Explainable artificial intelligence in ICU admission prediction for methanol poisoning patients: a comparative study of deep and machine learning models.

backgroundMethanol poisoning poses significant challenges due to its rapid progression and high mortality rate, necessitating timely and accurate ICU admission decisions. Explainable artificial intelligence (XAI) offers transparent insights into these decisions, enhancing trust and interpretability in predictive models. This study aims was to compare the effectiveness of deep learning and machine learning models in predicting ICU admissions for methanol poisoning patients using explainable AI techniques.

methodsThis study analyzed a dataset of 897 patient records collected from Loghman Hakim Hospital in Tehran, Iran, involving confirmed cases of methanol poisoning. Among these patients, 202 required ICU admission, whereas 695 did not. To develop predictive models, eight established approaches were implemented, including machine learning methods (Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Decision Tree (DT), and Random Forest (RF)) as well as deep learning architectures (Deep Neural Network (DNN), Feedforward Neural Network (FNN), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN)). To reduce the risk of overfitting, 10-fold cross-validation and systematic hyperparameter tuning were applied. Model interpretability was ensured using Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). The performance of all models was evaluated through multiple metrics, including accuracy, sensitivity, specificity, F1-score, and the area under the ROC curve (AUC).

resultsAmong all DL and ML models, the XGBoost ML model outperformed the others, achieving an accuracy of 92.0%, a precision of 92.0%, a recall of 97.0%, an F1-score of 95.0%, and a AUROC of 92.0%. Among DL models, the FNN emerged as the top performer with an accuracy of 91.0%, precision of 92.0%, recall of 96.0%, F1-score of 94.0%, and AUROC of 90.0%. Generally, ML models exhibited higher accuracy (87% - 92%) than DL models (89% - 91%), with similar precision (75% - 92% for ML, 91% - 92% for DL), recall (90% - 97% for ML, 94% - 96% for DL), F1-scores (82% - 95% for ML, 93% - 94% for DL), and AUROC scores (84% - 92% for ML, 89% - 90% for DL).

conclusionWhile both ML and DL models demonstrated strong performance, ML models, especially XGBoost, proved to be more effective in predicting ICU admissions for methanol poisoning patients. The proposed model is intended to support early ICU admission risk stratification during the initial emergency department assessment and should be used as a decision-support tool rather than a replacement for clinical judgment.