A robust and generalized framework in diabetes classification across heterogeneous environments.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Computational Epidemiology

Background:

• Diabetes mellitus (DM) poses a significant global health burden, with increasing incidence linked to lifestyle factors.
• Accurate and timely DM prediction is vital for women, especially during pregnancy and postpartum, to prevent complications.
• Existing predictive models often lack external validity across diverse populations and datasets.

Purpose of the Study:

• To develop and validate a robust machine learning (ML) framework for diabetes mellitus (DM) prediction.
• To assess the generalizability and performance of ML models across heterogeneous datasets (PIMA and BD).
• To investigate the impact of intra-dataset, inter-dataset, and partial fusion validation techniques on prediction accuracy.

Main Methods:

• Utilized two distinct datasets (PIMA and BD) for training and testing ML models.
• Employed intra-dataset, inter-dataset, and partial fusion validation strategies.
• Evaluated various ML models including Extreme Gradient Boosting (XGBoost), Random Forest (RF), Gradient Boosting (GB), and deep learning (DL).

Main Results:

• Intra-dataset validation: XGBoost (79% on PIMA), RF/GB (approx. 99% on BD).
• Inter-dataset validation: Ensemble model achieved 88% (PIMA-trained, BD-tested), but performance dropped to 74% (BD-trained, PIMA-tested).
• Partial fusion validation: DL model reached 98% accuracy when PIMA data was combined with 30% of BD data.

Conclusions:

• Dataset diversity and partial data fusion significantly enhance ML model robustness and generalizability for DM prediction.
• The proposed ML framework provides valuable insights into predicting diabetes across varied populations.
• Effective diabetes prediction strategies require careful consideration of data heterogeneity and fusion techniques.