Peripheral blood samples were obtained by clinical staff at the Department of Neurology, Vienna General Hospital, during routine diagnostic procedures. The study cohort included individuals diagnosed with Alzheimer’s disease (AD, n = 34), mild cognitive impairment in the context of AD (MCI, n = 31), and age- and sex-matched controls without neurological disease (HC, n = 21). Diagnostic classifications were based on NIA-AA criteria for biological AD definition, including neuropsychological testing, magnetic resonance imaging (MRI) and amyloid biomarker assessments (Albert et al., 2011; Jack et al., 2018). Cognitive function was assessed with the Neuropsychological Test Battery Vienna (NTBV) covering attention, language, executive functions, and episodic memory, with age-, education-, and sex-corrected z-scores derived from normative data (Lehrner et al., 2015; Pusswald et al., 2013). Global cognition was evaluated by Mini-Mental State Examination (MMSE) and Wortschatztest (WST) as an estimate of premorbid IQ, and depressive symptoms were assessed using the Beck Depression Inventory-II (BDI-II) (Kühner et al., 2007).

All patients underwent at least T1-weighted, T2-weighted or fluid-attenuated inversion recovery (FLAIR), and diffusion-weighted MRI sequences as part of routine diagnostics to evaluate atrophy, vascular lesions, and to exclude alternative pathologies or diffusion-restricted areas. Amyloid-PET was performed in 29 MCI and 28 AD patients using either [¹¹C] Pittsburgh compound-B (PiB, n = 52) or [¹⁸F] flutemetamol (Vizamyl®, n = 28) on a Siemens Biograph 64 True Point or GE Advance PET scanner. Approximately 400 MBq [¹¹C] PiB or 185 MBq [¹⁸F] flutemetamol were administered intravenously, with image acquisition starting 40 min (PiB) or 90 min (flutemetamol) post-injection and lasting ~20 min, following CT-based attenuation correction (Philippe et al., 2011). Scans were visually rated as amyloid-positive or -negative by an experienced nuclear medicine physician according to manufacturer guidelines.

CSF was collected from 19 MCI and 18 AD patients by lumbar puncture (L3/L4, L4/L5, or L5/S1) into polypropylene tubes and stored at −20 °C until biomarker analysis (Aβ42, pTau181, tTau) or immediately at −80 °C for research purposes. Concentrations of Aβ42, pTau181, and tTau were measured using commercial ELISAs (Innotest, Fujirebio), applying manufacturer cut-offs (Aβ42 < 500 pg./mL, pTau181 > 61 pg./mL, tTau > 300 pg./mL) (Vanderstichele et al., 2000; Vanmechelen et al., 2000). From these measures, the Innotest Amyloid Tau Index (IATI = Aβ42/(240 + 1.18 × tTau)) was calculated, with values <1 indicating AD pathology (Tabaraud et al., 2012).

Only patients with either positive CSF biomarkers or positive Amyloid PET and MRI results were included in the disease study cohort. For 19 MCI patients and 15 AD patients, both PET and CSF data was available. To distinguish sporadic from monogenic early-onset AD, individuals with age at onset ≤65 years and a positive family history who consented to genetic testing underwent whole-exome sequencing (WES). In total, 12 of 15 EOMCI patients (80%) and 14 of 21 EOAD patients (67%) were genetically evaluated, and no pathogenic variants in APP, PSEN1, or PSEN2 were identified; such cases were excluded from the cohort by design.

The remaining early-onset patients who did not undergo sequencing all had a Goldman score (GS) of 3.5–4, which indicates one relative with late-onset dementia (GS 3.5) or no known family history of dementia (GS 4), both of which carry a very low likelihood of autosomal-dominant AD (Goldman et al., 2005).

Further exclusion criteria included active systemic inflammatory or autoimmune diseases. Control participants had no history of neurological disease and no comorbidities known to affect systemic inflammation.

Peripheral blood sampling for RNA extraction was performed as part of the clinical diagnostic work-up. CSF collection and/or PET imaging occurred within the same diagnostic episode, though not necessarily on the same day. No participant was receiving systemic corticosteroids or immunosuppressive therapy at the time of sampling. Detailed information on chronic medications (e.g., statins, antihypertensives, NSAIDs) was not systematically recorded.

