Improving diabetes disease patients classification using stacking ensemble method with PIMA and local healthcare data

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Computational Biology

Background:

• Diabetes mellitus is a global health challenge with a high rate of undiagnosed cases.
• Early diagnosis and management are crucial for preventing or delaying diabetes complications.
• Machine learning and deep learning offer promising avenues for improved disease detection.

Purpose of the Study:

• To develop and evaluate stacking-based ensemble models for early diabetes mellitus classification.
• To improve classification accuracy and robustness by combining predictions from multiple models.
• To assess the performance of proposed models on diverse datasets, including real-world clinical data.

Main Methods:

• Proposed two stacking-based ensemble models: classical and deep neural network (NN) stacking.
• Utilized the PIMA Indian diabetes dataset, simulated data, and local healthcare facility data.
• Employed train-test split and cross-validation (CV) for model validation.

Main Results:

• The highest accuracy achieved was 95.50% with a stacking ensemble of three NN architectures using 5-fold CV on simulated data.
• For the Pima Indian Diabetes dataset, stacked accuracy reached 75.03% (train-test) and 77.10% (CV).
• Proposed methods demonstrated high performance on the primary dataset, with accuracy ranging from 92% to 95% and F1-scores from 88% to 96%.

Conclusions:

• The developed stacking ensemble models show significant potential for accurate and robust early diabetes detection.
• The proposed methods and datasets can advance machine learning applications in healthcare for diabetes management.
• This work contributes to the early identification and treatment of diabetes, improving patient outcomes.