Machine learning-based predictive clinical model for Shigella spp. infection in children with diarrhea

Diarrheal disease remains a significant cause of morbidity and mortality in children under five years of age in low and middle-income countries. Identifying the etiology of diarrheal episodes represents a significant challenge in contexts with limited access to laboratory diagnostic methods, where therapeutic decisions are often based solely on clinical criteria, such as the presence of blood in the stool recommended by the World Health Organization (WHO). In this scenario, identifying combinations of clinical signs and symptoms could contribute to more precise therapeutic decisions. In this study, we developed and internally validated a predictive model based exclusively on clinical variables to identify episodes attributable to Shigella spp. infection, using as a reference an etiological outcome defined by quantitative molecular methods (qPCR). Secondary data from the Malnutrition-Enteric Diseases (MAL-ED) cohort, a multicenter study conducted in eight countries (2009-2016), with longitudinal follow-up of 1,715 children, were used. Diarrheal episodes were reconstructed from a disease surveillance database. Subsequently, fecal samples were temporally linked to these episodes, allowing the incorporation of molecular etiological data, defining diagnostic positivity for Shigella spp. After eligibility criteria and data processing, a final analytical database was obtained with 3,342 episodes and nine clinical variables (age, sex, blood in stool, bowel movement frequency, diarrhea duration, dehydration, fever, vomiting, and hospitalization) selected after multicollinearity assessment. Five machine learning algorithms were evaluated, with performance estimated by internal validation. Logistic regression showed the best discrimination (AUC = 0.789) and good calibration (Brier score = 0.077). At a cutoff point of 0.46, the model achieved a sensitivity of 0.753 and a specificity of 0.708. In comparison, the WHO score showed inferior performance (AUC = 0.556; sensitivity = 0.147; specificity = 0.965). The high negative predictive value (0.96) highlights the model's ability to exclude cases not attributable to Shigella spp., suggesting potential utility as a tool to support primary clinical diagnostic decision-making and rational use of antibiotics in resource-limited settings.