Multi-center machine learning in imaging psychiatry: A meta-model approach.

Area of Science:

• Neuroimaging
• Machine Learning
• Psychiatric Diagnostics

Background:

• Automated diagnosis of psychiatric disorders using medical images is challenged by limited training data.
• Heterogeneous disorders like schizophrenia require large, diverse datasets for robust machine learning model generalizability.
• Existing multicenter studies face hurdles with data sharing due to legal/ethical concerns and computational complexity.

Purpose of the Study:

• To investigate the feasibility of creating a diagnostic meta-model by combining local support vector machine (SVM) classifiers without sharing raw medical data.
• To evaluate the performance of this distributed learning approach against models trained on pooled multicenter data.

Main Methods:

• Utilized a 4-center setup with 480 participants (schizophrenia patients and healthy controls).
• Trained SVM models using three types of structural MRI imaging features to differentiate patients from controls.
• Developed meta-models by combining locally trained SVM classifiers and compared them to a single model trained on all pooled data.

Main Results:

• The combined meta-model demonstrated high similarity to the model built on pooled data.
• Classification performance of the meta-model was comparable to the pooled data model across all imaging features.
• Both meta-model similarity and performance surpassed that of individual local models.

Conclusions:

• Distributed learning via meta-model combination is a practical alternative for building robust diagnostic models.
• This approach overcomes data sharing limitations, enabling larger and more informative analyses in psychiatric neuroimaging.
• Facilitates privacy-preserving, large-scale machine learning for psychiatric disorder diagnosis.