Machine learning for predictive analyses in health: An example of an application to predict death in the elderly in São Paulo, Brazil

Abstract

This study aims to present the stages related to the use of machine learning algorithms for predictive analyses in health. An application was performed in a database of elderly residents in the city of São Paulo, Brazil, who participated in the Health, Well-Being, and Aging Study (SABE) (n = 2,808). The outcome variable was the occurrence of death within five years of the elder’s entry into the study (n = 423), and the predictors were 37 variables related to the elder’s demographic, socioeconomic, and health profile. The application was organized according to the following stages: division of data in training (70%) and testing (30%), pre-processing of the predictors, learning, and assessment of the models. The learning stage used 5 algorithms to adjust the models: logistic regression with and without penalization, neural networks, gradient boosted trees, and random forest. The algorithms’ hyperparameters were optimized by 10-fold cross-validation to select those corresponding to the best models. For each algorithm, the best model was assessed in test data via area under the ROC curve (AUC) and related measures. All the models presented AUC ROC greater than 0.70. For the three models with the highest AUC ROC (neural networks and logistic regression with LAS-SO penalization and without penalization, respectively), quality measures of the predicted probability were also assessed. The expectation is that with the increased availability of data and trained human capital, it will be possible to develop predictive machine learning models with the potential to help health professionals make the best decisions.