From 3 January 2022 through 30 April 2024 every adult aged 60–79 years who presented to the Department or who was enrolled in one of community-based chronic-disease management programmes affiliated with the First Hospital of Anhui Medical University was screened prospectively for eligibility. Inclusion criteria were: (1) continuous residence within Hefei municipality for the preceding 5 years to ensure stable exposure assessment; (2) capacity to comprehend Mandarin or a local dialect sufficiently to complete interviewer-administered surveys; and (3) Mini-Mental State Examination (MMSE) score ≥ 24, ensuring reliable self-report. Individuals were excluded for: prior ischaemic stroke, intracerebral haemorrhage, or significant traumatic brain injury; congenital cerebrovascular malformation or intracranial neoplasm; current employment in heavy industry or mining; lifetime tobacco exposure ≥ 5 pack-years (pack-years = packs per day × years; lifetime non-smokers and < 5 pack-years were eligible); and any change to antihypertensive medication within the 6 months preceding enrolment. These criteria reduced extreme-age heterogeneity, limited dominant vascular risks unrelated to ambient exposure, and minimised confounding by secondary-prevention physiology or acute pharmacological fluctuations in blood-pressure variability. Of consecutively screened adults, 460 participants met all eligibility requirements and were included in the analytic cohort after written informed consent. This research protocol has been reviewed by the Ethics Committee of the First Affiliated Hospital of Anhui Medical University. After review, the Ethics Committee of the First Affiliated Hospital of Anhui Medical University determined that this study involves only anonymous questionnaire surveys and therefore does not require approval from the Ethics Committee to proceed.

Item generation followed a three-step process. First, an interdisciplinary panel of neurologists, environmental hygienists, and sleep psychologists reviewed international guidelines and PubMed/Embase literature (January 2000 – June 2021) to draft an initial pool of 72 items covering environmental exposure, sleep quality and conventional cerebral-haemorrhage risk modifiers. For this study, environmental exposure was defined a priori as the cumulative (chronic) inhaled burden of ambient and indoor air pollutants over months to years, operationalised by long-term residential conditions and habitual behaviours. The AEEQ therefore prioritises cumulative rather than episodic (acute) exposure: short-term peaks are noted where relevant but the instrument is designed to capture the long-run dose that plausibly contributes to chronic cerebrovascular stress. Second, two modified Delphi rounds involving 15 external experts were conducted; items with a median relevance rating < 4 on a 5-point Likert scale or with an inter-quartile range > 1.5 were dropped. Third, cognitive interviewing with older adults refined wording for comprehension and cultural resonance. The final 42-item Anhui Environmental Exposure Questionnaire (AEEQ) deliberately reflects item-pool parsimony: simulation and pilot timing indicated that approximately 40–45 items balanced respondent burden (~10–12 min) against content breadth across six domains. Retaining materially overlapping items would inflate completion time with minimal gain in construct coverage, whereas reducing the pool below ~35 items risked under-representation of key behavioural modifiers and sleep constructs. Within the ambient-exposure domain, mask use, window-opening habits, and stove-fuel choices were conceptualised as modifiers of pollutant infiltration and effective inhaled dose (e.g., altering indoor/outdoor penetration fraction, source intensity, and personal filtration), not as exposures in their own right; scoring therefore treats these behaviours as dose-attenuating or dose-amplifying features. Sleep items incorporated irregular bedtimes and evening chronotype to operationalise circadian misalignment, a hypothesised mechanism potentiating sympathetic surges and blood-pressure variability that complements the sleep-quality narrative. While a single item probes perceived traffic noise frequency for contextual awareness, the AEEQ does not include a stand-alone noise-exposure module or objective acoustic measurement; this was a pragmatic decision to limit respondent burden and because prior work suggests modest incremental variance after particulate metrics in similar settings—noise will be considered in future iterations.

Item-level content-validity indices (I-CVI) and the scale-level CVI (S-CVI/Ave) were computed from the second Delphi round. An a-priori threshold of ≥ 0.78 for I-CVI and ≥ 0.90 for S-CVI signified adequate expert agreement. Face validity was assessed qualitatively during cognitive interviews and via a 5-point global clarity rating completed by interviewers after pilot administration; mean scores > 4 signalled satisfactory readability.