The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of the Medical University of Vienna. Written informed consent, including publication of data in scientific journals and deposition in scientific databases, was obtained from all participants in the course of the inclusion in two existing registries: Dementia Registry RDA MUV (EK 1323/2018, approval date: 15.06.2018) and the BIOBANK MUV (EK 2195/2016, approval date: 17.02.2017).

DNA was isolated from whole blood collected in EDTA tubes using standard procedures. The APOE genotype was determined in genomic DNA from patients using TaqMan qPCR with probes for two single nucleotide polymorphisms (SNPs) in the APOE gene, rs429358 and rs7412. Fluorescence of VIC and FAM, tagging the two different SNPs in each probe, was measured, enabling identification of ε2, ε3, and ε4 alleles.

Blood was collected in PAXgene Blood RNA tubes (Qiagen, 762,165) and stored at −20 °C until processing. RNA was extracted using the PAXgene Blood RNA Kit (PreAnalytiX, 762,164) following the manufacturer’s protocol. RNA concentration and purity were assessed via NanoDrop spectrophotometry (Thermo Fisher Scientific), ensuring 260/280 ratios between 1.9 and 2.1. Samples were stored at −80 °C prior to cDNA synthesis. cDNA was synthesized from 500 ng to 1 μg total RNA using the iScript™ cDNA Synthesis Kit (Bio-Rad, 1,708,891) according to the manufacturer’s instructions.

Quantitative PCR (qPCR) was performed in technical duplicates on an AriaMx Real-time PCR System (Agilent), using SYBR Green- and probe-based assays as appropriate. For SYBR Green assays, reactions were assembled with SsoAdvanced™ Universal SYBR® Green Supermix (Bio-Rad, 1725272) and gene-specific primers (Sigma-Aldrich; sequences listed in Supplementary Table 1). The inflammatory and vascular transcripts included in this study were selected as part of a predefined, literature-based mechanistic panel reflecting pathways relevant to CHI3L1 biology. Specifically, IL1B and TNF represent canonical pro-inflammatory cytokines; MMP9 and LRP1 as mediators of matrix remodeling, vascular integrity, and BBB permeability; and TREM2 reflects peripheral myeloid activation pathways associated with genetic AD risk. These analyses were designed as exploratory correlations within this predefined mechanistic panel rather than an unbiased transcriptomic screen. APOE transcript quantification used PrimePCR™ Probe Assays (Bio-Rad) with Luna® Universal Probe qPCR Master Mix (NEB, M3003L), and TREM2 quantification employed a predesigned TaqMan Assay (ThermoFisher, Assay ID C_100657057_10).

Relative gene expression was calculated on the ΔCt scale. GAPDH served as the reference gene for SYBR Green-based assays, while RNase P was used for probe-based assays (TaqMan and PrimePCR). All statistical inference was performed on ΔCt values, with lower ΔCt corresponding to higher transcript abundance. Fold-change values were derived by normalizing each sample’s ΔCt to the median ΔCt of the healthy control group (2^-ΔΔCt) and were log₂-transformed for visualization. As this transformation can amplify variance, particularly at extreme values in small subgroups, fold-change plots are provided for graphical purposes only, whereas all statistical analyses and effect-size estimates were based on ΔCt values.

To verify the appropriateness of normalization procedures, we assessed the stability of the housekeeping genes GAPDH and RNase P across diagnostic groups, APOE ε4 genotypes, and age. Raw Ct values were compared using Kruskal-Wallis tests, Dunn’s post-hoc comparisons, and Spearman correlations, and coefficients of variation (CV) were calculated to quantify dispersion. Both genes showed narrow Ct distributions with low variability (GAPDH: mean 20.55 ± 1.88, CV = 9.1%; RNase P: mean 23.00 ± 1.59, CV = 6.9%), consistent with stable expression across samples. Neither GAPDH nor RNase P Ct values differed between APOE ε4 carriers and non-carriers (all p > 0.05), and Ct values did not correlate with age within HC, MCI, or AD (all p > 0.10). GAPDH Ct values were also stable across diagnostic groups, whereas RNase P showed a modest group-wise effect (Kruskal-Wallis p < 0.01), driven by a significant HC-AD difference in Dunn’s post-hoc testing (adjusted p = 0.0016), while HC-MCI and MCI-AD comparisons were non-significant (Supplementary Figure 2). Because RNase P was used exclusively for normalization of the probe-based TREM2 assay, we performed sensitivity analyses in which TREM2 expression was re-normalized to GAPDH alone and to the geometric mean of GAPDH and RNase P. Across all normalization strategies, CHI3L1-TREM2 correlations remained non-significant, demonstrating that technical differences in reference scaling did not materially influence outcomes.

