Machine Learning to Stratify Asymptomatic Alzheimer’s Disease Progression Risk with Cerebrospinal Fluid Biomarkers

• Published in: Alzheimer's & dementia Vol. 18; no. S6
• Main Authors: Tandon, Raghav, Watson, Caroline M, Seyfried, Nicholas T., Mitchell, Cassie S, Zhao, Liping, Lah, James J.
• Format: Journal Article
• Language: English
• Published: 01.12.2022
• ISSN: 1552-5260; 1552-5279
• DOI: 10.1002/alz.069445

Background Change in protein regulation begin years before Alzhiemer’s Disease (AD) functional symptom onset. The period where changes in proteins is detectable without overt cognitive dysfunction is often referred to as the asymptomatic AD stage. The ability to utilize biomarkers to determine who is most likely to transition to symptomatic AD is critical for potential future intervention development. Machine learning provides a means to stratify the asymptomatic AD population using measured biomarker levels. Method Seventy‐five peptides were measured in the cerebrospinal fluid (CSF) of subjects in the Emory Healthy Brain Study. Subjects labeled as symptomatic AD had a clinically confirmed AD diagnosis with cognitive deficit. Quantified total‐tau to amyloid beta ratio was used to label non‐symptomatic patients as either heathy control or asymptomatic AD. Recursive feature elimination (RFE) was used to identify a small set of predictive peptides best for classifying either symptomatic AD or healthy control AD. Asymptomatic AD subjects (n = 134) were clustered based on peptide levels using t stochastic nearest neighbor embedding (t‐SNE). Finally, the selected most predictive peptides and the k‐nearest neighbors (KNN) algorithm was utilized to calculate the proximity of asymptomatic AD subjects to healthy control or symptomatic AD. The calculated proximity represented the stratified AD progression. Result RFE successfully classified patient stage using <10 peptides (AUC > 95%, n =58 held out test subjects). Peptides mapped to the proteins NPTXR, YWHAZ, CHI3L1, VGF, APOE, SMOC1, and GAPDH were most discriminative. T‐SNE results revealed biomarker heterogeneity within the population, but cluster separation was sufficient to utilize KNN proximity as a stand‐in for approximated asymptomatic AD progression. Conclusion Machine learning of CSF biomarker levels successfully stratified asymptomatic subjects into AD‐like (more progressed) and control‐like (relatively less progressed) asymptomatic cases. Stratification of subjects based on biomarker levels could improve future clinical trial construction and assist in earlier intervention for higher risk subjects.