PI-120 - LEVERAGING MACHINE LEARNING FOR HIGH-DIMENSIONAL COVARIATE SCREENING IN ALZHEIMER'S DISEASE PROGRESSION MODELING.

S. Ibrahim^1,2, J. Lang³, I. Asiimwe⁴, S. Mouksassi^5,1, G. Pillai^1,6, S. Shcherbinin⁷, I. Gueorguieva⁸, F. Neuroimaging Initiative⁹; ¹Applied Pharmacometrics Training - Africa Program, c/o Pharmacometrics Africa NPC, Cape Town, South Africa, ²Cairo University, Cairo, Egypt, ³Eli Lilly and Company, 8 Arlington Square West, Downshire Way, Bracknell, Berkshire RG12 1PU, United Kingdom, ⁴Applied Pharmacometrics Training - Africa Program, c/o Pharmacometrics Africa NPC,, Cape Town, South Africa, ⁵Certara, PA, USA, ⁶Division of Clinical Pharmacology, University of Cape Town, Rondebosch 7701, Cape Town, South Africa, ⁷Eli Lilly and Company, 16 893 South Delaware Street, Indiana, United States, ⁸Eli Lilly and Company, , 8 Arlington Square West, Downshire Way, Bracknell, Berkshire RG12 1PU, United Kingdom, ⁹Data used in preparation of this abstract were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu), San Francisco, California, United States.

Background: Alzheimer's disease (AD) is an irreversible neurodegenerative disorder marked by amyloid plaque accumulation, a key early-changing Core 1 biomarker (Jack et al., 2024) which map onto either the amyloid beta or AD tauopathy pathway. Our analysis aims to use machine learning (ML) with disease progression (DP) models to screen predictors of amyloid accumulation in the ADNI database. Early detection of risk factors can significantly reduce AD incidence and is crucial for primary prevention (PP).

Methods: A natural amyloid accumulation DP model (1) identified baseline factors of age, gender, APOE4 genotype and disease state as important factors and the model was used to simulate virtual subjects based on 1291 ADNI participants. Thirty-seven covariates were screened in 100 simulated datasets using nine different ML methods as well as the traditional covariate selection methods (SCM, COSSAC, COSSAC/SAMBA) with and without accounting for highly correlated covariates. The performance of the methods was assessed using the F1 score (accuracy measure). The best method was then used to prescreen covariates in the original dataset followed by SCM.

Results: The ML AIC stepwise (AIC) method was the best method to screen the covariates on the amyloid baseline and slope even when correlated covariates were not accounted for (F1 score = 0.97 ± 0.08, & 0.52 ± 0. 09, respectively). The accuracy of AIC method improved after accounting of correlated covariates (F1 score = 1 ± 0, & 0.76 ± 0.25, respectively). SCM and COSSAC showed lower performance (F1 score = 0.82 ± 0.17/ 0.99 ± 0.025, & 0.68 ± 0.17/ 0.96 ± 0.097, without/with correlations respectively). Importantly, AIC runs had a lower computational burden compared to COSSAC and SCM (1.2 min ± 0.03, 37.6 min ± 26 & 6.84 hr ± 0.1 respectively). AIC as prescreening tool on the original dataset reduced the run time of COSSAC and SCM by 61% & 72.7% respectively. Further evaluation of the selected covariates led to an updated disease progression model with the effect of CSF amyloid-β 42 and phospho-tau on amyloid baseline and APOE4 genotype effect on amyloid progression.

Conclusion: ML methods ease computation and accelerate covariate screening. Simulations guide the choice of ML method and covariate selection. The updated DP model can aid in designing AD PP trials by targeting high-risk groups and quantifying disease progression.