Probabilistic Staging in Alzheimer's Disease with Deep Kernel Learning

1369 

Abstract Submission 

Vasiliki Tassopoulou¹

¹University of Pennsylvania, Philadelphia, PA

Vasiliki Tassopoulou  
University of Pennsylvania
Philadelphia, PA

Machine learning (ML) can significantly enhance the process of subject stratification and selection in clinical trials, offering a more sophisticated and data-driven approach to these critical aspects of clinical research. In this work, we propose to learn a continuous temporal function of progression that describes the transition of subjects from several diagnosis stages. In Alzheimer's Disease (AD), we usually have clinical stages of Cognitive Normal (CN), Mild Cognitive Impairment (MCI) and finally AD. In this study, we use a deep kernel learning framework with temporal single-task Gaussian Processes to learn the temporal function that describes the evolution of the clinical status. Our approach tailors the deep architecture to handle multimodal data including imaging, genomics and clinical information to learn a common embedding as the input to the Gaussian Process (GP) kernel.

Subjects scanned using the same scanner with more than four longitudinal MRI acquisitions from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and Baltimore Longitudinal Study of Aging (BLSA) study cohorts were used for method development. T1 structural MRIs were preprocessed and segmented into 145 gray (GM) and white matter (WM) regions of interest (ROIs) using previously described methods [1]. Models were trained using 145 ROI volumes at baseline along with clinical covariates such as the Age at baseline, the Sex, the APOE4 Allele, cognitive status (CN, MCI or AD), and the number of years of education as features. The output of the model is the temporal function of the cognitive status. The learned temporal function is a function of the baseline acquisition of the subject, which means that on inference time, we need only the first scan of the subject so as to get the temporal function of the clinical status. Through the predictive posterior distribution we can obtain an estimate of the staging curve along with the uncertainty bands. Models were trained using 5-fold cross validation is applied in subject level, which indicates a leave-subjects-out validation scheme.

In the first figure, we visualise the staging curve that is our model's output, along with the staging's uncertainty intervals. The red dashed lines are the critical points that indicate the progression from one stage to another. The 1 corresponds to the CN status, the 0 to the MCI and the -1 to AD. The critical threshold of 0.5 determines the progression from CN to AD, whereas the critical threshold -1 determine the progression from MCI to AD. In the first plot, we observe 6 test samples. The staging curve has been produced using only the baseline acquisition as input. In the plot, we visualize three stable subjects, one CN, one MCI and one AD, along with three converters to AD. In the stable subjects, we observe how the predicted staging curve remains stable. In the CN subject, the staging curve is above 1, on the MCI is between -0.5 and 0.5 and on the stable AD subject the test subjects that progress to AD. This plot showcases the potential of the model to predict staging curves. In the second plot, we calculate the correlation of the clinical biomarker with an imaging one,the SPARE-AD score [2], stratified based on the progression status. The CN stable have the smaller correlation with the SPARE-AD score, since the staging curve is stable and the SPARE-AD are noisy and we observse some variation. In the following progression status, the correlation increases and that indicates a positive relationship with the SPARE-AD score.

The two attached plots, showcase the potential of the Temporal Deep Kernel GP model to be used as a staging model for test subjects with only baseline acquisitions. Further validation, such as the correlation of the Staging Curve with other clinical variables, such as cognitive scores, is needed in order to prove further the validity of the produced staging curves. The work for further validating this model is part of our current and ongoing work.

Bayesian Modeling ¹

Multivariate Approaches ²

Other Methods

Machine Learning