Reconstructing subject-specific effect maps.

Area of Science:

• Neuroimaging analysis
• Machine learning in medicine
• Biostatistics

Background:

• Predictive models enable subject-specific inference from neuroimaging data for disease analysis.
• Local inference for detecting condition effects on individual measurements is underutilized due to noisy, dispersed detections.

Purpose of the Study:

• To propose a novel reconstruction method (RSM) to enhance subject-specific detections from predictive modeling, particularly binary classifiers.
• To reduce noise in local inference caused by sampling errors in classifier training.

Main Methods:

• RSM is a wrapper-type algorithm applicable to various binary classifiers for diagnostic purposes.
• Reconstruction is framed as a Maximum-A-Posteriori problem utilizing a classifier-specific prior model.

Main Results:

• RSM demonstrated higher detection accuracy on synthetic data compared to direct modeling or bootstrap averaging.
• On the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, RSM improved correlations between subject-specific detections and non-imaging Alzheimer's Disease markers.

Conclusions:

• RSM effectively reduces noise in subject-specific detections from predictive models.
• The method enhances detection accuracy and reliability, showing promise for clinical applications in neurodegenerative diseases.