Improving the performance of machine learning algorithms for health outcomes predictions in multicentric cohorts

Machine learning algorithms are being increasingly used in healthcare settings but their generalizability between different regions is still unknown. This study aims to identify the strategy that maximizes the predictive performance of identifying the risk of death by COVID-19 in different regions of a large and unequal country. This is a multicenter cohort study with data collected from patients with a positive RT-PCR test for COVID-19 from March to August 2020 (n = 8477) in 18 hospitals, covering all five Brazilian regions. Of all patients with a positive RT-PCR test during the period, 2356 (28%) died. Eight different strategies were used for training and evaluating the performance of three popular machine learning algorithms (extreme gradient boosting, lightGBM, and catboost). The strategies ranged from only using training data from a single hospital, up to aggregating patients by their geographic regions. The predictive performance of the algorithms was evaluated by the area under the ROC curve (AUROC) on the test set of each hospital. We found that the best overall predictive performances were obtained when using training data from the same hospital, which was the winning strategy for 11 (61%) of the 18 participating hospitals. In this study, the use of more patient data from other regions slightly decreased predictive performance. However, models trained in other hospitals still had acceptable performances and could be a solution while data for a specific hospital is being collected.