Machine learning for predicting medical outcomes associated with acute lithium poisoning

Area of Science:

• Medical informatics
• Toxicology
• Machine learning in healthcare

Background:

• Artificial intelligence and machine learning show promise in predicting medical outcomes.
• Acute lithium toxicity poses significant health risks, necessitating accurate outcome prediction.

Purpose of the Study:

• To evaluate the effectiveness of the random forest algorithm for predicting medical outcomes in acute lithium toxicity.
• To assess the model's predictive performance using established metrics.

Main Methods:

• Analysis of 2,760 acute lithium overdose cases from the National Poison Data System (NPDS) (2014-2018).
• Development and validation of a random forest model to predict serious (major effect, death) and minor medical outcomes.
• Performance evaluation using accuracy, recall (sensitivity), and F1-score; SHapley Additive exPlanations (SHAP) for feature importance.

Main Results:

• The random forest model achieved high accuracy (99% training, 98% validation, 98% test) and F1-scores.
• Exceptional performance for predicting important outcomes (100% accuracy, 96% sensitivity) and minor outcomes (96% accuracy, 100% sensitivity).
• Key predictive factors identified by SHAP include drowsiness/lethargy, age, ataxia, abdominal pain, and electrolyte abnormalities.

Conclusions:

• The random forest algorithm demonstrates high accuracy (98%) and sensitivity in predicting medical outcomes for acute lithium intoxication.
• The model effectively distinguishes between significant and minor outcomes, offering potential clinical utility.
• Further research is recommended to validate these predictive capabilities.