Evaluating machine learning pipelines for multimodal neuroimaging in small cohorts: an ALS case study

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Machine learning (ML) offers potential for analyzing multimodal neuroimaging data to identify biomarkers and improve neurological disorder diagnosis.
• Rare and heterogeneous diseases present challenges for ML due to small, limited datasets.
• Current research often focuses on optimizing ML classification models for small-cohort data.

Purpose of the Study:

• To systematically evaluate the impact of various ML pipeline configurations on classification performance.
• To assess the efficacy of scaling methods, feature selection, dimensionality reduction, and hyperparameter optimization.
• To determine the best strategies for analyzing multimodal MRI data in small-cohort neurological studies.

Main Methods:

• Evaluated ML pipeline components including scaling, feature selection, dimensionality reduction, and hyperparameter optimization.
• Utilized multimodal MRI data from a cohort of 16 Amyotrophic Lateral Sclerosis (ALS) patients and 14 healthy controls.
• Assessed classification performance based on different pipeline configurations.

Main Results:

• Subject-wise feature normalization showed a modest improvement in classification outcomes.
• Feature selection and dimensionality reduction steps provided limited utility.
• Hyperparameter optimization strategies yielded only marginal gains in performance.
• Overall influence of pipeline refinements on classification performance was modest.

Conclusions:

• For small-cohort studies, focusing on dataset enrichment (expanding cohort size, integrating modalities) is more beneficial than extensive pipeline tuning.
• Methodological framework provided to guide future research in ML for neuroimaging.
• Emphasizes the need for dataset enrichment to enhance clinical utility of ML models in rare neurological diseases.