine learning in health care can overlook situations in which machine learning might provide clinical utility. We believe that this narrow focus on generalisability should be replaced with wider considerations for the ultimate goal of building machine learning systems that are useful at the bedside.

Background Successful integrations of machine learning into routine clinical care are exceedingly rare, and barriers to its adoption are poorly characterized in the literature. Objective This study aims to report a quality improvement effort to integrate a deep learning sepsis detection and management platform, Sepsis Watch, into routine clinical care. Methods In 2016, a multidisciplinary team consisting of statisticians, data scientists, data engineers, and clinicians was assembled by the leadership of an academic health system to radically improve the detection and treatment of sepsis. This report of the quality improvement effort follows the learning health system framework to describe the problem assessment, design, development, implementation, and evaluation plan of Sepsis Watch. Results Sepsis Watch was successfully integrated into routine clinical care and reshaped how local machine learning projects are executed. Frontline clinical staff were highly engaged in the design and development of the workflow, machine learning model, and application. Novel machine learning methods were developed to detect sepsis early, and implementation of the model required robust infrastructure. Significant investment was required to align stakeholders, develop trusting relationships, define roles and responsibilities, and to train frontline staff, leading to the establishment of 3 partnerships with internal and external research groups to evaluate Sepsis Watch. Conclusions Machine learning models are commonly developed to enhance clinical decision making, but successful integrations of machine learning into routine clinical care are rare. Although there is no playbook for integrating deep learning into clinical care, learnings from the Sepsis Watch integration can inform efforts to develop machine learning technologies at other health care delivery systems.

IMPORTANCE The ability to accurately predict in-hospital mortality for patients at the time of admission could improve clinical and operational decision-making and outcomes. Few of the machine learning models that have been developed to predict in-hospital death are both broadly applicable to all adult patients across a health system and readily implementable. Similarly, few have been implemented, and none have been evaluated prospectively and externally validated.OBJECTIVES To prospectively and externally validate a machine learning model that predicts in-hospital mortality for all adult patients at the time of hospital admission and to design the model using commonly available electronic health record data and accessible computational methods. DESIGN, SETTING, AND PARTICIPANTSIn this prognostic study, electronic health record data from a total of 43 180 hospitalizations representing 31 003 unique adult patients admitted to a quaternary academic hospital (hospital A) from October 1, 2014, to December 31, 2015, formed a training and validation cohort. The model was further validated in additional cohorts spanning from March 1, 2018, to August 31, 2018, using 16 122 hospitalizations representing 13 094 unique adult patients admitted to hospital A, 6586 hospitalizations representing 5613 unique adult patients admitted to hospital B, and 4086 hospitalizations representing 3428 unique adult patients admitted to hospital C. The model was integrated into the production electronic health record system and prospectively validated on a cohort of 5273 hospitalizations representing 4525 unique adult patients admitted to hospital A between February 14, 2019, and April 15, 2019. MAIN OUTCOMES AND MEASURES The main outcome was in-hospital mortality. Model performance was quantified using the area under the receiver operating characteristic curve and area under the precision recall curve. RESULTS A total of 75 247 hospital admissions (median [interquartile range] patient age, 59.5 [29.0]years; 45.9% involving male patients) were included in the study. The in-hospital mortality rates for the training validation; retrospective validations at hospitals A, B, and C; and prospective validation cohorts were 3.0%, 2.7%, 1.8%, 2.1%, and 1.6%, respectively. The area under the receiver operating characteristic curves were 0.87 (95% CI, 0.83-0.89), 0.85 (95% CI, 0.83-0.87), 0.89 (95% CI, 0.86-0.92), 0.84 (95% CI, 0.80-0.89), and 0.86 (95% CI, 0.83-0.90), respectively. The area under the precision recall curves were 0.29 (95% CI, 0.25-0.37), 0.17 (95% CI, 0.13-0.22), 0.22 (95% CI, 0.14-0.31), 0.13 (95% CI, 0.08-0.21), and 0.14 (95% CI, 0.09-0.21), respectively. CONCLUSIONS AND RELEVANCEProspective and multisite retrospective evaluations of a machine learning model demonstrated good discrimination of in-hospital mortality for adult patients at the (continued) Key Points Question How accurately can a machine learning model predict risk of in-hospital mortality for adult patients when evaluated prospectively and externally? Findings In this p...