Accelerating Hyperparameter Tuning in Machine Learning for Alzheimer’s Disease With High Performance Computing

Abstract

Driven by massive datasets that comprise biomarkers from both blood and magnetic resonance imaging (MRI), the need for advanced learning algorithms and accelerator architectures, such as GPUs and FPGAs has increased. Machine learning (ML) methods have delivered remarkable prediction for the early diagnosis of Alzheimer’s disease (AD). Although ML has improved accuracy of AD prediction, the requirement for the complexity of algorithms in ML increases, for example, hyperparameters tuning, which in turn, increases its computational complexity. Thus, accelerating high performance ML for AD is an important research challenge facing these fields. This work reports a multicore high performance support vector machine (SVM) hyperparameter tuning workflow with 100 times repeated 5-fold cross-validation for speeding up ML for AD. For demonstration and evaluation purposes, the high performance hyperparameter tuning model was applied to public MRI data for AD and included demographic factors such as age, sex and education. Results showed that computational efficiency increased by 96%, which helped to shed light on future diagnostic AD biomarker applications. The high performance hyperparameter tuning model can also be applied to other ML algorithms such as random forest, logistic regression, xgboost, etc.

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

“…The HABS-HD dataset contains 1328 normal controls and 377 MCIs and ADs (prevalence rate = 22.1% and class imbalance index = 0.31). A previous workflow [1] for hyperparameter tuning with HPC could have encountered an out-of-memory problem for the big imbalanced HABS-HD data. This study aimed to (1) solve the out-of-memory problem, (2) improve the model performance for imbalanced data, and (3) increase computation efficiency.…”

“…We adopted 10 times repeated fivefold cross-validation [1] to reduce the noisy estimation of the optimal parameters of the ML model caused by a single run of the fivefold cross-validation [20]. Briefly, the fivefold cross-validation procedure where data samples are shuffled and stratified is repeated 10 times, and the mean performance across all folds from all runs is then used for the hyperparameter tuning.…”

“…However, applying ML and HPC to Alzheimer's disease (AD) research is still relatively new. There was only 1 hit [1] from a PubMed search of ML, HPC, and AD (and 84 hits from Google), although the number of publications pertaining to ML and AD has greatly increased, from 294 in 2020 to 1582 in 2022.…”

“…HPC provides great opportunities to look into the imbalanced data problem while accelerating hyperparameter tuning efficiently [18,19]. In our previous work, we found that HPC could be used to significantly reduce computational time while maintaining the necessary accuracy for balanced AD data [1]. However, applying the workflow from [1] to imbalanced big data may cause an out-of-memory problem because the addition of a third parameter causes the memory usage to increase by x times, where x is the length of the class weight (normally, x is between 10 and 100).…”

“…In our previous work, we found that HPC could be used to significantly reduce computational time while maintaining the necessary accuracy for balanced AD data [1]. However, applying the workflow from [1] to imbalanced big data may cause an out-of-memory problem because the addition of a third parameter causes the memory usage to increase by x times, where x is the length of the class weight (normally, x is between 10 and 100). Loading the big imbalanced data chunkby-chunk and then tuning it partially will not result in the same best parameters for all the chunks.…”