Performance of machine learning algorithms in diffusion tensor imaging of movement disorders: an exploratory...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Machine learning (ML) applied to diffusion tensor imaging (DTI) shows potential for detecting brain alterations in movement disorders.
• Existing studies exhibit significant methodological heterogeneity, hindering quantitative comparison and clinical translation.

Purpose of the Study:

• To characterize performance trends, methodological diversity, and variability in ML models trained on DTI data for movement disorder classification.
• This exploratory meta-analysis aimed to describe performance distributions and methodological patterns, not to infer a single pooled diagnostic effect due to extreme heterogeneity.

Main Methods:

• Systematic literature search of PubMed, Web of Science, and Scopus for studies using ML with DTI in movement disorders.
• Extracted accuracy, sensitivity, specificity, and AUC; employed random-effects modeling for descriptive summaries and subgroup analyses.
• Assessed study quality using JBI tools; used multiple imputation for missing metrics.

Main Results:

• Included 46 studies (2016-2024) on Parkinson's disease, Tourette syndrome, and essential tremor.
• Reported median AUC was high (≈0.91), but heterogeneity was extreme (I²=94.7%).
• Deep learning and radiomics models showed high accuracy but limited external validity due to small, single-center cohorts.

Conclusions:

• ML models with DTI exhibit high internal performance for movement disorders, but generalizability is limited.
• Current evidence is exploratory, hampered by small sample sizes, methodological fragmentation, and lack of standardized pipelines.
• Future progress requires data harmonization, multicenter collaborations, and federated learning for reproducible and interpretable models.