Clinical relevance of three‐range cutpoints in Alzheimer’s disease CSF biomarkers

Abstract

Listen

Translate article

AbstractBackgroundAlzheimer’s disease (AD) was biologically defined by the 2018 NIA‐AA Research Framework (RF), which recommends dichotomously categorizing biomarker status as normal or abnormal using single cutpoints. Nevertheless, the RF also states a three‐range approach might be useful in AD research. However, this potentially relevant strategy remains mostly unexplored in the field. We hereby propose a three‐range cutpoint system to interpret CSF biomarker measurements. We hypothesized the intermediate group would have a different cognitive trajectory than negative/positive groups and that three‐range models would prognostically outperform binary models.MethodWe included 1278 non‐demented individuals (CU: n=575; MCI: n=703) from the ADNI with baseline CSF Elecsys® biomarkers. 20% of these individuals (who had baseline [18F]AV45 Aβ‐PET) were randomly allocated to a “training set” to define three‐range cutpoints – which were later evaluated in a “test set” comprising the remainder 80% participants. Three‐range cutpoints were defined with two‐graph receiver operating characteristics (TGROC) for each CSF biomarker (Aβ1‐42, p‐tau, t‐tau, p‐tau/Aβ1‐42, t‐tau/Aβ1‐42). Cutpoints were generated with 100,000 bootstraps, combining conventional or earlier‐centiloid [18F]AV45 thresholds to cognition in five different schemes for each biomarker. Linear mixed‐effects models (LMM) were employed to choose the best three‐range scheme and also to main analyses. The primary outcome of LMMs was the mPACC composite and participants were followed for up to 6 years. Three‐range models were compared within themselves, to binary and to continuous models based on AIC, BIC, log‐likelihood and R2.ResultLMMs indicated all three‐range approaches outperformed models with binarized or linear‐continuous form of biomarkers and, interestingly, were closer in information metrics to more sophisticate spline‐continuous models, with p‐tau/Aβ1‐42 being the best three‐range predictor (Table 1). For the latter model (Figure 1), slope‐contrast analyses indicated the intermediate group had a faster rate of decline than the negative (β‐estimate: 2.04, t=4.88, p= 3.5x10‐6) but slower than the positive group (β‐estimate: 4.68, t=10.4, p=1x10‐14).ConclusionOur findings demonstrate that the proposed three‐range system for interpreting CSF biomarkers in AD prognostically outperforms binary cutpoints for all evaluated biomarkers, supporting the existence of an intermediate zone beyond the current normal/abnormal paradigm. This approach has clear potential applications to clinical trial recruitment and to clinical practice.