Hyperparameter Tuning with High Performance Computing Machine Learning for Imbalanced Alzheimer's Disease Data.

Area of Science:

• Computational neuroscience
• Biomedical data science
• Machine learning applications in healthcare

Background:

• Accurate machine learning (ML) detection of Alzheimer's disease (AD) remains challenging.
• Class imbalance in AD datasets poses a significant hurdle for ML algorithms.
• Existing ML models often assume evenly distributed data, which is not typical for AD research.

Purpose of the Study:

• To develop and evaluate a high-performance computing (HPC) based hyperparameter tuning workflow.
• To address the challenge of imbalanced data in mild cognitive impairment (MCI) and AD detection.
• To optimize Support Vector Machine (SVM) models for AD-related datasets.

Main Methods:

• Implemented a single-node multicore parallel mode for hyperparameter tuning (gamma, cost, class weight) of SVM models.
• Utilized R packages (bigmemory, foreach, doParallel) on the Texas Advanced Computing Center (TACC) Lonestar6 system.
• Employed 10x repeated fivefold cross-validation for robust model evaluation.

Main Results:

• Achieved a dramatic reduction in computational time by up to 98.2% for SVM hyperparameter tuning.
• Improved cross-validation performance, with positive predictive value (PPV) at 16.42% and negative predictive value (NPV) at 92.72% (base rate 12%).
• Demonstrated enhanced agility, simplicity, and productivity in processing imbalanced AD data.

Conclusions:

• A single-node multicore parallel structure combined with high-performance SVM hyperparameter tuning offers efficient and fast computation.
• The proposed workflow effectively handles imbalanced data challenges in AD applications.
• This approach shows significant potential for improving ML-based diagnostic tools for Alzheimer's disease.