To assess potential batch-related or temporal drift effects, we examined associations between measurement order and Ct values of GAPDH and RNase P as well as ΔCt values of CHI3L1. GAPDH Ct values showed a weak negative association with measurement order in the AD group [ρ = −0.492, p = 0.003; 95% CI (−0.791, −0.093)], but the corresponding ΔCt values were unaffected [ρ = 0.038, p = 0.829; 95% CI (−0.362, 0.436)], indicating successful normalization. RNase P Ct values showed no significant temporal structure in any group, and no systematic drift was observed for CHI3L1 ΔCt values across diagnostic categories.

RNA quality was assessed in a representative subset of samples (n = 5 per diagnostic group) using High Sensitivity RNA ScreenTape analysis (Agilent). All samples demonstrated integrity acceptable for qPCR analyses (mean RIN = 5.8 ± 2.0) with no group-wise differences. To evaluate potential effects of RNA quality on qPCR performance, correlations between RIN and Ct values for GAPDH, RNase P, and CHI3L1 were examined; no significant associations were observed (all p > 0.05; 95% CIs spanning zero), indicating that variability in RNA integrity did not influence housekeeping gene amplification or relative quantification. Amplification efficiency (E) was determined from a 5-point dilution series and accepted if 90–110% (R² ≥ 0.99). Primer specificity for SYBR Green assays was verified by melt curve analysis. Duplicate reactions with SD > 0.5 Ct were repeated, and no-template as well as no-reverse transcription controls were included on each plate. Samples from all diagnostic groups (HC, MCI, and AD) were distributed across plates to avoid group-plate confounding. All plates were processed under identical cycling conditions, reactions were run in technical duplicates, and ΔCt normalization minimized plate-to-plate variation. Amplification efficiencies were consistent across plates, confirming robust assay performance.

Statistical analyses were conducted using GraphPad Prism 8 (GraphPad Software, San Diego, CA).

Analysis hierarchy and exploratory framework: To align analytic complexity with the limited cohort size (n = 86), analyses were structured according to a predefined hierarchy.

Primary analysis: Group-wise comparison of CHI3L1 ΔCt stratified by APOE ε4 carrier status and AAO (Mann Whitney U; corrected with Holm-Šidák).

Secondary analyses:

All stratified and correlation results are interpreted as exploratory, with emphasis on effect sizes and confidence intervals rather than statistical significance.

Data normality was assessed using the Shapiro–Wilk test. Given the non-normal distribution of gene expression data (ΔCt values), non-parametric tests were applied for group comparisons and correlations. All analyses were performed on ΔCt values, which represent relative cycle differences.

The final cohort consisted of 31 biomarker-confirmed MCI patients, 34 biomarker-confirmed AD patients, and 21 age- and sex-matched HC (Table 1). For consistency, all statistical tests were based on ΔCt values, with lower ΔCt reflecting higher expression. Fold-change values shown in figures reflect only visual scaling and were not used for hypothesis testing. To evaluate potential demographic influences, we first assessed the relationship between CHI3L1 ΔCt values, age at blood draw, and sex. In unadjusted analyses, no significant sex differences were observed within any diagnostic group, and bivariate correlations with age were weak (Supplementary Figure 1). However, when considered in a multivariable regression model including diagnostic group, APOE status, age, and sex, age emerged as a significant predictor of CHI3L1 ΔCt (β = +0.064, p = 0.004, q = 0.020), indicating lower CHI3L1 expression with increasing age. Sex, APOE ε4 status, and diagnostic cohort were not significant independent predictors (all p > 0.44). Likewise, the APOE ε4 × cohort interaction showed no association (β = −1.058, p = 0.060, q = 0.150) (see Supplementary Table 2A).

Comparison of CHI3L1 transcript levels across HC, MCI, and AD groups revealed no significant differences (p = 0.174). Higher median CHI3L1 ΔCt values (indicating lower expression) within the MCI group relative to HC and AD did not reach significance (HC vs. MCI, p = 0.486 and MCI vs. AD, p = 0.241 after Dunn’s correction; Figure 1A).

