Neuroimaging-based classification of PTSD using data-driven computational approaches: A multisite big data study from
Xi Zhu1, Yoojean Kim2, Orren Ravid2
1Department of Psychiatry, Columbia University Medical Center, New York, NY, USA; New York State Psychiatric Institute, New York, NY, USA.
Neuroimage
|October 20, 2023
Summary
Machine learning models show limited accuracy in classifying Post-Traumatic Stress Disorder (PTSD) across diverse brain datasets. A denoising variational autoencoder (DVAE) framework offers improved generalizability for multi-site neuroimaging data.
Area of Science:
- Neuroimaging
- Computational Psychiatry
- Machine Learning
Background:
- Data-driven computational approaches are advancing psychiatric disorder diagnosis.
- Current machine learning studies for PTSD are limited by small, homogeneous samples and varied methodologies, hindering result generalization.
- This study utilizes large, heterogeneous brain datasets from the ENIGMA-PGC PTSD Working Group to classify PTSD versus controls and assess generalizability.
Purpose of the Study:
- To classify individuals with Post-Traumatic Stress Disorder (PTSD) versus controls using large-scale, multi-site brain imaging data.
- To evaluate the generalizability and reproducibility of traditional machine learning methods and a denoising variational autoencoder (DVAE) framework.
- To establish a baseline classification performance for PTSD in large, heterogeneous neuroimaging datasets.
Main Methods:
- Analysis of structural MRI (s-MRI), resting-state fMRI (rs-fMRI), and diffusion MRI (d-MRI) data from 3,477, 2,495, and 1,952 individuals, respectively.
- Identification of brain features distinguishing PTSD from controls using traditional machine learning.
- Evaluation of the denoising variational autoencoder (DVAE) for classification performance and generalizability via leave-one-site-out cross-validation.
Main Results:
- Classification performance for PTSD versus controls across multiple sites was modest (60% AUC for s-MRI, 59% for rs-fMRI, 56% for d-MRI).
- Performance improved significantly when classifying PTSD from healthy controls without trauma history (75% AUC).
- The DVAE framework demonstrated better generalization to unseen datasets compared to traditional methods, maintaining classification performance while reducing feature dimensionality.
Conclusions:
- Large-scale neuroimaging datasets provide a baseline for PTSD classification, highlighting the impact of control group selection on performance.
- The DVAE framework shows superior generalizability across multi-site data, suggesting its potential for more robust, clinically applicable neuroimaging-based diagnostic models.
- DVAE classification models are less site-specific, enhancing their applicability in diverse clinical settings.