Construct validity was examined through a split-sample strategy. The cohort was randomised 1:1 into exploration (n = 230) and confirmation (n = 230) using a computer-generated sequence. In the exploration sample, exploratory factor analysis with principal-axis extraction and promax rotation identified latent dimensions; factors with eigenvalues > 1.0 and items with loadings ≥ 0.40 were retained. Confirmatory factor analysis was then performed in the hold-out sample with maximum-likelihood estimation. Model fit was adjudicated on χ²/df < 3, comparative-fit index (CFI) ≥ 0.95, Tucker–Lewis index (TLI) ≥ 0.95, root-mean-square error of approximation (RMSEA) ≤ 0.06, and standardised root-mean-square residual (SRMR) ≤ 0.08.

Internal consistency was quantified by Cronbach’s α for each domain and for the total scale, with ≥ 0.70 considered acceptable. Split-half reliability employed the Guttman coefficient. Test–retest stability was evaluated in a convenience subsample of 212 participants who completed the AEEQ a second time 14 ± 3 days after baseline; two-way mixed-effects intraclass correlation coefficients (ICC (1, 2)) with 95% confidence intervals were calculated, and values ≥ 0.75 denoted good repeatability.

Concurrent validity was explored by correlating AEEQ air-pollution domain scores with annual-mean residential PM₂.₅ concentrations derived from a land-use regression model (1 km × 1 km resolution) for the calendar year preceding baseline, and by correlating the sleep-quality domain with the Pittsburgh Sleep Quality Index (PSQI). Annual means were selected a priori to align with the questionnaire’s chronic-exposure emphasis and to reflect cerebrovascular stress accumulation over months to years more faithfully than seasonal or monthly averages. Pearson or Spearman coefficients, as appropriate, were interpreted as follows: < 0.30 weak, 0.30–0.49 moderate, ≥ 0.50 strong. Objective noise metrics were not paired with the questionnaire in this iteration because of data-collection constraints and anticipated limited incremental variance beyond PM₂.₅ in this setting.

Questionnaire-validation guidelines recommending a minimum of 5–10 participants per item informed recruitment targets; with 42 items, a lower bound of 210 respondents was required for factor analysis. Although initial screening exceeded this benchmark, post-review eligibility refinements yielded an analytic sample of 460, which remains comfortably above recommended ratios for both exploratory and confirmatory factor analysis and permits assessment of reliability and criterion validity with adequate precision.

Trained interviewers administered the AEEQ in a quiet clinic room, recording responses on tablet computers equipped with range and logic checks. The MMSE was administered prior to questionnaire completion to ensure cognitive eligibility. Blood pressure, body-mass index, and medication lists (including antihypertensive classes and any regimen changes within 6 months) were abstracted from electronic medical records on the same day. Smoking history was captured via standardised items allowing pack-year computation to enforce the < 5 pack-year threshold. Occupational history characterised current or recent heavy-industry/mining employment. Geographic coordinates of each residence were geocoded for linkage with satellite-based pollution data. To address environmental inequity, interviewers additionally recorded proxies of socio-economic position and housing type (education, household income bands, housing tenure and building type), recognising that these characteristics may concentrate higher AEEQ scores in disadvantaged groups with public-health ramifications. Data entry accuracy was verified by double extraction of a 5% random sample, achieving < 0.2% discrepancy.