To further investigate whether age at onset (AAO) modulates peripheral CHI3L1 expression, we stratified the MCI and AD cohorts into early-onset (EO; AAO < 65 years), and late-onset (LO; AAO ≥ 65 years) subgroups. Kruskal-Wallis testing revealed a significant overall group effect when comparing HC (n = 21), EOMCI (n = 15), LOMCI (n = 16), EOAD (n = 21), and LOAD (n = 13) groups (H = 12.02, p = 0.017), with an effect size of η² = 0.10 (95% CI 0.00 to 0.13). This effect was primarily attributable to increased CHI3L1 expression in the EOAD subgroup, which exhibited the lowest ΔCt values (indicating highest expression levels) across all groups (Figure 1B). Although post hoc pairwise comparisons did not reach significance after multiple testing correction, the expression pattern suggests selective upregulation of peripheral CHI3L1 in EOAD. Given the modest cohort size, these subgroup patterns are interpreted as exploratory and descriptive rather than confirmatory.

We subsequently investigated whether APOE ε4 carrier status modulates peripheral CHI3L1 expression. All statistical tests are based on ΔCt values, with lower ΔCt reflecting higher expression. Fold-change values shown in figures reflect only visual scaling and were not used for hypothesis testing. In HC and MCI groups, CHI3L1 expression did not differ between APOE ε4 + (HC, n = 5; MCI, n = 18) and APOE ε4- (HC, n = 16; MCI, n = 13), indicating no genotype effect during prodromal dementia stages. While median CHI3L1 ΔCt values (indicating higher expression) within the AD group were lower in APOE ε4 carriers (n = 23; median = 0.60, IQR: −1.20 to 3.4) compared to non-ε4 carriers (n = 12; median = 3.25, IQR: 1.30 to 4.10), significance was lost after Holm-Šidák correction (p = 0.061; Hodges-Lehmann median difference = −1.7; 95% CI = −3.6 to 0; Figure 2A). Because the global multivariable regression (Supplementary Table 2A) indicated that APOE ε4 effects may differ across diagnostic groups, we performed stratified regression analyses in EOAD and LOAD. In EOAD, APOE ε4 carrier status was the only significant predictor of CHI3L1 expression (β = −2.155, 95% CI -4.019 to −0.291, p = 0.025), however, significance was lost after FDR correction for multiple testing (q = 0.075). In contrast, no predictors - including APOE ε4 - were significant in the LOAD subgroup (all p > 0.30; Supplementary Table 2B).

To investigate the interaction between APOE ε4 genotype and AAO, we stratified the MCI and AD cohorts into EO and LO subgroups. Mann–Whitney U tests with Holm-Šidák correction revealed no significant differences in CHI3L1 expression between APOE ε4 + and APOE ε4- carriers in EOMCI (ε4+: n = 6, median = 2.90, IQR: 1.73 to 3.43; ε4-: n = 9, median = 2.80, IQR: 2.45 to 4.10), LOMCI (ε4+: n = 12, median = 2.75, IQR: 0.93 to 3.80; ε4-: n = 4, median = 1.90, IQR: 1.10 to 3.08), or LOAD (ε4+: n = 9, median = 3.45, IQR: 0.85 to 4.33; ε4-: n = 4, median = 2.50, IQR: 1.30 to 4.38) subgroups (all p > 0.999).

Notably, in the EOAD subgroup, APOE ε4 + showed significantly lower ΔCt values (higher expression) of CHI3L1 expression compared to their non-carrier counterparts (ε4+: n = 13, median = −0.20, IQR: −1.95 to 1.20; ε4-: n = 8, median = 3.30, IQR: 0.73 to 4.10; p = 0.026; Hodges-Lehmann median difference = −3.3; 95.54% CI -5.3 to −0.4; Figure 2B). To assess robustness given the small subgroup sizes, we performed non-parametric bootstrap resampling (10,000 replicates) to derive confidence intervals for both the Hodges-Lehmann shift and Hedges’ g. The bootstrap-based Hodges-Lehmann estimate (−2.6; 95% CI -4.5 to −0.4) showed a comparable magnitude and direction, confirming the stability of the median difference across computational methods. The corresponding standardized effect size was large (Hedges’ g = −1.09; 95% bootstrap CI -2.61 to −0.26). Collectively, these results support the robustness of the observed genotype effect despite limited subgroup size.

We examined associations between CHI3L1 and transcripts from a predefined, literature-based mechanistic panel of disease-related genes as hypothesis-driven exploratory analyses to characterize peripheral inflammatory pathways potentially linked to CHI3L1 expression. For this we performed Spearman correlation analyses and FDR was controlled using the Benjamini-Hochberg procedure.

