Plasma GFAP vs. NfL for AD vs. other dementias

Answer

Across five verified studies, plasma NfL and GFAP disagree on which better discriminates Alzheimer's disease (AD) from frontotemporal dementia/FTLD, and the disagreement is not explained within any single study. In a Spanish cohort, NfL alone outperformed GFAP alone for AD vs FTD (AUC 0.78 vs 0.70), and their ratio did better than either (GFAP/NfL AUC 0.86) 66a0e27c. In a Chinese cohort, GFAP alone was the stronger single discriminator for AD vs FTLD (AUC 0.894), with plasma GFAP roughly 2-fold higher in AD than in FTLD, while NfL was elevated across all patient groups but was not the top discriminator 2a6e0f94. Two further cohorts show NfL, not GFAP, consistently selected into the best-performing multi-marker panel for AD vs FTD (AUC 0.90-0.94), with GFAP entering the optimal panel in only one of two replication cohorts f5e2560c, and a large Bayesian analysis found NfL the stronger single biomarker overall (mean AUC 0.85) while GFAP did not top any single-marker ranking fd40c9da. A fifth cohort did not isolate a single-marker AUC for AD vs FTLD but found combined plasma+neurophysiology panels reaching AUC 0.98, with GFAP differing significantly between groups c99492e1. No verified source reports an NfL fold-change for this contrast; the only structured fold-change figure in the corpus is GFAP's ~2-fold elevation in AD vs FTLD 2a6e0f94.

Summary

The corpus does not settle whether GFAP or NfL is the better single biomarker for separating AD from FTD/FTLD: one cohort favors NfL, another favors GFAP, and two others favor NfL indirectly through panel-selection frequency. What is consistent across cohorts is that a combined panel, or the GFAP/NfL ratio specifically, outperforms either marker alone wherever it was tested. Readers should not treat either marker as a clear individual winner; the corpus supports using them together, or as a ratio, rather than picking one.

Evidence

• Gomez-Tortosa et al. 2025 - Head-to-head plasma NfL vs GFAP in 108 AD, 73 FTD, 54 controls: NfL AUC 0.78 vs GFAP AUC 0.70 for AD vs FTD; GFAP/NfL ratio AUC 0.86. 66a0e27c
• Sun et al. 2025 - 120 dementia patients (51 AD, 54 SIVD, 15 FTLD) plus 27 controls: GFAP AUC 0.894 for AD vs FTLD, ~2-fold higher plasma GFAP in AD than FTLD; NfL elevated across all groups but not top discriminator. 2a6e0f94
• Thijssen et al. 2022 - Two independent cohorts (n=160, n=152): NfL selected into the optimal AD-vs-FTD panel in both cohorts (AUC 0.90-0.94); GFAP selected in only one. f5e2560c
• Costa et al. 2026 - Bayesian classification, 97 AD / 255 FTD / 70 HC, 41 model combinations: NfL top single biomarker overall (mean AUC 0.85); pTau217+NfL best two-marker AD-vs-FTD panel (AUC 0.90). fd40c9da
• Benussi et al. 2022 - 202 participants (AD, FTLD, controls), plasma + TMS neurophysiology: best AD-vs-FTLD model (combined panel) AUC 0.98; no single-marker GFAP-vs-NfL AUC isolated for this exact contrast. c99492e1

Methodology

Single-marker AUCs for the AD-vs-FTD/FTLD contrast were compared only where a study reported both GFAP and NfL individually for that exact contrast (Gomez-Tortosa, Sun). Studies reporting only multi-marker panel AUCs (Benussi, Costa, Thijssen) were used to establish which marker is selected into the winning panel, not pooled with the direct head-to-head AUCs, since panel selection and standalone discrimination are different measurements and cannot be averaged together. The single structured fold-change figure was reported as-is rather than generalized to NfL, for which no comparable figure exists in the corpus.

Limitations

• Only two of five sources (Gomez-Tortosa, Sun) report an isolable single-marker AUC for GFAP vs NfL in this exact contrast, and they disagree on both the values and the ranking; the corpus cannot say which is more generalizable.
• No verified source reports an NfL fold-change (AD vs FTLD/FTD) to compare against GFAP's ~2-fold figure.
• Cohorts differ by ethnicity, FTLD subtype mix, and control composition, any of which could drive the disagreement, but no verified source in this corpus tests these confounds directly.