Aim: Predicting mortality is important for intensivists, yet conventional disease severity scores may not consistently predict mortality in patients with Coronavirus Disease 2019 (COVID-19). We aimed to develop a machine learning-based mortality prediction model for COVID-19 patients admitted to the intensive care unit (ICU).
Study Design: This study employs a retrospective and prospective longitudinal design. We retrospectively screened a total of 436 COVID-19 patients admitted to the ICU between March 15, 2020, and December 31, 2021. The worst laboratory results and vital signs within the first 24 hours of ICU admission were recorded. We selected 29 inputs to develop a model using machine learning (ML), employing an artificial neural network (ANN) as the decision model. For model testing, we prospectively followed 108 patients from January 1, 2022, to March 31, 2022.
Results: Our model predicted mortality with an 88% sensitivity and specificity. Conventional disease severity scores predicted mortality with lower sensitivity and specificity than our model did: 71% sensitivity and 70% specificity for the Acute Physiology and Chronic Health Evaluation II (APACHE-2), and 75% sensitivity and 75% specificity for both the Simplified Acute Physiology Score II (SAPS-2) and APACHE-4. Our model demonstrated greater discriminative power for mortality with an area under the curve (AUC) of 0.93 (95% confidence interval [CI], 0.87-0.98) compared to conventional disease severity scores. Respiratory support within the first 24 hours of ICU admission was identified as the most important factor affecting mortality.
Conclusions: In scenarios such as epidemics, where conventional disease scores fall short in predicting mortality, machine learning models can be developed to reliably forecast disease outcomes.