Harmonization for Parkinson’s Disease MRI Classification
Author Information
Author(s): Saqib Mohammed, Silvina G. Horovitz, Ikuko Miyazaki
Primary Institution: University of Pennsylvania
Hypothesis
The scanner-specific classifier will perform well for its own scanner but will perform poorly for other scanners.
Conclusion
Batch effects remain a major issue for classification problems, and solving for batch effects in a classifier must avoid circularity and overly optimistic results.
Supporting Evidence
- Multi-scanner classifiers that considered neurobiological batch effects could achieve a test AUC of 0.902.
- Single scanner classifiers achieved a mean AUC of 0.651 ± 0.144.
- Classifiers trained with the NeuroComBat model where the group was used were more performant compared to classifiers trained without.
Takeaway
This study looks at how to better classify Parkinson's disease using MRI scans from different machines, finding that using data from multiple scanners can lead to problems if not handled correctly.
Methodology
The study used a multi-dataset cohort of T1 MRI scans from 372 subjects, applying the ComBat harmonization model to classify subjects with Parkinson’s disease and healthy volunteers.
Potential Biases
Differences in clinical populations between sites can introduce bias.
Limitations
The study is limited by the need for scanners with enough data and the potential for data leakage in the harmonization process.
Participant Demographics
The cohort included 216 subjects with Parkinson’s disease and 156 healthy volunteers from 11 different scanners.
Statistical Information
Statistical Significance
p<0.05
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website