The Antibiotic Resistance Microbiology Dataset (ARMD) consists of fully de-identified electronic health record data from Stanford Healthcare [1, 2]. To ensure patient privacy and comply with data-sharing policies, the dataset employs the fol- lowing de-identification measures: (i) unique ran- domly generated identifiers for patients and cul- ture orders, (ii) temporal de-identification through jittering of timestamps while preserving temporal relationships, (iii) age censoring with predefined bins and grouping of patients aged 90 years or older, (iv) gender encoding as binary values with- out explicit male/female labels, and (v) exclusion of all direct patient identifiers. All demographic

and clinical details are provided in a de-identified format.

To ensure balanced representation for model development, a targeted sampling strategy was employed. From the full ARMD dataset [1, 2], a cohort of 1,000 patients was selected based on their culture results, comprising 500 with ‘Suscep- tible’ outcomes and 500 with ‘Resistant’ outcomes. These patients corresponded to a total of 997 unique clinical encounters. This predefined cohort served as the foundation for all subsequent data extraction and integration processes.

For each relevant dataset—microbiology cultures cohort, prior medications, demographics, labora- tory results, vital signs, antibiotic class exposure, and comorbidity—records were filtered to include only those associated with the selected cohort. Each file was then preprocessed independently, involving handling of missing values, encoding of categorical variables, and generation of statisti- cal and temporal features. After preprocessing, the individual datasets were merged into a unified file using consistent patient and culture identi- fiers through a series of left joins. This integration preserved all available information for the target cohort and resulted in a final dataset of shape (4,982,642 rows, 47 columns). The distribution of susceptibility categories is shown in Fig. 1.

The final dataset contains a total of approx- imately 5 million records, with 2,697,708 labeled as Susceptible and 2,284,934 as Resistant. This results in a mild class imbalance (approximately 54% vs 46%), which is generally not severe enough to negatively impact model performance. Most modern machine learning algorithms, especially tree-based models like XGBoost and Random For- est, are robust to such small imbalances and can handle them effectively without additional rebalancing techniques. The merged dataset was subsequently split into training and testing sub- sets using an 80/20 ratio, yielding a training set of shape (3,986,113 rows, 46 columns) and a test- ing set of shape (996,529 rows, 46 columns), which were used for model development and evaluation.

The dataset integrates information from multiple clinical domains. Key features were extracted and engineered as follows:

with usable coverage. Models capable of han- dling missing data, such as XGBoost, were used to mitigate the impact of remaining sparsity.

rate, respiratory rate, temperature, systolic and diastolic blood pressure) were provided as pre-aggregated values without raw time- stamped measurements. For each variable, we retained the median, first, and last values when available. We then computed delta features (e.g., delta heartrate) representing the change between the first and last measurements to cap- ture temporal trends. Due to significant sparsity in the first/last readings, most delta features were missing for a large subset of patients. To balance feature richness and sample size, we retained rows with at least one valid delta mea- surement while preserving the robust median summaries. This resulted in a time-aware, yet broadly applicable, representation of physiolog- ical states across patients.

We selected a series of gradient-boosting models and Random Forest due to their effectiveness in handling imbalanced datasets, high-dimensional feature spaces, and missing values. These mod- els are known for their ability to capture complex nonlinear relationships, making them well-suited for predicting antimicrobial resistance (AMR) based on diverse clinical, microbiological, and treatment-related variables [1, 2].

Each model was trained on the ARMD dataset and evaluated using an 80/20 train-test split [1, 2]. The models included XGBoost, LightGBM, Ran- dom Forest, and HistGradientBoostingClassifier. Performance was assessed by classifying resistant

versus susceptible cases using precision, recall, F1-score, and ROC AUC as evaluation metrics.

All analyses were performed using Python (3.11.5) with open-source libraries including scikit-learn, XGBoost, and LightGBM. To ensure reproducibil- ity, experiments were run with fixed random seeds. Class imbalance was addressed in XGBoost using the scale pos weight parameter, while other models were trained with default settings. Large language models (ChatGPT, OpenAI) were used only for language editing and formatting; all out- puts were reviewed and verified by the authors.

