From raw data to actionable insights: preprocessing real-world data for machine learning in diabetes care.

Area of Science:

• Healthcare Data Science
• Clinical Informatics
• Machine Learning in Medicine

Background:

• Real-World Data (RWD) requires extensive preprocessing for machine learning (ML) readiness.
• Applying ML to RWD holds potential for improving patient management in healthcare.
• Type 2 diabetes patient data over 20 years was utilized to explore preprocessing impacts.

Purpose of the Study:

• To investigate the effect of data preprocessing pipelines on ML models applied to RWD.
• To evaluate ML model performance and explanations across different preprocessing strategies.
• To identify predictors of HbA1c decrease in type 2 diabetes patients using ML.

Main Methods:

• Three distinct data preprocessing pipelines were established.
• Logistic Regression (LR), XGBoost, and Decision Tree Classifier (DTC) models were applied and tuned for precision.
• Shapley Additive Explanations (SHAP) were used for model interpretability.

Main Results:

• No significant variation in precision scores for LR and XGBoost across experimental conditions.
• Decision Tree Classifier (DTC) performance improved with missing data imputation.
• Exploratory Data Analysis and SHAP confirmed that worse baseline values predict HbA1c decrease.

Conclusions:

• ML model performance and explanations were robust across preprocessing variations.
• Worse baseline health metrics are significant predictors of HbA1c decrease in type 2 diabetes patients.
• ML insights from RWD can inform clinical discussions and enhance patient management strategies.