BLSA Plasma Biomarkers and Prediction of Amyloid PET Status

Abstract

AbstractBackgroundIt’s crucial to quantify and compare the accuracies of plasma biomarkers in predicting amyloid PET status.MethodThe sample included a cross‐sectional sample of 212 participants from Baltimore Longitudinal Study of Aging. Amyloid PET status was determined from Pittsburgh compound B (PiB) PET using a Gaussian mixture model (140 PiB– and 72 PiB+). Using the Quanterix SIMOA Neuro‐4‐plex, three plasma biomarkers (amyloid‐beta 42 to 40 ratio [Aβ42/Aβ40], glial fibrillary acidic protein [GFAP], and neurofilament light chain [NfL]) were evaluated. We used receiver operating characteristic (ROC) curve analysis to estimate area under the curve (AUC) for each of the three biomarkers in predicting amyloid PET status. AUC estimates were cross validated using leave‐one‐out method. Finally, from a set of potential features (i.e., the three plasma biomarkers, age and APOE‐ε4 carrier status), we applied stepwise logistic regression to find the best subset of features to predict amyloid PET status.ResultParticipant characteristics are presented in Table 1. Each of the 3 biomarkers performed better than a random classifier, with GFAP having the highest validated AUC (0.692), followed by Aβ42/Aβ40 (AUC=0.677), and NfL (AUC=0.625) (Figure 1). The only statistically significant AUC difference is between GFAP and NfL (p = 0.029) (Table 2). Based on stepwise logistic regression, the model with Aβ42/Aβ40, GFAP and APOE‐ε4 status was selected as the final model with a moderate AUC of 0.732 (Figure 2). Odds ratios associated with each predictor are reported in Table 3.ConclusionOf the three Quanterix SIMOA plasma biomarkers, Aβ42/Aβ40 and GFAP have similar AUCs and are better at predicting amyloid PET status compared with NfL, although all AUCs are modest. The final multivariable logistic regression model, which included Aβ42/Aβ40, GFAP and APOE‐ε4 status, had a modest AUC of 0.732. Future studies with larger samples and more biomarkers, especially p‐tau measure, will be necessary to develop better prediction models.