OPTIMIZING TRANSFORMER-BASED LEARNING MODEL WITH TABTRANSFORMER FOR PREDICTING ANTIBIOTIC SUSCEPTIBILITY FROM MICROBIOLOGY MEDICAL RECORDS

Antimicrobial Resistance (AMR) has become a growing threat due to the increase in infections that are unresponsive to conventional therapies. Therefore, the development and optimization of Transformer-based Deep Learning using TabTransformer was employed to model the complex interactions between categorical features. This model was trained to predict antibiotic susceptibility at the individual culture level using the Antibiotic Resistance Microbiology Dataset (ARMD). To address the challenge of highly imbalanced data, the methodology applied includes extensive feature engineering to create historical and clinical variables, as well as the use of Focal Loss during training. After optimization, the final model demonstrated excellent discriminatory ability, with an Area Under the ROC Curve (AUC-ROC) of 0.93 and balanced classification performance, yielding a macro average F1-score of 0.82. Interpretability analysis using SHAP confirmed that patient clinical history and prior drug exposure were the most dominant predictive factors. These findings suggest that the Transformer-based Deep Learning architecture using TabTransformer, combined with clinically relevant feature engineering, can produce a reliable and evidence-based predictive tool