Development and validation of a simple machine learning tool to predict mortality in leptospirosis

Abstract

Predicting risk factors for death in leptospirosis is challenging, and identifying high-risk patients is crucial as it might expedite the start of life-saving supportive care. Admission data of 295 leptospirosis patients were enrolled, and a machine-learning approach was used to fit models in a derivation cohort. The comparison of accuracy metrics was performed with two previous models—SPIRO score and quick SOFA score. A Lasso regression analysis was the selected model, demonstrating the best accuracy to predict mortality in leptospirosis [area under the curve (AUC-ROC) = 0.776]. A score-based prediction was carried out with the coefficients of this model and named LeptoScore. Then, to simplify the predictive tool, a new score was built by attributing points to the predictors with importance values higher than 1. The simplified score, named QuickLepto, has five variables (age > 40 years; lethargy; pulmonary symptom; mean arterial pressure < 80 mmHg and hematocrit < 30%) and good predictive accuracy (AUC-ROC = 0.788). LeptoScore and QuickLepto had better accuracy to predict mortality in patients with leptospirosis when compared to SPIRO score (AUC-ROC = 0.500) and quick SOFA score (AUC-ROC = 0.782). The main result is a new scoring system, the QuickLepto, that is a simple and useful tool to predict death in leptospirosis patients at hospital admission.