Machine Learning Methods to Predict Diabetes Complications.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining
• Health Informatics

Background:

• Machine learning (ML) algorithms extract patterns from data, enhancing data mining pipelines.
• Predictive modeling for type 2 diabetes mellitus (T2DM) complications is crucial for patient management.
• Electronic health records (EHRs) offer a rich data source for developing predictive models.

Purpose of the Study:

• To develop and validate predictive models for T2DM complications using a data mining pipeline.
• To forecast the onset of retinopathy, neuropathy, and nephropathy at 3, 5, and 7 years post-initial hospital visit.
• To create specialized, clinically translatable models for T2DM complication prediction.

Main Methods:

• Utilized a data mining pipeline within the EU-funded MOSAIC project.
• Handled missing data using Random Forest (RF) and addressed class imbalance.
• Employed Logistic Regression with stepwise feature selection for predictive modeling.
• Included variables: gender, age, time from diagnosis, BMI, HbA1c, hypertension, and smoking habit.

Main Results:

• Achieved a maximum prediction accuracy of 0.838 for T2DM complications.
• Identified distinct sets of predictive variables for each complication and time scenario.
• Developed specialized models tailored to specific complications (retinopathy, neuropathy, nephropathy).

Conclusions:

• The developed ML models accurately predict T2DM complications.
• Specialized models derived from EHR data are easily translatable to clinical practice.
• This approach supports proactive patient management and personalized T2DM care.