A validated model for early prediction of group A streptococcal aetiology in necrotising soft tissue infections using minimal patient data

Abstract Background Necrotising soft tissue infections (NSTI) are life-threatening conditions caused by diverse bacteria. Treatment strategies have remained largely universal and unchanged, and only modest improvements in patient outcomes have been observed. Emerging insights into NSTI pathogenesis may enable more targeted approaches. Because microbial aetiology is central to guiding appropriate therapy, we aimed to develop and externally validate machine learning models capable of predicting microbial aetiology using only data available at an early stage. In parallel, we explored whether similar models could predict selected clinical endpoints related to surgical management, patient handling, and organ support. Methods We used data from the INFECT study, an international multicentre prospective cohort investigating NSTI characteristics and pathogenesis. A total of 409 adults with surgically confirmed NSTI were enrolled between February 2013 and June 2017 from five Scandinavian hospitals. More than 700 clinical variables were collected from hospital admission to intensive care unit entry. Machine learning models were developed to predict the presence of Streptococcus pyogenes (GAS, Group A streptococcus ) and five clinical endpoints: risk of amputation, size of skin defect, maximum skin defect size, length of intensive care (ICU) stay, and need for renal replacement therapy. Unsupervised variable selection was implemented, and Shapley Additive explanations were used for model interpretability. External validation employed a retrospective multicentre cohort of 216 NSTI patients treated in 11 Dutch hospitals between January 2013 and December 2017. Results Eight presurgical variables (age, diabetes, affected area, prior surgical intervention, and blood creatinine and haemoglobin concentrations) were sufficient for predicting GAS aetiology with high discriminatory power. Performance was good in both the development cohort (ROC-AUC 0.828; 95% CI 0.763–0.883) and the external validation cohort (ROC-AUC 0.758; 95% CI 0.696–0.821). Prediction of clinical endpoints related to surgical management, ICU stay, and organ support was unsuccessful. Conclusions We developed and externally validated a model predicting GAS aetiology in NSTI using presurgical data alone. Early identification of GAS may improve clinical handling and support tailored decisions on treatment and infection control, including management of close contacts and reduction of hospital transmission risk.