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

Area of Science:

• Microbiology
• Bioinformatics
• Artificial Intelligence

Background:

• Matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectrometry (MS) offers rich data beyond basic bacterial identification.
• Antimicrobial resistance (AMR) is a growing global health threat requiring rapid diagnostic solutions.

Purpose of the Study:

• To develop a deep learning system (ANTIBIOTIC) for bacterial identification and AMR prediction using MALDI-TOF MS data.
• To assess the performance of deep learning models on both internal and external datasets.
• To create an antibiotic recommendation chatbot integrating bacterial identification, AMR prediction, and large language models.

Main Methods:

• Utilized public (DRIAMS) and hospital (NTUHYL) MALDI-TOF MS datasets (over 89,000 records).
• Developed 26 LightGBM models for bacterial species identification.
• Built 248 Temporal Convolutional Network models for AMR prediction across bacteria-antibiotic combinations.
• Integrated models with a large language model for chatbot functionality.

Main Results:

• Bacterial identification models achieved high median AUCs (0.99 internal, 0.96 external).
• AMR prediction models showed initial high AUC (0.94 internal) but declined on external data (0.55), improving to 0.61 after fine-tuning.
• The ANTIBIOTIC system successfully integrated identification, prediction, and recommendation capabilities.

Conclusions:

• Deep learning effectively extracts valuable information from MALDI-TOF MS for bacterial identification and AMR prediction.
• AMR prediction models necessitate regular updates with recent data to sustain performance.
• The ANTIBIOTIC system demonstrates a promising approach for enhanced clinical microbiology diagnostics.