Analyses were conducted with Stata 18.0 and Mplus 8.9. Continuous variables are summarised as mean ± SD or median (inter-quartile range), categorical variables as counts and percentages. Bartlett’s test of sphericity and the Kaiser–Meyer–Olkin statistic appraised factorability. For internal consistency, 95% CIs around α were generated by bootstrapping (1,000 replications). Test–retest ICCs employed two-way mixed models with absolute-agreement definition. Hypothesised correlations for criterion validity were tested via two-tailed significance with Bonferroni correction for the dual comparisons (α = 0.025). Missing item responses (< 1% overall) were imputed using domain-specific means when ≤ 10% of a domain’s items were blank; questionnaires exceeding that threshold were deemed incomplete and excluded from analysis. In modelling frameworks, mask use, window-opening, and stove-fuel items were treated conceptually as dose modifiers that influence pollutant infiltration and personal filtration rather than as independent exposures; this rationale guided their role within the composite AEEQ score. Given biologic plausibility that antihypertensive therapy may buffer pollutant-induced blood-pressure variability, prespecified interaction terms between the AEEQ composite and antihypertensive use were tested, with supportive stratified analyses. To probe distributional equity, socio-economic and housing variables were included as covariates and examined for effect modification. Finally, we acknowledged classical measurement-error bias: any non-differential misclassification in questionnaire responses would be expected to bias associations toward the null, underscoring the conservative nature of estimated hazard ratios.

Between 3 January 2022 and 30 April 2024, 4,326 older adults were screened; after applying the revised eligibility criteria (age 60–79 years; continuous ≥ 5 years residence; MMSE ≥ 24; lifetime non-smoker or < 5 pack-years; no prior ischaemic stroke, intracerebral haemorrhage or significant traumatic brain injury; not currently employed in heavy industry/mining; and no change to antihypertensive medication within the preceding 6 months), 460 participants were enrolled and analysed. Median residence duration within Hefei municipality was 16 years (IQR 13–20). Questionnaire completeness was high: < 1% of all items were missing; 3 questionnaires (0.7%) required domain-mean imputation for ≤ 10% of items per the a-priori rule; no questionnaire exceeded the exclusion threshold. Baseline characteristics are summarised in Table 1: mean age 68.2 ± 5.6 years; 214 (46.5%) men; body-mass index 24.2 ± 3.1 kg m⁻²; hypertension 281 (61.1%); diabetes 92 (20.0%); current smoker 28 (6.1%) (all < 5 pack-years); ≥ 1 alcoholic drink per week 149 (32.4%). Annual-mean residential PM₂.₅ (calendar year prior to baseline) was 46.9 ± 5.8 μg m⁻³. A total of 239 (52.0%) participants were on long-term antihypertensive therapy at baseline.

Delphi round-2 appraisal yielded item-level CVIs ranging 0.86–1.00 (median 0.94), with S-CVI/Ave = 0.96, surpassing a-priori thresholds (I-CVI ≥ 0.78; S-CVI/Ave ≥ 0.90). Interviewer global-clarity ratings during cognitive testing averaged 4.6 ± 0.3 on the 5-point scale. Table 2 presents domain-specific indices.

Factorability prerequisites were satisfied (Kaiser–Meyer–Olkin 0.91; Bartlett’s χ² 3,085, df = 861, p < 0.001). In the exploration sample (n = 230), exploratory factor analysis (principal-axis, promax) recovered six factors with eigen-values > 1.0, together explaining 66.1% of total variance. The first three factors were labelled ambient air-pollution, indoor air-quality/fuel use, and sleep quality; eigen-values were 8.10 (19.3%), 5.61 (13.4%), and 4.76 (11.3%), respectively, with the remaining factors (cardiometabolic risk modifiers, lifestyle, medication/comorbidity) contributing 3.48 (8.3%), 3.02 (7.2%), and 2.50 (6.0%) (see Table 3). Primary loadings ≥ 0.40 were observed for 41/42 items; a single sleep-maintenance item (night-time bathroom visits) loaded at 0.39 and was retained for clinical relevance (see Table 4). The scree plot is shown in Figure 1. Confirmatory factor analysis in the validation sample (n = 230) showed excellent fit (χ²/df 2.19; CFI 0.965; TLI 0.958; RMSEA 0.041, 90% CI 0.036–0.047; SRMR 0.052). Standardised latent-factor correlations ranged from 0.22 to 0.45, consistent with related but non-redundant constructs (see Tables 5, 6).