Genes of interest included classical inflammatory mediators (IL1B, TNF, TREM2) and BBB-associated factors (LRP1, MMP9). Spearman correlations were performed on ΔCt values, with lower ΔCt reflecting higher expression.

In the HC group, no significant correlations were observed between CHI3L1 expression and any of the examined markers. In the MCI cohort, CHI3L1 expression correlated positively with MMP9 (r = 0.631, q < 0.001; CI 0.33 to 0.82). In AD, CHI3L1 exhibited positive correlations with MMP9 (r = 0.663; q < 0.0001; CI 0.41 to 0.82), IL1B (r = 0.480; q = 0.007; CI 0.17 to 0.71), TNF (r = 0.395; q = 0.024; CI 0.06 to 0.65) and LRP1 (r = 0.745; q < 0.0001; CI 0.54 to 0.87).

Stratification by AAO revealed no significant correlations in EOMCI. In LOMCI, CHI3L1 correlated with TNF (r = 0.690; q = 0.008; CI 0.29 to 0.89) and MMP9 (r = 0.780; q = 0.008; CI 0.39 to 0.93). Within the AD group, correlations were observed in EOAD with TNF (r = 0.480; q = 0.040; CI 0.06 to 0.76) LRP1 (r = 0.785; q < 0.0001; CI 0.53 to 0.91) and MMP9 (r = 0.805; q < 0.0001; CI 0.56 to 0.92), but not in LOAD (Table 2).

When stratifying by APOE ε4 status, no significant correlations were identified in MCI APOE ε4 − individuals. In contrast, MCI APOE ε4 + participants showed correlations between CHI3L1 and IL1B (r = 0.753; q = 0.003; CI 0.41 to 0.91), TNF (r = 0.662; q = 0.005; CI 0.27 to 0.87), LRP1 (r = 0.536; q = 0.027; CI 0.08 to 0.81), TREM2 (r = −0.511; q = 0.030; CI -0.79 to −0.04), and MMP9 (r = 0.746; q = 0.003; CI 0.38 to 0.91) (Figures 3A–E). In the AD cohort, no significant associations were found in APOE ε4 − individuals. In APOE ε4 + AD individuals, CHI3L1 correlated with IL1B (r = 0.582; q = 0.009; CI 0.21 to 0.81), LRP1 (r = 0.799; q < 0.0001; CI 0.57 to 0.91), and MMP9 (r = 0.585; q = 0.009; 0.19 to 0.82) (Figures 3F–J).

Notably, exploratory sex-stratified analyses revealed that the association between CHI3L1 expression and inflammatory markers was strongest in female APOE ε4 + individuals. In this subgroup, CHI3L1 expression correlated significantly with MMP9 (r = 0.788; q = 0.010; CI 0.37 to 0.94), IL1B (r = 0.705; q = 0.015; CI 0.24 to 0.91), TNF (r = 0.757; q = 0.010; CI 0.34 to 0.93), and TREM2 (r = −0.646; q = 0.025; CI -0.89 to −0.13) in MCI (Supplementary Figures 3A–E) and with IL1B (r = 0.726; q = 0.011; CI 0.30 to 0.91), LRP1 (r = 0.833; q = 0.002; CI 0.53 to 0.95), and MMP9 (r = 0.706; q = 0.022; CI 0.20 to 0.91) in AD (Supplementary Figures 3F–J). These correlations were not observed in males or in APOE ε4- females, suggesting a potential sex-by-genotype interaction in peripheral immune activation (Supplementary Table 3).

Exploratory associations with CRP and lipids showed that CHI3L1 expression did not correlate with CRP or lipid markers in healthy controls or MCI (all p > 0.05), arguing against baseline systemic inflammation or dyslipidemia as primary drivers of group differences. In AD, modest inverse correlations were observed with total cholesterol (Spearman r = −0.476, q = 0.032; CI -0.74 to −0.09), as well as HDL (r = −0.444, q = 0.045; CI -0.73 to −0.03) and LDL (r = −0.532, q = 0.032; CI -0.78 to −0.14), whereas no significant associations were detected in EOAD or LOAD when analyzed separately. These data suggest limited confounding by subclinical inflammation in controls and prodromal disease, with evidence for an inflammatory-metabolic interplay in established AD (Supplementary Table 4).