Machine learning to predict ICU mortality from vital signs for different diagnoses and COVID-19

ABSTRACT

INTRODUCTION: Vital signs (VS) are important indicators of disease severity and clinical deterioration. However, the predictive scope of VS for ICU mortality is unknown and there are no validated system for early and real-time prediction of ICU mortality from VS data alone. In this study we aimed to develop and validate a Machine Learning (ML) classifier to predict ICU mortality from continuous VS data. METHODS: We used de-identified patient VS data obtained from our eSearch (Philips Healthcare) database to encode 7 continuous VS time series and use of 5 VS monitoring devices. Mean, standard deviation, autocorrelation, and the trend of the mean were used to encode VS time series variations and were adjusted to the entire ICU stay, and 6, 12, or 24 hours before death. Our approach did not encode diagnoses but agnostically classified based on VS features. Performance of the models was determined on a naïve cohort and an independent sample of patients with COVID-19. RESULTS: A total of 19,266 ICU stays prior to COVID were studied including 17,339 in the training cohort, and 1,927 in the naïve validation cohort with ICU mortalities of 9%. An independent sample of 548 patients with COVID-19 with mortality of 22% was also used for validation. For the entire stay, and 6-, 12-, and 24-hours in advance, the ML classifier achieved AUCs and PRCs of 0.97 - 0.81 and 0.78 - 0.40, respectively in the naïve population obtained prior to COVID, and AUCs and PRCs of 0.92 - 0.80 and 0.81 - 0.58, respectively, for the COVID cohort. Notably, a differential ranking of features was found for mortality predictions in the COVID-19 sample, as well as in 9 other specific diagnoses. The effectiveness of this approach compared favorably with six other ML methods and with the DRS (Philips) mortality predictions. CONCLUSIONS: A data-driven ML algorithm developed from composite vital sign data alone made ICU mortality predictions with model performance on a naïve ICU test population, as well as on a COVID-19 patient population, that rivals other prediction models using more complex data domains. Shapley Additive exPlanations provided interpretability and clinical validation of the ML model related to the specific features in the ICU subpopulations.