Multi task learning based early prediction model for antibiotic resistance using multi institutional cohort data.

Area of Science:

• Medical Informatics
• Computational Biology
• Infectious Diseases

Background:

• Antibiotic resistance is a critical global health threat.
• Inappropriate antibiotic use accelerates resistance development.
• High-cost data labeling hinders medical research.

Purpose of the Study:

• To develop and compare machine learning models for predicting antibiotic resistance.
• To address the challenge of high-cost data labeling in medical records.
• To evaluate multi-task learning approaches against single-task models.

Main Methods:

• Retrospective study using electronic medical records from 59,551 patients.
• Development of single-task learning (Logistic Regression, XGBoost, LightGBM, CatBoost, MLP) and multi-task learning (Hard, Soft Parameter Sharing) models.
• Performance evaluation using Area Under the ROC Curve (AUC), Precision-Recall Curve (PRC), and calibration slope.

Main Results:

• The hard parameter sharing multi-task learning model achieved the highest AUC (79.63) and PRC (80.26) for five of nine antibiotic classes in external validation.
• Shapley Additive Explanations identified previous antibiotic resistance as the key predictor.
• Model performance varied by subgroup: hard parameter sharing excelled with prior culture data, while soft parameter sharing was better without it.

Conclusions:

• Multi-task learning models offer generalized, stable, and superior performance in predicting antibiotic resistance compared to traditional models.
• This approach provides a novel solution for handling partial targets in medical datasets.
• The findings support the use of advanced machine learning for combating antibiotic resistance.