Data Driven Staging of Genetic Frontotemporal Dementia by Neuroimaging Data

229 

Abstract Submission 

Mahdie Soltaninejad^1,2, Yasser Iturria Medina^1,2, Reza Rajabli^1,2, Gleb Bezgin^1,2, Simon Ducharme^1,2,3

¹McGill University, Montreal, Canada, ²Montreal Neurological Institute, Montreal, Canada, ³Douglas Mental Health University Institute, Montreal, Canada

Mahdie Soltaninejad  
McGill University|Montreal Neurological Institute
Montreal, Canada|Montreal, Canada

Yasser Iturria Medina, YIM  
McGill University|Montreal Neurological Institute
Montreal, Canada|Montreal, Canada

Reza Rajabli  
McGill University|Montreal Neurological Institute
Montreal, Canada|Montreal, Canada

Gleb Bezgin  
McGill University|Montreal Neurological Institute
Montreal, Canada|Montreal, Canada

Simon Ducharme  
McGill University|Montreal Neurological Institute|Douglas Mental Health University Institute
Montreal, Canada|Montreal, Canada|Montreal, Canada

Frontotemporal dementia (FTD) is a complex disorder marked by substantial clinical, genetic, and pathological variations. Clinical presentations include behavioral changes and/or language impairment [1]. Heterogeneity poses a significant challenge for treatment development, emphasizing the need for precise biomarkers to track disease progression. Despite advancements, current biomarkers demonstrate notable variability among FTD variants, constraining their individual utility in disease staging. Our study bridges this gap by incorporating a multitude of biomarkers in progression modeling. We employ the contrastive trajectory inference (cTI) algorithm [2] to analyze multi-modal features in FTD. This approach offers a comprehensive exploration of disease staging, leveraging neuroimaging data to uncover complex patterns. Unlike previous FTD investigations that frequently oversimplify progression with a single disease trajectory assumption, our model acknowledges the potential existence of multiple disease trajectories.

Our study utilized a dataset obtained from the Genetic Frontotemporal dementia Initiative (GENFI), comprising 922 MRI scans from individuals with genetic FTD and 630 scans from healthy controls. The T1w and T2w scans of all participants were processed using the MINC toolkit in order to measure white matter hyperintensities [3]. Additionally, cortical thickness measurements were obtained through Freesurfer v7.1.1 [4]. Post image processing, data harmonization was achieved through the application of the COMBAT algorithm [5], and data standardization was performed by calculating z-scores. Subsequently, we employed the cTI method which is an unsupervised machine learning algorithm for staging and subtyping high dimensional data. This process commenced with feature selection and dimension reduction facilitated by contrastive principal component analysis. Trajectory assignment was achieved through the utilization of a minimum spanning tree, wherein subjects were categorized into distinct disease trajectories, and a corresponding disease score was calculated for each individual. Validation of our staging methodology was conducted through the comparison of disease scores with key clinical metrics. Furthermore, we employed power analysis to determine the necessary sample size for clinical trials utilizing our disease score, and we conducted a comparative analysis with commonly utilized clinical scores.

Robust correlations (p<0.001) were observed between disease scores and critical clinical measures, including scores of MMSE (r=-0.45), Digit Symbol Substitution Test (r= -0.40), Boston Naming Test (-0.41), Verbal Fluency Task (r=-0.30), MiniSEA test (r=-0.29), and Trail Making Test (r=0.43). Notably, disease score exhibited a significant correlation with the estimated year of onset (p<0.001). Subtyping analysis identified three distinct categories: healthy controls, and two subtypes among FTD mutation carriers. Our power analysis revealed a significant reduction in required sample size when utilizing our cTI disease score as opposed to relying on clinical and neuropsychological scores.

Employing a data-driven method on neuroimaging data, we were able to derive individualized disease scores within a heterogenous group of individuals with FTD. Calculated disease scores exhibited correlations with a comprehensive array of clinical and neuropsychological assessments, including evaluations of behavioral symptoms, attention, memory, language, and executive functions. The encouraging indications from our findings suggest that data-driven approaches on neuroimaging features hold promise as an effective method for personalized assessment in patients and disease monitoring in clinical trials. Specifically, our cTI score demonstrated its value in guiding the planning of clinical trials for FTD. Future work will delve into the factors influencing subtypes, providing valuable insights for personalized interventions in the realm of FTD.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ¹

Aging

Multivariate Approaches ²

Aging

Degenerative Disease

Machine Learning

MRI

Statistical Methods

Other - FTD; Disease Progression; contrastive Trajectory Inference