Prediction of antimicrobial resistance from MALDI-TOF mass spectra using machine learning: a validation study.

Area of Science:

• Clinical microbiology and infectious diseases
• Computational biology and bioinformatics
• Analytical chemistry and mass spectrometry

Background:

• Matrix-Assisted Laser Desorption-Ionization-Time of Flight (MALDI-TOF) mass spectrometry is a rapid method for bacterial identification.
• Predicting antimicrobial resistance (AMR) using MALDI-TOF data with machine learning (ML) can expedite antimicrobial susceptibility testing (AST).
• Early AMR prediction is vital for timely, targeted antibiotic treatment, potentially reducing mortality in severe infections.

Purpose of the Study:

• To validate the performance of ML models for AMR prediction using MALDI-TOF data from routine diagnostics and public databases.
• To assess the long-term performance and stability of these ML models over an 18-month period.
• To determine the optimal strategies for training and retraining ML models to ensure sustained accuracy in AMR prediction.

Main Methods:

• Collected and analyzed MALDI-TOF mass spectra from Escherichia coli (n=7,897), Klebsiella pneumoniae (n=2,444), and Staphylococcus aureus (n=4,664).
• Trained and cross-validated six distinct ML classification models (LR, MLP, SVM, RF, LGBM, XGB) for AMR prediction.
• Prospectively monitored model performance over 18 months post-training, evaluating data from both local (Germany) and external (Switzerland) sources.

Main Results:

• ML models demonstrated comparable performance (AUROC ≥ 0.8) on both local and public datasets for predicting specific AMR patterns.
• Optimal prediction was achieved for oxacillin resistance in S. aureus (RF, 0.85), ciprofloxacin in E. coli (XGB, 0.83), and piperacillin-tazobactam in K. pneumoniae (XGB, 0.81).
• Model performance significantly decreased when training and test data were geographically or temporally dissimilar, highlighting the need for local adaptation and regular retraining.

Conclusions:

• ML-based AMR prediction using MALDI-TOF spectra shows significant promise for clinical diagnostics, achieving good predictive performance (AUROC ≥ 0.8).
• The accuracy of ML models is highly dependent on the training data's locality and temporal relevance.
• Regular retraining of ML classifiers with up-to-date, local MALDI-TOF data is essential to maintain high performance levels for AMR prediction.