Generalizable machine learning models for rapid antimicrobial resistance prediction in unseen health care settings

The deployment of machine learning in clinical settings is often hindered by the limited generalizability of the models. Models that perform well during development tend to underperform in new environments, limiting their clinical utility. This issue affects models designed for the rapid identification of antimicrobial resistance, which is essential to guide treatment decisions. Traditional susceptibility tests can take up to 3 days, whereas integrating matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectrometry with machine learning has the potential to reduce this to 1 day. However, model performance declines drastically in hospitals or time frames outside the training data.

To improve robustness, we develop advanced feature representations using masked autoencoders (MAEs) for MALDI-TOF spectra and chemical language models and SELF-referencing embedded strings (SELFIES) for antimicrobials. Cross-validated on data from 4 medical institutions, our models demonstrate improved performance and stability. The MAE and SELFIES encodings increase the area under the precision-recall curve by 4% when evaluated on unseen time periods, while the MAE and Molformer language model encodings improve it by 10% when applied across different hospitals.

These results underscore the value of combining deep learning with chemical and spectral information to build generalizable, high-impact clinical artificial intelligence.