Prediction of COVID-19 in-hospital mortality in older patients using artificial intelligence: a multicenter study

Abstract

Background: Once the pandemic ended, SARS-CoV-2 became endemic, with flare-up phases. COVID-19 disease can still have a significant clinical impact, especially in older patients with multimorbidity and frailty.

Objective: This study aims at evaluating the main characteristics associated to in-hospital mortality among data routinely collected upon admission to identify older patients at higher risk of death.

Methods: The present study used data from Gerocovid-acute wards, an observational multicenter retrospective-prospective study conducted in geriatric and internal medicine wards in subjects ≥60 years old during the COVID-19 pandemic. Seventy-one routinely collected variables, including demographic data, living arrangements, smoking habits, pre-COVID-19 mobility, chronic diseases, and clinical and laboratory parameters were integrated into a web-based machine learning platform (Just Add Data Bio) to identify factors with the highest prognostic relevance. The use of artificial intelligence allowed us to avoid variable selection bias, to test a large number of models and to perform an internal validation.

Results: The dataset was split into training and test sets, based on a 70:30 ratio and matching on age, sex, and proportion of events; 3,520 models were set out to train. The three predictive algorithms (optimized for performance, interpretability, or aggressive feature selection) converged on the same model, including 12 variables: pre-COVID-19 mobility, World Health Organization disease severity, age, heart rate, arterial blood gases bicarbonate and oxygen saturation, serum potassium, systolic blood pressure, blood glucose, aspartate aminotransferase, PaO2/FiO2 ratio and derived neutrophil-to-lymphocyte ratio.

Conclusion: Beyond variables reflecting the severity of COVID-19 disease failure, pre-morbid mobility level was the strongest factor associated with in-hospital mortality reflecting the importance of functional status as a synthetic measure of health in older adults, while the association between derived neutrophil-to-lymphocyte ratio and mortality, confirms the fundamental role played by neutrophils in SARS-CoV-2 disease.

Keywords: COVID-19; artificial intelligence; in-hospital mortality; mobility; neutrophil-to-limphocyte ratio.

FIGURE 1

(A) C-statistic distributions and (B) predictive performance drop for the mortality prediction model computed on the training dataset. (C) Kaplan-Meier survival estimates for the training dataset, stratified according to categories of mortality risk.

FIGURE 2

Risk prediction scores for the 12 predictors included in the training model. Figure legend: *p < 0.05, **p < 0.01, ***p < 0.001 vs. low-risk, walks independently, or WHO disease severity class 4; *p < 0.05, **p < 0.01, ***p < 0.001 vs. intermediate-risk, assisted mobility or WHO disease severity class 5 for Dunn’s post-hoc analysis.

FIGURE 3

Model validation: (A) C-statistic (in green), (B) Kaplan-Meier survival function.