Introduction: Hydrofluoric acid (HF) exposure leads to severe and life-threatening injuries, necessitating prompt and accurate prognosis for effective treatment. Despite advancements, machine learning (ML) in predicting outcomes for HF exposure is underexplored. This study aims to develop a robust AI-driven model to forecast hospitalization and mortality in HF exposure cases. Methods: A comprehensive dataset of 163 patients treated for HF burns was utilized, including demographic data, injury specifics, and initial treatment responses. Data, extracted from case series and reports following PRISMA guidelines (1979-2020), were preprocessed to handle missing values and encode categorical variables. Due to significant class imbalance, under-sampling techniques were applied. Multiple ML algorithms, including Logistic Regression, Random Forest, and Gradient Boosting Machines, were evaluated. Models were trained using stratified k-fold cross-validation and optimized through grid search. Statistical analysis was performed using Python (version 3.8) with libraries like scikit-learn, pandas, and numpy. Initial treatments included topical calcium gluconate application, intravenous calcium gluconate, and Hexafluorine irrigation. Results: The Random Forest model, after hyperparameter tuning and data balancing, outperformed others, achieving an accuracy of 93.94%, precision of 94%, recall of 100%, F1-score of 97%, sensitivity of 93%, specificity of 90%, and an ROC AUC of 0.9355. Key predictive features included age, total body surface area (TBSA) affected, and initial treatment type. Older age and higher TBSA were linked to worse outcomes, defined as increased mortality and hospitalization. Topical calcium gluconate was the most effective initial treatment, followed by intravenous calcium gluconate and Hexafluorine irrigation. The confusion matrix showed excellent balance, maintaining high recall without sacrificing precision. Conclusion: The AI-driven model enhances clinical decision-making for HF exposure by providing accurate prognostic information. By addressing class imbalance and leveraging advanced ML techniques, the model improves predictive performance. This tool can help clinicians make informed treatment decisions, optimize resource allocation, and improve patient outcomes.
Shin et al. (Sun,) studied this question.