A. Parkinson’s disease
a. Freezing of Gait (FoG)
Freezing-of-gait (FOG) is defined as a brief, episodic absence or marked reduction of the forward progression of the feet despite the intention to walk. PD-FOG is considered one of the most debilitating motor symptoms affecting almost 50% of patients with PD and is one of the most common causes of falls and subsequent morbidity and mortality. Several competing hypotheses have been proposed to understand PD-FOG. However, these models are incomplete and only partially explain the phenotype of PD-FOG. In general, FOG has been attributed to overloading across neural networks in an attempt to compensate for reduced motor function, which may lead to the inability to “set-shift” among the different neural networks. We utilized rsfMRI data and graph theoretical approaches and found that the PD-FOG patients showed a noticeable reorganization of hub regions. However, we could not prove or disprove any of the commonly accepted hypotheses for PD-FOG. We are currently developing novel computational models using deep-learning approaches to analyze the MRI data for PD-FOG to identify predictive biomarkers of PD-FOG.
Freezing of gait (FOG) is a major clinical challenge in Parkinson’s disease (PD), and our study investigated the potential of predicting PD-FOG using an unbiased machine learning (ML) framework combining structural MRI and clinical data. Thirty-seven participants (16 PD-FOG, 21 PD-nFOG) underwent isotropic 1 mm³ T1-weighted MRI, from which brain morphometric features—including subcortical volume, cortical volume, mean curvature, area, local gyrification index, and thickness—were extracted using FreeSurfer7. Participants were split into discovery (13 PD-FOG, 17 PD-nFOG) and independent testing (3 PD-FOG, 4 PD-nFOG) cohorts. Using Random Forest (RF), Support Vector Machine (SVM)-Linear, and SVM-Non-Linear models with least absolute shrinkage and selection operator (LASSO) for feature reduction, we evaluated predictive performance across individual and combined feature sets. Among all models, the SVM-linear approach achieved the best performance on the independent dataset (AUC = 0.71, precision = 75%, sensitivity = 75%, specificity = 66.67%). Notably, cortical area measures from 24 FreeSurfer-derived regions—distributed across frontal, temporal, parietal, and occipital lobes—emerged as the key predictors of PD-FOG, while other measures, either alone or combined, showed no predictive value. Importantly, these cortical area features significantly correlated with clinical and physical therapy metrics of FOG, supporting their biological relevance. Together, these findings demonstrate that widespread cortical involvement, captured through regional cortical area measures, can serve as robust predictors of PD-FOG when modeled with SVM-linear ML approaches.
b. Mild Cognitive Impairment (MCI)
Approximately 50-80% of Parkinson’s disease (PD) patients develop PD-dementia (PDD) within twelve years of diagnosis. However, no reliable method yet exists to predict PDD. Identifying pathophysiology-based biomarkers that could identify PD patients at high risk for PDD reliably is critical for better understanding the pathophysiological processes underlying PDD. We have invested our efforts in approaching the identification of predictive biomarkers of PDD through multimodal imaging utilizing both MRI-derived measures (diffusion-weighted MRI (dMRI), resting-state functional MRI (rsfMRI), and T1-MRI measures) and non-imaging measures such as demographics (Eg, age, sex), clinical (Eg, disease duration and severity), genetics (Eg, LRRK2), and CSF measures (Eg, Total Tau, β-Amyloid, α-synuclein) since each of these measures has been shown to independently predict PDD. Our current analyses show that advanced dMRI measures may be more predictive of mild cognitive impairment (MCI) in PD than the conventional single tensor dMRI measures. Our initial results have also shown that dopamine affects the functional connectivity in PD-MCI differently than PD with normal cognition. A deeper understanding of these effects, along with developing novel tools for analyzing rsfMRI and developing a deep-learning model by fusing these modalities to identify reproducible biomarkers of PDD, is currently underway.