Each model was tested on a held-out dataset containing 996,529 records. Table 1 summarizes precision, recall, and F1-score for both classes. In addition, the ROC AUC for each model is reported in Table 2. The comparative ROC curves of the four models are shown in Fig. 5.

demonstrated the highest precision (80%) for resistant cases, suggesting they are better at reducing false positives—important when avoiding unnecessary antibiotic use.

mance, achieving a resistant precision of 75% and recall of 66%. but showed slightly lower

overall performance compared to the boosting models, it was still highly competitive, espe- cially for situations that favored model inter- pretability and robustness over the marginal performance gains.

strong performance, with recall reaching 85% in LightGBM and HistGradientBoosting. This high recall is beneficial for confidently identify- ing patients who are unlikely to need aggressive antimicrobial therapy.

To enhance the interpretability of the model, SHAP analysis (SHapley additive explanations) was applied to the XGBoost model. SHAP values explain the individual features’ contribution to the model’s predictions by measuring the influence on the output probability. Figure 6 is a SHAP sum- mary plot indicating the most influential features for the classification of the antibiotic resistance. The plot reveals that antibiotic type and bacterial organism were among the most influential features in the model’s predictions. aligning with known microbiological principles. Clinical variables, such as median systolic blood pressure (median sysbp), bicarbonate (median hco3) and platelet count (median plt) were also exhibited significant influ- ence. Transplant history (has transplant) emerged as a key risk factor, notably highlighting the effect of immunosuppression on resistance development.

These insights validate the model’s clinical rele- vance and illustrate how SHAP enhances trans- parency, connecting the predictions with medi- cal meaningful features. Such interpretability is important for building trust in AI-driven decision support tools, allowing clinicians to not only rely on predictions but also understand the rationale behind them. especially in critical contexts like antimicrobial resistance.

The consistently high performance of all four mod- els is encouraging for AMR prediction from elec- tronic health records [1, 2]. XGBoost achieved the highest recall (88%) for resistant cases, indicating it correctly identified 88% of true resistant infec- tions. However, its lower precision (71%) suggests 29% of predicted resistant cases were false posi- tives—a trade-off that may justify overtreatment

in high-risk scenarios. LightGBM and HistGradi- entBoosting achieved the highest precision (80%) for resistant cases, meaning only 20% of pre- dicted resistant cases were false positives. This reduces unnecessary antibiotic use, a key goal of stewardship programs aiming to curb resistance. Notably, both models handled null values natively and trained efficiently—HistGradientBoosting’s lower memory overhead may advantage resource- constrained settings. While Random Forest had lower AUC (0.8764) than boosting models, its moderate precision (75%) and recall (66%)—com- bined with inherent interpretability (e.g., feature importance)—make it viable for clinical audits or regulatory contexts where explainability is man- dated, despite marginally lower accuracy. Light- GBM and HistGradientBoosting produced identi- cal performance metrics (precision/recall/AUC), likely due to their shared gradient-boosting foun- dations. However, LightGBM’s histogram-based optimization (evidenced by training logs showing row-wise multi-threading) may offer faster train- ing on datasets exceeding millions of records. All models natively handled missing values with- out imputation—XGBoost and LightGBM assign nulls to optimal branches during splitting, while Random Forest uses surrogate splits. This pre- serves data integrity, critical for EHRs where missingness may reflect clinical decisions (e.g., omitted tests).

The choice of model for clinical deployment should be guided by institutional risk tolerance and treat- ment priorities:

1 in 5 predicted resistant cases is a false alarm, minimizing collateral damage to the microbiome and resistance selection pressure.

Implementation requires integration with clinical workflows:

therapy when resistance is unlikely.

This data-driven approach could reduce mortality from untreated resistance while curbing unneces- sary antibiotic use - addressing both arms of the AMR crisis.