Prediction of amyloid and tau status in nondemented older adults using tree-based ensemble models

174 

Abstract Submission 

Hwamee Oh¹, Younghoon Seo²

¹Brown University, Providence, RI, ²Bowdoin College, Brunswick, ME

Hwamee Oh, Ph.D.  
Brown University
Providence, RI

Younghoon Seo  
Bowdoin College
Brunswick, ME

Current gold standards for monitoring brain amyloidosis and tauopathy, the prominent pathological features of Alzheimer's disease (AD), are based on positron emission tomography (PET) scans. Given the expensive and invasive nature of amyloid and tau PET scans, predicting amyloid and tau status in pre-dementia older adults with AD pathologies using more affordable and accessible measures can facilitate early intervention and clinical trials by reducing the screen failure rate. The goal of the present study was to develop interpretable tree-based ensemble models to predict PET-based amyloid and tau burden using non-invasive and widely available variables.

The amyloid (Aβ; n = 1062) and tau (n = 410) PET datasets consisted of individuals with normal cognition and mild cognitive impairment from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Amyloid PET with the [18F]Florbetapir tracer was used as the gold-standard measure for binary amyloid status classification with established positivity cutoffs (< 1.11), while tau PET with the [18F]Flortaucipir tracer was used for the three-stage (low, intermediate, and high) tau status determination using cut-off values found based on pre-established protocols (lower cut-off: < 1.27; upper cut-off: < 1.44). For each subject, we obtained the demographic data, neuropsychological data, apolipoprotein (APOE) ε4 genotype, and volumetric MRI measures, as well as plasma Aβ 42/40 ratio for a subset of the amyloid sample (n = 285). We trained random forest (RF), extreme gradient boosting machine (XGBoost), and light gradient boosting machine (lightGBM) models using different combinations of the features, and measured the model performance using area under the receiver operating curve (AUROC). Shapley Additive exPlanations (SHAP) values were used to rank feature importance.

The performance of baseline non-imaging model showed modest performance for Aβ (RF = 0.665, XGB = 0.650, LGBM = 0.659). Subsequent additions of features improved the predictive performance, with the model using demographic data, cognitive data, and volumetric MRI measures demonstrating the highest performance (RF = 0.762, XGB = 0.763, LGBM = 0.761). Meanwhile, the baseline model achieved modest performance for the three-stage tau classification (RF = 0.643, XGB = 0.654, LGBM = 0.643), and the further addition of features improved the performance, with the feature combination of demographic data, cognitive, volumetric MRI measures, and continuous Aβ PET standardized uptake value ratios (SUVRs) achieving very good performance (RF = 0.799, XGB = 0.801, LGBM = 0.800). SHAP summary plots for Aβ classification revealed the most important features being age, entorhinal cortex volume, and several neuropsychological and functional measures. For tau groups, the low tau group was characterized by low Aβ load, high global cognition scores, and higher hippocampal and middle temporal gyrus volume, intermediate group by higher age, intermediate global cognition, and higher memory scores, and high group by higher Aβ load, more impaired functional scores, lower age, and lower memory scores.

Tree-based ensemble machine learning models achieved modest to very good performance in predicting amyloid and tau status among nondemented older adults. These results suggest that using noninvasive and widely available measures are promising to be used as pre-screening filter for AD clinical trials.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ¹

PET Modeling and Analysis ²

Degenerative Disease

Machine Learning

MRI

Neurological

Positron Emission Tomography (PET)

Other - Alzheimer’s disease, biomarkers, amyloid, tau