Objective:Machine learning studies of PTSD show promise for identifying neurobiological signatures of this disorder, but studies to date have largely excluded Black American women, who experience disproportionately greater trauma and have relatively higher rates of PTSD. PTSD is characterized by four symptom clusters: trauma reexperiencing, trauma avoidance, hyperarousal, and anhedonia. A prior machine learning study reported successful PTSD symptom cluster severity prediction using functional MRI data but did not examine white matter predictors. White matter microstructural integrity has been related to PTSD presence and symptoms, and unexplored metrics such as estimates of tract shape may provide unique predictive utility. Therefore, this study examines the relationship between white matter tract shape and PTSD symptom cluster severity amongst trauma-exposed Black American women using multiple machine learning models.Participants and Methods:Participants included 45 Black American women with PTSD (Mage=40.4(12.9)) and 89 trauma-exposed controls (Mage=39.8(11.6)). Shape and diffusion metrics for the cingulum, corpus callosum, fornix, inferior longitudinal fasciculus, superior longitudinal fasciculus, and uncinate fasciculus were calculated using deterministic tractography. Current symptom severity was calculated using the PTSD Symptom Scales. Input features included tract metrics, questionnaire responses, and age. The following regression models were generated: least absolute shrinkage and selection operator (LASSO), ridge, elastic net, and gaussian process (GPR). Additionally, two forms of latent-scale GPR, one without (lsGPR) and with (sp-lsGPR) node selection via spike and slab priors, were calculated. The performance of regression models was estimated using mean square error (MSE) and R2.Results:sp-lsGPR performed at or above other models across all symptom clusters. LASSO models were comparable to sp-lsGPR for avoidance and hyperarousal clusters. Ridge regression and GPR had the weakest performance across clusters. Scores for sp-lsGPR by cluster are as follows: reexperiencing Mmse=0.70(0.17), Mr2=0.56(0.13); avoidance Mmse=0.75(0.17), Mr2= 0.51(0.13); hyperarousal Mmse=0.57(0.18), Mr2=0.66(0.12); anhedonia Mmse=0.74(0.27), Mr2=0.57(0.13). The top three ranked posterior inclusion probabilities for white matter tracts across sp-lsGPR models include four sections of the cingulum, three sections of the corpus callosum, the right fornix, the left inferior longitudinal fasciculus, the first segment of the right superior longitudinal fasciculus, and the right uncincate fasciculus. The greatest posterior inclusion probability value for the sp-lsGPR models was the left frontal parahippocampal cingulum for the hyperarousal cluster.Conclusions:Results support the combined predictive utility of white matter metrics for brain imaging regression models of PTSD. Results also support the use of sp-lsGPR models, which are designed to balance interpretable linear models and highly-flexible non-linear models. The sp-lsGPR model performance was similar across clusters but was relatively better for the hyperarousal cluster. This finding contrasts with prior machine learning work using functional data which was unable to predict hyperarousal scores above chance (MR2=0.06). These diverging findings highlight the importance of examining both functional and structural data in PTSD populations. Differing findings may also be related to sample characteristics as the prior study was conducted in China. Black American women and Chinese individuals have unique lived experiences that may differentially impact brain structure and function. Future work should continue to include diverse research samples to account for such experiences.