Mortality prediction in patients with hyperglycaemic crisis using explainable machine learning: a prospective,...

Area of Science:

• Medical Informatics
• Clinical Prediction Modeling
• Machine Learning in Healthcare

Background:

• Hyperglycaemic crisis significantly increases short- and long-term mortality risk.
• Individualized risk assessment is crucial for managing patients post-admission.
• Predictive modeling can aid in identifying high-risk individuals.

Purpose of the Study:

• To develop an explainable machine learning model for predicting 3-year mortality in hyperglycaemic crisis patients.
• To provide individualized risk factor assessment for these patients.
• To enhance understanding of key mortality predictors.

Main Methods:

• Trained and validated five machine learning algorithms on hyperglycaemic crisis patient data (2016-2020).
• Employed tenfold cross-validation for internal validation and external validation on unseen data.
• Utilized SHapley Additive exPlanations (SHAP) for model interpretability and feature importance analysis.

Main Results:

• The Light Gradient Boosting Machine model demonstrated superior performance (AUC 0.89).
• Key predictors for 3-year mortality included advanced age, elevated blood glucose, and high blood urea nitrogen.
• A total of 337 patients were analyzed, with a 3-year mortality rate of 13.6%.

Conclusions:

• An explainable AI model can accurately predict mortality and visualize individual risk factors.
• Advanced age, metabolic disorders, and impaired renal/cardiac function are critical for non-survival.
• The model offers valuable insights for personalized patient management and intervention strategies.