Machine learning for COVID-19 mortality prediction: enhancing cart models with node-specific odds ratios

The objective of the study is to evaluate a classification and regression tree (CART) model combined with odds ratio (OR) to identify key predictors of COVID-19 mortality. Data from 1,432 patients hospitalized at Hospital General de México during the pandemic's 1^st year were analyzed.

A CART model was constructed using demographic, clinical, and laboratory data collected at admission. The model's performance was evaluated using multiple criteria, and ORs were calculated for each node to measure mortality.

Mechanical ventilation emerged as the strongest predictor of mortality. Patients intubated at admission with hospital stays under 17.5 days had the highest mortality rate (99%, OR = 80). Conversely, non-intubated patients hospitalized over 5.5 days with glomerular filtration rates above 66.5 had the lowest mortality (5%, OR = 0.26). The model showed excellent performance, with an F1 score of 0.918, accuracy of 0.903, and area under the curve of 0.955.

The CART model, combined with OR calculations, offers a reliable and comprehensible tool for predicting COVID-19 mortality risk. This model provides practical utility in various healthcare settings, including resource-limited contexts, by focusing on readily available clinical parameters. The results emphasize the significant influence of mechanical ventilation on patient outcomes and the need for timely interventions in high-risk patients.