Machine Learning-Driven Biomarker Discovery for Depression and PTSD in Traumatic Brain Injury

Abstract

Introduction Machine Learning holds significant promise in advancing precision psychiatry. Post-psychiatric complications such as PTSD and depression are common after a Traumatic Brain Injury (TBI) (Mayer & Quinn, 2022, Ahmed et al., 2017). Yet, we still lack standard diagnostic criteria for post-TBI psychiatric complications, leaving many individuals undiagnosed and without appropriate healthcare. Through the Enhancing Neuroimaging Genetics through Meta-Analysis (ENIGMA) Brain Injury working group, we addressed this issue by applying state-of-the-art machine learning and imaging analysis techniques to identify specific and localized neural markers of psychiatric illness in TBI. The findings of this research can assist in developing sensitive, personalised biomarkers for early diagnosis of psychiatric disorders in TBI, guiding treatment strategies (Siqueira Pinto et al., 2023). Methods Machine learning using logistic regression with Elastic Net regularization was applied to segmented 3D T1-weighted and diffusion MRI data of the brain to classify 1) individuals with TBI only (n= 547) vs healthy controls with no TBI (HC) (n=150) 2) TBI with psychiatric diagnosis vs HC (TBI/PTSD: n = 196, HC: n = 132; TBI/Depression: n = 194, HC: n = 150; TBI/PTSD & Dep: n = 144, HC: n = 123). Age, sex, and intracranial volume were included as covariates in all models. Participants consisted of n=73 females and n = 624 males (mean age: 47.2 ± 15.6 years). The dataset consisted of LIMBIC-CENC, ADNI-DoD and Duke University datasets. Data was harmonised across consortia using the ComBat algorithm and consisted mostly of deployment-related TBI. White matter features were segmented using the JHU White Matter atlas in a TBSS approach; grey matter cortical and subcortical features were segmented using FreeSurfer. The total number of grey and white matter features included in each model was 239. Results Neuroimaging data classified individuals with TBI and depression vs HCs returning an area under the curve (AUC) of 0.66. The cingulum section adjoining the hippocampus (CGH) was a top discriminant feature - revealing reductions in mean diffusivity (MD), axial diffusivity (AD), radial diffusivity (RD) in right and left CGH and increases in fractional anisotropy (FA) in the cingulum in the cingulated cortex (CGC) predominantly in the left hemisphere. The TBI/Depression/PTSD vs HC model returned an AUC of 0.64 again showing reductions in MD, AD, RD in CGH. The TBI/PTSD vs HC model returned an AUC of 0.58 with reductions in MD, AD and RD in tracts such as ALIC, CST and CGH. The TBI-only vs HC model returned an AUC of 0.59. There were increases in FA across a range of limbic and association tracts, including pathways involved in emotion regulation and cognitive processing, while reductions in MD, AD, and RD were observed in projection and cingulum-related tracts. All models performed significantly better than a null model. Conclusion The results from four machine learning models identify distinct neuroimaging biomarkers associated with traumatic brain injury (TBI) and psychiatric comorbidities. In TBI and depression, disruptions in the microstructural organization of the CGH may contribute to both cognitive and emotional symptoms commonly seen in post-TBI depression. The cingulum is critical for emotional processing, mood regulation, and linking the hippocampus to other emotion-related areas. Its involvement in depression is well established. In this group, increased FA in the CGC may reflect compensatory structural changes in response to CGH damage. As one of the last white matter tracts to mature— reaching peak FA around 42 years old—the cingulum may be particularly vulnerable to environmental impacts (Dennis et al., 2023). Its disruption could serve as a neurobiological marker for post-TBI depression, aiding in early identification of at-risk patients and enabling targeted interventions to mitigate long-term psychological consequences.