A validated model for early prediction of group A streptococcal aetiology and clinical endpoints in necrotising soft tissue infections

Abstract

ObjectivesTo develop and externally validate machine learning models for predicting microbial aetiology and clinical endpoints, encompassing surgery, patient management, and organ support in Necrotising Soft Tissue Infections (NSTI).MethodsPredictive models for the presence of Group A Streptococcus (GAS) and for five clinical endpoints (risk of amputation, size of skin defect, maximum skin defect size, length of ICU stay, and need for renal replacement therapy) were built and trained using data from the prospective, international INFECT cohort (409 patients, 2013-2017), implementing unsupervised variable selection, and comparing several algorithms. SHapley Additive exPlanations (SHAP) analysis was used to interpret the model. GAS predictive models were externally validated using data from a Dutch retrospective multicenter cohort from the same calendar period (216 patients).ResultsEight variables available pre-surgery (age, diabetes, affected anatomical locations, prior surgical interventions, and creatinine and haemoglobin levels) sufficed for prediction of GAS aetiology with high discriminatory power in both the development (ROC-AUC: 0.828; 95%CI 0.763, 0.883) and validation cohort (ROC-AUC: 0.758; 95%CI 0.696, 0.821). The prediction of clinical endpoints related to surgical, patient management, and organs support aspects was unsuccessful.ConclusionAn externally validated prediction model for GAS aetiology before organ support aspects was unsuccessful, having implications for targeted treatment decisions of NSTI.