Internal-consistency estimates met or exceeded the a-priori criterion (α ≥ 0.70) for every domain. Cronbach’s α (with 95% bootstrap CIs; 1,000 replications) were: ambient air-pollution 0.86 (0.83–0.88), indoor air-quality/fuel 0.83 (0.79–0.86), sleep quality 0.84 (0.81–0.87), cardiometabolic modifiers 0.80 (0.76–0.84), lifestyle 0.78 (0.74–0.82), medication/comorbidity 0.79 (0.75–0.83); total scale 0.90 (0.88–0.92). Guttman split-half coefficients paralleled these values (0.77–0.87; total 0.88). In the 14 ± 3-day retest subsample (n = 212), two-way mixed-effects ICCs indicated good-to-excellent stability: domains 0.83–0.88; total scale 0.86 (0.82–0.89). Details appear in Table 7.

As prespecified, the AEEQ air-pollution exposure domain correlated strongly with annual-mean residential PM₂.₅ (Pearson r = 0.62, 95% CI 0.560–0.673, p < 0.001), and the sleep-quality domain correlated inversely with PSQI (Spearman ρ = −0.56, 95% CI − 0.620 to −0.494, p < 0.001). Using the dose-modifier specification (mask use, window-opening, stove-fuel) embedded in scoring improved the air-pollution domain’s correlation with PM₂.₅ by Δr = 0.06 relative to an unmodified specification. Criterion-validity coefficients are summarised in Table 8.

Over a median 24 months (IQR 19–28) of follow-up (≈ 920 person-years), 16 incident cerebral haemorrhages were confirmed by neuro-imaging (3.5%). Event incidence rose monotonically across quartiles of the baseline AEEQ composite (quartiles defined within the 460-participant cohort; 115 participants per quartile):

Q1 (lowest exposure) 1/115 (0.9%); Q2 2/115 (1.7%); Q3 3/115 (2.6%); Q4 (highest) 10/115 (8.7%); χ² trend = 12.95, p = 0.0047. Assuming roughly equal person-time per quartile (~230 person-years each), the corresponding incidence rates were 4.3, 8.7, 13.0, and 43.5 per 1,000 person-years (see Table 9). In multivariable Cox models adjusted a-priori for age, sex, hypertension, alcohol use, and anticoagulant therapy, each 1-SD increase in the AEEQ composite was associated with a 47% higher hazard of cerebral haemorrhage (HR 1.47, 95% CI 1.12–1.93). The time-dependent C-statistic for the reference clinical model was 0.72 (0.66–0.78) and improved to 0.80 (0.74–0.85) after adding the continuous AEEQ Z-score (likelihood-ratio p < 0.001); the net reclassification improvement (NRI) was 0.15 (p = 0.045) (see Table 10; Figure 2). A prespecified drug-exposure interaction was observed: the per-SD AEEQ–haemorrhage association was stronger among participants not using antihypertensives (HR 1.62, 95% CI 1.10–2.39) than among antihypertensive users (HR 1.31, 95% CI 0.95–1.81); interaction p = 0.048. Treating mask use, window-opening and stove-fuel explicitly as dose modifiers within the composite (as per Methods) produced slightly steeper exposure–response slopes than treating them as stand-alone exposures (ΔHR per SD + 0.06).

Findings were unchanged after excluding the 3 participants with domain-level imputation (HR per SD = 1.49, 95% CI 1.13–1.96), and when the AEEQ was modelled as tertiles rather than quartiles (tertile-3 vs. tertile-1 adjusted HR = 2.71, 95% CI 1.09–6.72; p-trend = 0.008). Proportional-hazards assumptions were not violated (global Schoenfeld p = 0.36). To address environmental inequity, we profiled AEEQ distributions by socio-economic position and housing type. Participants in the lowest education stratum (≤ primary) had higher mean AEEQ composites than those with ≥ secondary education (mean difference 0.42 SD, 95% CI 0.21–0.63), and renters exceeded homeowners by 0.35 SD (95% CI 0.12–0.58). In a mutually adjusted model, both education and tenancy remained independent correlates of higher AEEQ (both p < 0.01). Results are detailed in Table 11. Because noise exposure was not measured as a stand-alone module in this iteration (per Methods), the single traffic-noise item served only contextual purposes; it correlated moderately with the ambient domain (r = 0.41) but did not materially improve fit indices or predictive performance when added (ΔCFI +0.002; ΔAUC +0.003).