Mild traumatic brain injury (mTBI) affects more than 2.5 million people per year in the United States. MRI based volumetry has been used to study structural changes associated with mTBI. Cross-sectional and longitudinal studies on brain volumetry in mTBI observed group level changes in regional volumes such as hippocampus and caudate (1-3). However, inferences at single-subject level remain to be determined. The purpose of this work is to derive single-subject level inferences for volumetry features in mTBI at multiple time points after initial trauma using machine learning methods. T1-weighted MRI scan with 1mm isotropic resolution was performed on 78 uncomplicated mTBI subjects and 23 controls using GE 3T Signa MR750 scanner and a 32 channel brain coil under institutional guidelines. The mTBI subjects were scanned three days (32 subjects), seven days (61 subjects), one month (56) and three months (42 subjects) post injury. The scan instances at multiple times after injury are labelled as encounters 1, 2, 3 and 4 respectively. The T1-weighted images were used to segment the brain to derive whole-brain, supratentorial, hippocampus, thalamus, caudate, putamen, pallidus and amygdala volumes (4). Percentage asymmetry between left and right sub-cortical region volumes was computed as 100 × (left volume – right volume)/ [0.5 × (left volume + right volume)]. A suffix of L for left volume, R for right volume and Asym for asymmetry was applied in the labels used to show the volumetry results. The volumes were covariate corrected for age, gender, handedness and years of education using multiple linear regression model. Logistic Regression (LR) models were used at each encounter to identify important volumetric changes at single-subject level. A ten-fold cross-validation frame-work was used to reduce the effects of over-fitting on the LR models.Figure 1 shows the coefficients of the LR model with optimal single-subject level classification performance at different encounters. The average accuracy of the LR models from ten-fold cross-validation at different encounters was between 57 to 73%. The lowest performance was at encounter 1 (3 days post injury), where none of the volumetry feature LR coefficients had significant p-values. Supratentorial (encounters 2, 3, and 4), whole-brain (encounters 2, 3, and 4) and pallidus (3, and 4) volumes had significant p-values (p<0.05) for LR coefficients at later encounters. The coefficients of the derived LR models need to be interpreted together due to the presence of multi co-linearity between different volumetry features. Figure 2 shows the effect of multi co-linearity at encounter 4, which was also observed at encounters 1, 2 and 3 (results not shown). Single-subject level inferences are needed for clinical translation of findings. In this work Logistic regression models were used to identify important volumetry features that jointly describe the mTBI effects at single-subject level. The classification accuracy using LR was higher at encounters 2 and 3 (~70%) compared to 1 and 4 (~60%). The lower performance accuracy at encounters 1 and 4 could be because of either: i) non-significant volumetric differences between controls and mTBI subjects or ii) non-linear relation between volumes and mTBI at durations before 7 days and after 30 days (LR is linear model). One of the limitations of this work is the variability of subject return for subsequent encounters. Only 17 of 78 subjects were present consistently for all encounters. The variation in the results across different encounters could be due to the variation in the subjects. However, the volumes were covariate corrected for age, gender, handedness and years of education of the subjects scanned at different encounters. Importantly, the LR models were verified using ten-fold cross-validation framework. These factors suggest that the changes observed across encounters may be primarily from the changes caused by mTBI longitudinally. The interpretation of effects of mTBI on individual volumetry features using logistic regression coefficients is non-trivial due to multi co-linearity. Future work would use Classification and regression tree (CART) (5,6) models to build non-linear models and also to provide interpretability of changes and interactions between volumetry features in clinical setting. Pallidus, supratentorial and whole brain volumetry features together provide single-subject level signature for mTBI at multiple time-points after injury.