Class-Balanced Deep Learning with Adaptive Vector Scaling Loss for Dementia Stage Detection.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Alzheimer's disease (AD) causes irreversible cognitive decline, with Mild Cognitive Impairment (MCI) as a precursor.
• Early detection of AD and related dementias is critical for intervention and slowing disease progression.
• Class imbalance in machine learning models for cognitive states (CN, MCI, AD) necessitates balanced accuracy metrics.

Purpose of the Study:

• To introduce VS-Opt-Net, a novel method combining vector-scaling (VS) loss and Bayesian optimization within the STREAMLINE pipeline.
• To enhance the performance and balanced accuracy of machine learning models for classifying cognitive normal (CN), Mild Cognitive Impairment (MCI), and Alzheimer's disease (AD) subjects.
• To address class imbalance and improve generalization in deep network training for dementia detection.

Main Methods:

• Utilized MRI-based brain regional measurements as features for binary classifications (CN vs MCI, AD vs MCI).
• Incorporated the vector-scaling (VS) loss function into the STREAMLINE machine learning pipeline.
• Employed Bayesian optimization for hyperparameter tuning of both the VS loss function and the deep learning model.
• Compared the balanced accuracy of VS-Opt-Net against other class-balanced machine learning models and loss functions.

Main Results:

• Hyperparameter optimization using Bayesian methods significantly improved the balanced accuracy of the deep neural network with VS loss.
• The VS-Opt-Net model demonstrated superior performance compared to other models on the Alzheimer's disease dataset.
• Feature importance analysis revealed VS-Opt-Net's capability to identify key biomarker differences across dementia stages.

Conclusions:

• VS-Opt-Net effectively enhances model performance and balanced accuracy in classifying cognitive states, particularly for imbalanced datasets.
• The proposed method, leveraging VS loss and Bayesian optimization, offers a promising approach for early Alzheimer's disease detection using neuroimaging data.
• VS-Opt-Net aids in understanding neurobiological distinctions between different stages of cognitive impairment and dementia.