Studies have shown that both active and retired athletes with repeated head trauma are more likely to suffer from cognitive decline and loss of executive and attention functions when compared to age-matched healthy controls^{1, 2}. Several cross-sectional MRI studies have shown structural differences in such athletes but there has been no results reported in longitudinal studies of such active fighters^{3, 4}. The professional fighters brain health study (PFBHS) is a longitudinal study of active professional fighters with age-matched healthy controls using multimodal MRI methods⁵. In this study, we show that the features predicting cognitive decline at baseline are sensitive and specific over time in our cohort of longitudinal active fighters.Baseline Subjects: 331 subjects (21-controls (20male (M); 30.1±8.28years), 119-boxers (108M; 29.71±7.06years), 152-mixed-martial-arts-fighters (143M; 29.05±5.04years), 22-martial-arts-fighters (19M; 28.36±5.04years) and 17-mixed-fighters (15M; 30±5.66years)) were recruited at the center. Years of education, number of professional and amateur fights, race, gender and year of first fighting were recorded at every visit. Each subject went through various neuropsychological assessment tests to measure psychomotor speed (PSY) and processing (P) speed. These scores were standardized⁶ (PSYSTANS/PSTANS) and the fighters were categorized as “impaired” (99-subjects) or “non-impaired” (211-subjects). Timepoint1 subjects: 79 subjects (6-controls, 19-impaired-fighters and 54-nonimpaired-fighters) were scanned again after at-least one year. Data Acquisition: All subjects, at both timepoints, were scanned with a 3T Siemens Verio scanner. Single-shot-EPI-sequence was used to acquire diffusion weighted images (DWIs) with 71 diffusion directions and b-value of 1000 s/mm²; TR/TE/Resolution=7000ms/91ms/2.5mm³. Sagittal-MPRAGE T1-weighted images were also acquired for every subject within the same session with TR/TE/FA/Resolution= 2300ms/2.98ms/9^o/1mm³. Data Processing: Standard processing steps were used to fit diffusion tensors after eddy current distortion correction in FSL⁷. After registering each subject to MNI152 template using tbss_2_reg and tbss_3_postreg procedures from TBSS⁸, various diffusion metrics such as fractional anisotropy (FA), axial diffusivity (AxD), transverse diffusivity (RD) and mean diffusivity (MD) were estimated for all the 20 major white matter (WM) tracts from JHU WM atlas⁹. Using AAL template in MNII152 space, MD was also estimated for all the 116 gray matter AAL structures. Various other structural measures such as volume and thickness were extracted for every subject for every label in Collin’s atlas using Freesurfer v5.3.0¹⁰. Feature selection: A total of 314 features (113 structural measurements, 196 diffusion measurements, gender, age, years of education, years of professional fighting and type of fighters) were used for supervised feature classification using an in-house feature selection algorithm combining least absolute shrinkage and selection operator (LASSO)¹¹ and radial basis functional network (RBFN)¹². Relationship between the features and the cognitive score: The mean of the extracted features at the baseline were plotted against PSTANS and PSYSTANS to explore if the features predicted cognitive decline in the fighters. The change in the extracted features over time were also plotted against change (Δ) in PSTANS and ΔPSYSTANS to explore the sensitivity and specificity of the predicted features on cognitive decline. Nonparametric t-tests was used to perform statistical comparisons, controlled for age and gender.Machine learning algorithm detected 6/314 features (FA of FMajor and inferior longitudinal fasciculus (ILF), left cerebellum white matter volume, left thalamic volume, thickness of left lateralorbitofrontal and left medialorbitofrontal) that predicted cognitive decline in our cohorts (prediction accuracy=0.7; Area under the curve=0.724). Mean values of all the extracted features normalized with respect to controls is shown in the left panel of Fig. 1. Boxplot of all the features (baseline and timepoint1) for the 3 groups (right panel of Fig. 1 and Fig 3) shows a significant decline (p<0.05) in the values for impaired fighters as compared to nonimpaired fighters and controls. PSTANS, ΔPSTANS, PSYSTANS and ΔPSYSTANS showed a significant relationship (p<0.05) with the predicted features (baseline and timepoint1) (Fig 2a, Fig 2b, Fig 4a and Fig 4b). Of note, significant difference in the slope (p<0.05) was observed for FA of FMajor between impaired and nonimpaired fighters (baseline and timepoint1) and left thalamic volume between control and impaired fighters (timepoint1).Using a novel machine learning approach combining LASSO and RBFN on a multimodal MRI dataset, our results show 6 features that predict cognitive decline in our cohort of professional fighters. Our results also report for the first time that the features predicting cognitive decline at baseline using machine learning algorithm are sensitive and specific over time in our longitudinal cohort of active professionals. The predictive features showed a significant relationship between the neuropsychological assessment tests and hence can be used as a surrogate to predict/monitor cognitive decline in patients with traumatic brain injuries.