Investigating survival in ALS patients using machine learning and Deformation-Based Morphometry

1415 

Abstract Submission 

Isabelle Lajoie¹, Sanjay Kalra², Mahsa Dadar³

¹McGill University, Montréal, QC, ²University of Alberta, Edmonton, Alberta, ³McGill University, Montreal, QC

Isabelle Lajoie  
McGill University
Montréal, QC

Sanjay Kalra  
University of Alberta
Edmonton, Alberta

Mahsa Dadar  
McGill University
Montreal, QC

Amyotrophic lateral sclerosis (ALS) is a neurodegenerative disorder with no known cure, leading to progressive loss of motor function and ultimately the death of the patient. Relationships between brain-related changes in ALS, symptoms and survival outcomes remain unknown. Deformation-based morphometry (DBM) is an MRI-based technique that can quantify regional atrophy (Ashburner, 1998). Here, we used DBM to localize cross-sectional and longitudinal disease-related brain changes and investigate whether ALS-specific brain changes detected by DBM can improve the performance of machine learning survival models.

Data includes T1-weighted and clinical assessments of 192 ALS patients and 163 healthy controls from the Canadian ALS Neuroimaging Consortium (CALSNIC) study (Kalra, 2020) acquired longitudinally across multiple centers. 104 ALS patients were classified as either short (N=42) or long (N=61) survivors based on the time between disease onset and outcomes (within or more than 24 months). Survival outcome was defined as death or requirement of permanent assisted ventilation for at least 22 hours. All T1-weighted MRIs were preprocessed and quality controlled. DBM maps were obtained by computing the Jacobian determinant of the estimated non-linear deformation fields to the ICBM152 template (Fonov, 2009). DBM values greater than one indicate localized expansion whereas values smaller than one indicate atrophy. To investigate the disease-related brain changes, age and sex were regressed out from the voxel-wise DBM maps of ALS patients based on the estimates of matched controls (Dadar M, 2020). A voxel-wise linear mixed-effects (LME) model was then employed with intercept and follow-up time as variables of interest to evaluate cross-sectional and longitudinal brain changes, respectively. Patients' ID and scanner site were included as categorical random variables. To investigate cross-sectional differences between short and long survival, an additional LME was employed with a categorical fixed variable contrasting survival groups. All results were corrected for multiple comparisons using False Discovery Rate with a threshold of 0.05. To assess how much the DBM features contribute to prognostic accuracy, the features that were significant predictors of survival in Cox univariate analysis were first identified from an initial set of features comprising 1064 regional DBM values (from Schaefer (Glen, 2021), Allen (Hawrylycz, 2012) and JHU (Wakana, 2007) atlases) and 14 clinical features. A logistic regression model was trained and tested on i) regional DBM features, ii) clinical features, and iii) clinical and DBM features combined. Nested cross-validation was employed for hyperparameters optimization and evaluation of performance generalization.

Figure 1.A shows significant cross-sectional bilateral atrophy in the motor cortex, the corticospinal tract (CST), along with an overall pattern of ventricular enlargement in ALS. Figure 1.B shows additional longitudinal changes in the somatomotor region as well as ventricular and sulcal enlargement. Figure 1.C and 1.D show cross-sectional changes in short and long survival groups, respectively, with more atrophy (such as in CST and corpus callosum) and ventricular enlargement observed in short survival. Figure 1.E shows that a portion of the corpus callosum is significantly more atrophied when comparing short to long survivors. Figure 2 shows mean area under the curve (AUC) and mean accuracy (ACC) for the respective set of features: A) AUC=0.86±0.14;ACC=81±10%, B) AUC=0.77±0.17; ACC=67±14% and C) AUC=0.84±0.14;ACC=80±16%. The results demonstrate that the regional DBM features enhance the accuracy of survival prediction and reduce result variability.

This study supports the utility of DBM features in localizing brain volume changes in ALS and improving prognostic accuracy, offering a deeper understanding of the mechanisms underlying disease progression, survival, and clinical disability.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Classification and Predictive Modeling ¹

Multivariate Approaches

Data analysis

Degenerative Disease

Machine Learning

STRUCTURAL MRI

Other - Deformation-based Morphometry, Amyotrophic lateral sclerosis