Predicting Mortality in Diabetic ICU Patients Using Machine Learning and Severity Indices.

Area of Science:

• Medical Informatics
• Clinical Research
• Predictive Analytics

Background:

• Diabetes is a major health concern linked to severe complications like renal, cardiovascular, and neuropathic issues.
• These complications increase healthcare costs, ICU admissions, and mortality risk.
• Limited research exists on predicting ICU outcomes based on diabetes, its management, and comorbidities.

Purpose of the Study:

• To develop and validate predictive models for mortality risk in intensive care unit (ICU) patients with diabetes.
• To identify key factors influencing mortality risk among diabetic ICU patients.
• To compare the efficiency of a reduced variable set against complex machine learning models.

Main Methods:

• Utilized the MIMIC-III database for patient data.
• Applied machine learning and binomial logistic regression modeling.
• Developed predictive models to assess mortality risk.

Main Results:

• The developed models demonstrated good predictive performance with AUC values of 0.787 and 0.785.
• A concise set of five variables (HbA1c, mean glucose, admission diagnoses, age, admission type) effectively predicted risk.
• This five-variable model achieved robust classification, outperforming models requiring up to 35 variables.

Conclusions:

• Predictive modeling can accurately assess mortality risk in diabetic ICU patients.
• A simplified approach using five key variables offers efficient and effective risk prediction.
• This study highlights the importance of specific diabetes-related and demographic factors in ICU patient outcomes.