•Dataset: Baseline data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) were analysed (n = 550), comprising 289 cognitively normal (CN) participants, 192 individuals with mild cognitive impairment (MCI), and 69 individuals with Alzheimer's disease (AD).

•Variables: analysis incorporated key plasma biomarkers (pTau217, Aβ42/40, GFAP, and NfL), neuroimaging measures, demographic characteristics, and cognitive performance assessed using the ADAS-Cog13.

•Causal discovery approach: The Fast Causal Inference (FCI)1 algorithm was applied using non-linear kernel conditional independence tests (α = 0.05) to infer potential causal relationships from observational data while allowing for latent confounding.

•Incorporation of prior knowledge: Biological and temporal constraints were imposed by specifying demographic variables as root nodes (having no incoming edges) and ADAS-Cog13 as a sink node (having no outgoing edges).

 

Overall structure: The inferred causal network was broadly consistent with established AD pathophysiology, supporting biological plausibility.

pTau217: Showed a direct effect on cognitive outcomes not mediated by other measured variables and was consistently upstream of multiple cognitive measures and GFAP (64% bootstrap frequency).

NfL: NFL unexpectedly appeared upstream of GFAP in 66% of bootstrap runs, contrasting with its typical characterisation as a late-stage neurodegeneration marker.

Stability: NfL relationships with other biomarkers were generally unstable (max edge frequency with pTau217: 42%), suggesting potential sensitivity to unobserved confounding and warranting cautious interpretation.

 

Interpretation: our findings support the role ofpTau217 as a causally proximal biomarker of cognitive decline and are consistent with prior evidence linking GFAP more closely to amyloid–tau pathology than NfL.

NfL positioning: The tentative upstream placement of NfL relative to GFAP suggests a possible earlier involvement in neuroinflammatory processes than typically assumed.

Implication: This motivates further prospective and longitudinal studies to determine whether the observed causal ordering is stable across disease stages and cohorts.