Enhanced Multi-Model Machine Learning-Based Dementia Detection Using a Data Enrichment Framework: Leveraging the...

Area of Science:

• Computational medicine and health informatics.
• Machine learning applications in clinical diagnostics.

Background:

• Early dementia diagnosis is crucial for effective intervention in older adults.
• Dementia often co-occurs with cardiovascular diseases, complicating diagnosis.
• Existing diagnostic methods may struggle to differentiate these conditions effectively.

Purpose of the Study:

• To develop a multiclass classification framework to distinguish dementia from cardiovascular comorbidities (heart failure, aortic valve disorder).
• To leverage "blessing of dimensionality" for improved predictive performance and feature accessibility.
• To enhance model generalizability and performance on minority classes.

Main Methods:

• Utilized a dataset of 26,474 electronic health records from Thai hospitals.
• Employed clinically informed feature augmentation and the borderline synthetic minority oversampling technique (SMOTE) for data enrichment and class imbalance.
• Evaluated multiple machine learning models (e.g., XGBoost, Random Forest, TabNet) on original and enriched datasets using standard performance metrics.

Main Results:

• All evaluated machine learning models demonstrated consistent performance improvements on the enriched dataset compared to the original.
• Feature augmentation and SMOTE effectively enhanced model generalizability and minority class performance.
• The study confirmed the benefits of increased dimensionality, guided by domain expertise, for diagnostic accuracy.

Conclusions:

• The proposed framework effectively differentiates dementia from cardiovascular diseases.
• Data enrichment strategies significantly boost the performance of machine learning models for complex clinical diagnoses.
• Leveraging dimensionality with clinical insights offers a promising avenue for improving early dementia detection.