Predicting Acute Lithium Poisoning with Machine Learning

Assessing the severity of acute lithium toxicity can be challenging. Clinicians need reliable tools to classify serious cases that require intensive intervention versus minor cases that can be managed conservatively. A random forest model trained on U.S. National Poison Data System (NPDS) records offers a robust predictive algorithm in this context. With 2,760 acute overdose records, the model achieves 98% F1-score and nearly perfect precision and sensitivity across outcome classes, marking a significant step for healthcare analytics in toxicology.

Background on Lithium Toxicity and NPDS Data

Lithium is a cornerstone treatment for bipolar disorder but carries a narrow therapeutic index. Acute poisoning can lead to severe neurological and cardiovascular risks when serum levels exceed 1.5 mEq/L. The National Poison Data System (NPDS) aggregates cases from U.S. poison centers, creating a large dataset with 133 binary features and one continuous age variable. This rich database enables training of predictive models to forecast medical outcomes, enhancing clinical decision support.

Model Development and Methods

• Data Imputation and Balancing: Missing symptom data was addressed using Markov Chain Monte Carlo imputation. Synthetic Minority Oversampling Technique (SMOTE) balanced serious and minor outcome classes to avoid bias.
• Feature Selection: Recursive Feature Elimination with Cross-Validation (RFECV) identified top predictors among 131 symptom features and patient age.
• Random Forest Classifier: Ensemble of decision trees optimized with hyperparameters like n_estimators, max_depth, and max_features. Training used 70% of data, with 15% for validation and 15% for testing.

Performance Results

The random forest model outperformed deep learning and other traditional classifiers. On the test set, it reached 98% accuracy and F1-score. For serious outcomes, it achieved 100% precision and 96% recall; for minor outcomes, it achieved 96% precision and 100% recall. The receiver operating characteristic (ROC) curve AUC-ROC reached 0.99, reflecting excellent discrimination.

Interpreting Predictions with SHAP Values

SHAP (SHapley Additive exPlanations) analysis sheds light on feature importance. Top predictors include drowsiness/lethargy, patient age, ataxia, abdominal pain, and electrolyte imbalance. For each case, SHAP values quantify the contribution of each symptom to the final classification, enabling transparent model insights and supporting clinician trust.

Clinical Integration and Impact

Integrating this predictive algorithm into emergency department triage systems can accelerate risk stratification. By highlighting key features through SHAP, the tool guides clinicians toward early interventions for high-risk patients, reduces misclassification, and optimizes allocation of critical care resources. Embedding the model in electronic health records and poison center workflows bridges data-driven insights with patient management.

Conclusion and Future Directions

This study demonstrates the potential of machine learning and digital technologies in medical toxicology. A random forest model leveraging NPDS data predicts acute lithium poisoning outcomes with high accuracy. Future work should validate results across diverse populations and explore real-time deployment. Ethical oversight, data privacy, and continued performance monitoring will be essential to harness AI benefits safely in healthcare settings.

Extending this framework to other toxic exposures, such as organophosphates and methanol, can unify predictive toxicology. Collaboration between AI researchers, toxicologists, and clinicians will refine models and expand their scope. Ensuring ethical considerations and minimizing algorithmic bias will be crucial as predictive models become integral to acute care decision making.

Key points

• Random forest model on NPDS data achieves 98% accuracy and test F1-score.
• SHAP analysis highlights drowsiness, age, ataxia, abdominal pain, and electrolyte imbalance as top predictors.
• Integration into clinical triage systems accelerates risk stratification and reduces misclassification.