Machine Learning Model with Fourier-Transform Infrared Spectroscopy (FTIR) as a Proof-of-Concept Tool for Predicting Group A Streptococcus (GAS) emm-Type in the Pediatric Population

Background: Since 2022, invasive Group A Streptococcus (GAS) infections have increased, mainly due to the spread of specific emm-types, such as emm1. As therapy may depend on the emm-type, rapid and cost-effective identification is crucial. Fourier-transform infrared spectroscopy (FTIR) has emerged as a promising alternative to sequencing for GAS typing. We applied machine learning (ML) to FTIR spectra to build a predictive model for emm-type identification. Methods: Twenty-four GAS strains were analyzed by whole-genome sequencing and FTIR. The model was trained on twenty-one strains (emm-types: 1, 3, 4, and 6), using leave-one-out cross validation (LOOCV). To test the model's ability to avoid false positive results, the model was also tested with three strains belonging to emm-types not included in the training of the model (emm-types: 12, 89, and 75). Results: An artificial neural network trained for 400 cycles achieved the highest accuracy (90.7%) out of the thirteen different models tested. When the three strains belonging to emm-types not included in the model were predicted with this model, it produced low score values, confirming its ability to avoid false positive results. Conclusions: We developed a preliminary and proof-of-concept model capable of accurately predicting the four most-prevalent emm-types in our setting, including the highly virulent emm1. These findings support FTIR combined with ML as a rapid, low-cost tool for GAS typing, with potential for real-time clinical applications to guide timely treatment decisions. However, as a proof-of-concept study, the relatively small sample size and limited emm-type diversity underline the need for further validation with larger and more diverse datasets.