Design-induced artifacts when 'disease clocks' are plugged into second-stage analyses of symptom onset

Background and Aims: Plasma phosphorylated tau 217 (p-tau217), including %p-tau217, has emerged as a robust biomarker of Alzheimer's disease (AD) pathology, with increasing interest in its longitudinal behavior. In "Predicting onset of symptomatic Alzheimer's disease with plasma p-tau217 clocks," Petersen et al. applied disease clock models, Sampled Iterative Local Approximation (SILA) and Temporal Integration of Rate Accumulation (TIRA), to estimate age at plasma %p-tau217 positivity and reported that this measure predicts age at onset of symptomatic AD. We aimed to determine whether this apparent predictive performance reflects biomarker information or arises from structural artifacts in the analysis. Methods: We analyzed digitized data from published figures and decomposed the clock-derived predictor into baseline age and estimated time from %p-tau217 positivity. We quantified shared and unique explained variance between baseline age and the clock-derived predictor using commonality analysis. To further disentangle structural and biomarker contributions, we evaluated a null scenario in which the biomarker-derived timing component was replaced with randomly generated values drawn over the observed range, preserving the predictor distribution while removing biomarker information. Results: The reported predictive performance was largely driven by structural artifacts arising from bounded follow up and constraints among the variables. Restriction to individuals who progressed during limited follow up, together with constraints on the allowable timing of events, induced a strong association between baseline age and age at symptom onset. In ADNI, baseline age alone explained substantially more variance in age at onset than the clock-derived predictors (R2=0.78 vs. 0.337 and 0.470 for TIRA and SILA). The estimated time from %p-tau217 positivity contributed minimal additional information, and randomized predictors yielded comparable performance to baseline age alone (R2=0.79). Conclusion: The apparent predictive ability of plasma %p-tau217 disease clocks is driven largely by structural age relationships rather than independent biomarker signal. The plasma %p-tau217 timing component provided minimal predictive value, and its combination with age obscured these structural dependencies. These findings underscore the need for careful evaluation of constructed predictors and outcomes in longitudinal analyses of disease progression.