B. Harmonization
Diusion MRI (dMRI) is the most widely used MRI technique to understand the relationship between white matter (WM) structures and aging. However, dMRI studies are poorly replicated, and the variability depends just not on the scanners but also the protocols. Furthermore, most dMRI studies are limited by the small sample size thereby limiting the generalizability of the conclusions. Hence, several large-scale multicenter studies with harmonized protocols are designed to increase the statistical power. However, the scanner variability due to local and temporal scanner characteristics can still exist which may result in high inter- and intra-scanner variability. Several techniques exist to reduce or remove the scanner/site variability such as meta-analysis where the site/scanner is coded as a dummy variable in the general linear statistical model (GLM), and mega-analysis where all sites/scanners jointly contribute to estimate the population dierence. Two mega-analysis models, namely (a) ComBat which is a batch harmonization technique that adds scanner eect as a multiplicative eect and is an extension of the GLM; and (b) rotation invariant spherical harmonics (RISH) features that remove scanner/site-specic biases while accounting for nonlinearity in dMRI signals , are widely used to harmonize diusion tensor imaging (DTI) metrics such as fractional anisotropy (FA), axial diusivity (AxD), radial diusivity (RD), and mean diusivity (MD). To our knowledge, no study has systematically compared these meta- and mega-analyses techniques for dMRI harmonization that can inform the sensitivity and specicity of various harmonization techniques. Hence, in this study, we utilized dMRI data from four dierent sites that were collected on dierent healthy participants over 85 years old across four dierent 3T MRI scanners from SIEMENS ® (PRISMA at the University of Alabama at Birmingham (UAB) and the University of Florida (UF), SKYRA at the University of Arizona (UA), and VIDA at the University of Miami (UM)) along with an NIH toolbox cognitive battery (NIH-TB-CB) to understand the correlations between dMRI-derived WM measures and NIH-TB-CB-derived crystallized (CM) and uid memory (FM).
C. Glymphatic Index
Sports-related traumatic brain injuries (TBIs) account for 1.2–30.3% of all TBI cases in the United States, with boxing and martial arts being major contributors. Repeated head impacts (RHI) are recognized as a risk factor for neurodegenerative and neuropsychiatric disorders. The recently discovered glymphatic system, essential for clearing brain waste such as amyloid-β and tau proteins, has been shown to be disrupted in individuals with cognitive decline. This system can be evaluated using Diffusion Tensor Imaging (DTI) through the Along the Perivascular Space (ALPS) method, which measures the ALPS index. We hypothesized that RHI would impair glymphatic function, leading to lower ALPS indices in cognitively impaired fighters compared to non-impaired fighters, and that total ALPS would correlate with the number of knockouts among impaired fighters. This longitudinal study included 280 male professional fighters (boxers and mixed martial artists) and 20 age-matched healthy controls, with recorded variables including age, race, and intracranial volume. Ninety-five fighters were classified as cognitively impaired based on neuropsychological assessments of processing and psychomotor speed. MRI data were acquired using a 3T Siemens Verio scanner with a 32-channel head coil, with diffusion-weighted imaging (b = 0 and b = 1000 s/mm²) across 71 directions. Diffusion metric images, including fractional anisotropy (FA) and directional diffusivity maps along the x-, y-, and z-axes, were generated using FSL software. The DTI-ALPS index was computed by comparing diffusivity values along the perivascular space with those of projection and association fibers at the level of the lateral ventricle body. Group comparisons and correlations between DTI-ALPS indices and the total number of knockouts were analyzed using PALM in FSL, with results significant at a familywise error (FWE)–corrected p < 0.05. Results showed that non-impaired fighters had significantly lower right and total glymphatic indices compared to impaired fighters. The relationship between glymphatic index and knockout history differed significantly between groups, with impaired fighters showing a non-significant negative correlation between total knockouts and glymphatic index. Contrary to our hypothesis, impaired fighters exhibited higher glymphatic indices than non-impaired fighters; however, glymphatic function deteriorated significantly with increasing knockouts. Understanding the impact of RHI on the glymphatic system is critical for early detection and management of neurodegenerative risk in contact sport athletes, and this study demonstrates the potential of the DTI-derived ALPS index as a noninvasive biomarker for assessing glymphatic dysfunction and guiding early intervention.
D. Repetitive Head Impact
Participants exposed to repeated head impacts (RHI) such as those involved in American football, professional boxing, etc. are at a higher risk of developing a brain condition called chronic traumatic encephalopathy (CTE) that leads to dementia. However, currently, the only definitive way of diagnosing CTE is through neuropathological testing. Our lab has focused its efforts on utilizing the combination of multimodal MRI and machine-learning techniques to identify participants at-risk for developing CTE. We use the clinical presentation and a battery of neuropsychological testing to evaluate whether the participant presents symptoms of cognitive decline when compared to a participant of the same sex, education, age, and other demographical factors. We developed our technique with a cohort of active professional fighters and found that seven MRI-derived features predicted fighters having clinical cognitive impairment with ~75% accuracy, both at baseline and follow-up using only the weights identified at baseline. We are currently in the process of increasing our predictive ability by adding other MRI measures such as resting-state fMRI connectivity changes and blood changes using deep-learning methodologies, as these changes have independently predicted cognitive decline due to RHI.
