JMIR Publications

ABSTRACT

Background:

Machine learning has been extensively applied in medical prediction. However, existing machine-driven rational drug use review systems predominantly rely on codifiable information while neglecting unstructured data such as electronic health records (EHRs) documented in natural language, leading to the oversight of potential medication-related risks.

Objective:

To address the aforementioned limitations, this study proposes and validates a text classification-based machine learning framework for the automatic identification of "prophylactic" and "non-prophylactic" medication indications in electronic health records (EHRs), thereby facilitating the assessment of the rationality of prophylactic proton pump inhibitor (PPI) use.

Methods:

In this study, a text annotation approach was adopted, wherein clinical pharmacists performed paragraph-by-paragraph annotation on 3695 medical record segments. Six algorithms—logistic regression, support vector machine (SVM), random forest, decision tree, LightGBM, and CatBoost—were utilized for model training. To address the class imbalance issue in the dataset, SMOTE (Synthetic Minority Oversampling Technique) was employed for oversampling to balance the sample distribution. The model performance was evaluated on an independent test set with an 8:2 training-to-test split ratio. The trained model was then used to validate the test set, predicting whether the text content was indicative of preventive intent or non-preventive intent, thereby determining the rationality of the clinical prophylactic use of proton pump inhibitors (PPIs). Finally, the SHAP (SHapley Additive exPlanations) method was applied to interpret the model's decision-making process.

Results:

The results indicated that all six trained machine learning models exhibited satisfactory performance in the predictive classification of text. Among these models, LightGBM and CatBoost, following SMOTE oversampling, achieved the optimal performance.

Conclusions:

In conclusion, the natural language processing (NLP) text classifier trained based on machine learning can effectively capture medication indications in medical records, providing high-reliability decision support for rational drug use review. Clinical Trial: Not Applicable.This study is a retrospective data analysis based on electronic health records, without involving human subjects or interventional clinical trials, thus trial registration is not applicable.