Deep Learning Model Correctly Predicts Time to Alzheimer’s Disease Conversion

Identifying presymptomatic and early mild cognitive impairment (MCI) fast progressors is a crucial component for designing the next generation of Alzheimer’s disease (AD) clinical trials. Innovations in deep learning (DL) in imaging have been hypothesized to improve the power to detect patterns of disease pathophysiology. Researchers proposed a new modeling strategy with a novel multi-input, multi-modal DL model, according to a study presented at the 13th Clinical Trials on Alzheimer’s Disease Congress 2020.

A total of 1,016 patients from the Alzheimer’s Disease Neuroimaging Initiative study with available Abeta (AV-45) positron emission tomography (PET) and T1 magnetic resonance imaging (MRI) scans were included in the study. Researchers built a DL classifier that discriminates between AD and control patients based on the T1 MRI and Abeta PET input. Patients with an AD (n=162) or control (n=351) diagnosis at the time of scanning were selected and further split into training (90%) and holdout (10%) cohorts.

A data augmentation strategy was used for model training. A baseline comparison model was comprised of three commonly measured features derived from T1 MRI (i.e., average hippocampal and ventricular volumes) and from Abeta PET (i.e., standard uptake value ratio normalized by whole cerebellum uptake) to predict AD or control status with a generalized linear model. They also applied the DL and baseline models to 492 patients with MCI who were not previously used for model training.

Patients were stratified by quartiles for their predicted risk for both models and evaluated by the hazard ratio (HR) of the time to conversion to AD in the highest (>75 percentile; Q4) and lowest risk (<25 percentile; Q1) groups. Researchers evaluated the conversion rate at a two-year landmark to understand model performance in a time period commonly used for AD clinical trials.

Researchersobserved similar performance in discriminating AD and control patients in both the DL (accuracy, 0.922) and baseline (accuracy, 0.941) models in the holdout sample. The difference in accuracy was explained by misclassifying four patients in the DL versus three in the baseline model, according to the researchers. When stratifying independent MCI samples by risk, they found that the baseline model had superior performance (Q4 vs. Q1; HR, 22.94) to the DL model (HR, 10.70).

In the DL model, 29.1% of patients in the highest-risk quartile had converted to AD at the two-year landmark; conversion in the lowest-risk group was 5.1%. In the baseline model, 35% of patients in the highest-risk quartile had converted to AD at the two-year landmark; conversion in the lowest-risk group was 0.9%. For comparison, 12.6% of patients are estimated to have converted from MCI to AD at the two-year landmark in the full dataset (i.e., without stratifying on predicted risk). “Our DL model successfully stratified [patients with] MCI by conversion rate and identified a high-risk cohort with a considerably higher AD conversion rate than without stratifying by risk,” the researchers noted.

“Compared with the baseline model, our model had slightly worse performance stratifying [patients with] MCI; this illustrates the challenges of training a DL model from scratch (i.e., without prior knowledge) on limited sample size,” the researchers concluded. “In [the] future, we plan to expand our proof-of-concept to additional datasets to build a more robust DL model incorporating different patient populations, disease stages, and endpoints.”

The study was funded by Roche.

Presentation: P028: A Multi-input, Multi-modal Deep Learning Model to Predict Time to Conversion to Alzheimer’s Disease. Presented at the 13th Clinical Trials on Alzheimer’s Disease Congress 2020, Nov. 4-7, 2020.