A deep learning system for bacterial identification and resistance prediction from MALDI-TOF data

Matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectrometry (MS) data contain underutilized information beyond species identification. We developed a deep learning system-termed ANTIBIOTIC-that identifies bacterial species, predicts antimicrobial resistance (AMR), and recommends antibiotics directly from raw MALDI-TOF MS data. Using a public dataset (DRIAMS, 2015-2018) and 89,026 MS records from a local hospital dataset (NTUHYL, 2017-2023), we built 26 LightGBM models for identifying the most prevalent bacterial species and 248 Temporal Convolutional Network models for predicting AMR across bacteria-antibiotic combinations. For bacterial identification, models achieved a median area under the curve (AUC) of 0.99 internally and 0.96 on temporally split external data. For AMR prediction, median AUC was 0.94 internally but declined to 0.55 on temporal external data, improving to 0.61 after fine-tuning with recent data. Integrating these models with a large language model enabled an antibiotic recommendation chatbot. These findings demonstrate that deep learning can leverage MALDI-TOF MS data for large-scale bacterial identification and AMR prediction, though AMR models require periodic updating to maintain performance.