Cross-Trial Knowledge Transfer in Alzheimer's Disease: A Machine Learning Approach Using the EXPEDITION Trials
Babak Khorsand1, Elham Ghanbarian1, Bhargav Nallapu2, Richard Lipton2, Ali Ezzati1
1Department of Neurology, University of California, Irvine, 2Department of Neurology, Albert Einstein College of Medicine
Objective:
To develop methods for enriching Alzheimer's disease (AD) treatment trials with individuals likely to show clinically meaningful cognitive decline (CMCD) if treated with placebo.
Background:
Up to 45% of participants in the placebo arm of AD trials show no CMCD over 80 weeks. Enriching trial populations for individuals likely to decline should increase the power to detect treatment effects.
Design/Methods:
We utilized data from patients with mild AD dementia from the placebo arms of the EXPEDITION1 and EXPEDITION2 trials (N=1025) to develop predictive models. Participants were categorized as Cognitively Stable (CS, defined as an ADAS-Cog14 change of <4 at week 80) or as having CMCD, defined as an ADAS-Cog14 change of ≥4 at trial completion (week 80s) (N=459). Neural Networks, Support Vector Machine, Naïve Bayses, K Nearest Neighbors, and Random Forest were trained to distinguish between CS and CMCD participants using a combination of demographic data, clinical data including Clinical Dementia Rating test (CDR), Geriatric Depression Scale test (GDS), Mini-Mental State Examination test (MMSE), and biomarkers, including APOE4 genotype and volumetric MRI. The trained models were then applied to classify participants in an external validation sample from the EXPEDITION3 study (N=1072).
Results:
Eligible participants from the EXPEDITION3 trial (N=838) had an average age of 72.7 ± 7.8 years; 59% were female. CS status was observed in 44.9% of participants at the final visit while CMCD occurred in 55.1%. Highest performing model showed a sensitivity of 78%, specificity of 52%, and AUC of 61%. The positive predictive values (PPVs) were 9% higher than the base prevalence of CMCD observed at the end of the trial.
Conclusions:
Our findings suggest that predictive models could enhance the design of Alzheimer's Disease trials by allowing for selective exclusion of participants likely to show clinically meaningful cognitive decline over 80 weeks of follow-up.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.