Biomarkers.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Machine learning (ML) models are crucial for Alzheimer's disease (AD) staging but often yield conflicting results due to unclear methodologies.
• This limits their clinical utility and patient stratification in clinical trials.
• A novel automated ML (autoML) approach is proposed to address these limitations by mitigating bias and enhancing predictive accuracy.

Purpose of the Study:

• To develop and evaluate a novel automated ML (autoML) framework for accurate Alzheimer's disease (AD) stage classification.
• To mitigate experimenter bias and arbitrary decisions in predictive modeling.
• To enhance the clinical utility of AD staging for patient stratification in clinical trials.

Main Methods:

• An AutoML-multiverse framework was developed, navigating 20,000 ML pipelines using Bayesian Optimization.
• Data-driven ensembles were constructed by integrating diverse pipelines into stacked models.
• Classification tasks included cognitively normal (CN) vs. AD, AD vs. mild cognitive impairment (MCI) vs. CN, and stable MCI (sMCI) vs. progressive MCI (pMCI), using structural MRI and clinical data from the ADNI dataset.

Main Results:

• Standalone ML models showed limited predictive performance; stacked ensembles offered marginal improvements.
• AutoML-derived ensembles consistently achieved superior accuracy across all diagnostic tasks.
• The AutoML ensemble achieved 77.55% balanced accuracy in distinguishing sMCI from pMCI, demonstrating the clinical relevance of MRI data.

Conclusions:

• The AutoML-multiverse framework demonstrates superior predictive performance compared to individual ML pipelines.
• It enables data-driven ensemble construction, reducing bias and arbitrary decision-making.
• The framework shows potential clinical utility in trials by integrating neuroimaging for precise patient stratification, especially for distinguishing disease progression rates.