We performed a benchmark consistency study of a synthetic persona dataset intended for auxiliary public health informatics and surveillance planning [12,13]. The primary aim was not to predict CHS municipal-level unmet health care rates, but to assess whether synthetic persona rows reproduced sociodemographic distributions observed in CHS [16,17]. We compared persona distributions with CHS complex-survey benchmarks and selected KOSIS aggregate indicators that also informed persona generation [18]. Table 1 summarizes the benchmark comparison layers, data sources, analytic metrics, and intended interpretation.

The synthetic persona dataset was treated as the validation target. We used NVIDIA Nemotron-Personas-Korea, an open-source Korean-language synthetic persona dataset released under a CC-BY 4.0 license [15]. Version 1.0 was released on April 20, 2026. The dataset card published by NVIDIA on Hugging Face describes 1 million rows, with structured fields including sex, age, marital status, family type, education level, occupation, district, province, and country, as well as 7 narrative persona text fields [15]. The analytic unit in this study was the row-level structured demographic record, not the narrative text fields. The dataset card identifies the following sources: KOSIS census tables covering sex, province, industry, occupation, travel, and leisure; Supreme Court data on names, birth year, and sex; National Health Insurance Service Health Checkup Information as of December 31, 2024; materials from the Korea Rural Economic Institute’s 2024 Food Consumption Behavior Survey; and seed data and domain expertise contributed by NAVER Cloud during dataset design [15]. It does not provide a common reference year for every source. Consistent with synthetic data evaluation guidance, the dataset was evaluated as a potential auxiliary analytic layer, not as a record-level representation of real individuals or an official population estimate [12,13].

The CHS 2021-2025 was used as the main survey benchmark because CHS is designed to produce local community health statistics using a complex survey design [16,17]. The final 2025 CHS microdata were officially released on February 25, 2026 [17]. We included all 1,156,122 adult respondents with positive person weights and valid design identifiers: 229,242 in 2021, 231,785 in 2022, 231,752 in 2023, 231,728 in 2024, and 231,615 in 2025. CHS person weights were used to estimate weighted benchmark distributions [19]. The pooled design comprised 28,309 analytic strata and 112,299 analytic primary sampling units (PSUs). Survey year was appended to the CHS stratum and cluster identifiers. The unmet health care denominator was used only for analyses of the unmet health care indicator.

KOSIS public statistics–based indicators were used as seed-source–informed aggregate benchmarks because KOSIS provides official Korean statistical indicators across population and regional domains [18]. This analysis used 2 municipal-level indicators that were directly relevant to vulnerable-profile surveillance and available in the local analysis files: aged population share and older-adult living-alone household share. In this study, municipal-level areas correspond to Korean si/gun/gu units, including cities, counties, districts, and wards. The aged population indicator used 2025 KOSIS counts of persons aged 65 years or older divided by the total population. The older-adult living-alone indicator used the KOSIS older-adult one-person-household percentage; the local KOSIS export available for this indicator was 2024 and was carried forward into the 2025 comparison table. The KOSIS-persona comparison used 252 matched municipal-level areas: 216 exact si/gun/gu matches, 35 parent-city values repeated across administrative-gu rows where KOSIS reported a parent city, and one Sejong province-city match. Indicator definitions and linkage notes are provided in Table S2 in Multimedia Appendix 1. Because KOSIS denominators differ from persona adult rows and because KOSIS/public statistics were among the persona seed sources, KOSIS comparisons were interpreted as seed-source–informed regional ranking checks rather than fully independent absolute-level validation.

Common core variables were harmonized across CHS and persona records: current province code, current municipal-level area code, age group, sex, marital status, living-alone status, and coarse education. These variables were selected because they are available in both the persona dataset and CHS and correspond to common demographic and socioeconomic stratifiers used in health inequality monitoring [3,15-17]. Age was grouped into the following 5 categories: 19 to 29 years, 30 to 44 years, 45 to 64 years, 65 to 74 years, and 75 years or older. Education was classified into low, middle, and high categories. Economic activity and occupation were retained as sensitivity variables rather than core variables.

Economic activity and occupation required special handling. Persona economic activity was derived from generated occupation text, and persona occupation was a generated detailed job-title string, whereas CHS uses survey-coded economic activity and occupation variables [15-17]. We therefore excluded both economic activity and occupation from the core manuscript analysis and treated them as sensitivity variables. A separate occupation recoding quality assurance compared the original keyword mapping with an expanded sensitivity mapping.

Detailed variable mapping is provided in Table S1 in Multimedia Appendix 1. Briefly, region was harmonized to current province and municipal-level area codes, with historical code changes and renamed or reorganized administrative areas mapped to current codes through a reviewed crosswalk, age was grouped using identical cutoff points, sex and marital status were mapped to shared categories, and education was classified into 3 analytically comparable levels. CHS living-alone status was derived from household variables, and persona living-alone status was derived using a prespecified exact-match rule for the Korean-language family-type label denoting living alone. All 39 distinct family-type labels were audited, and the only ambiguous label was evaluated using alternative recoding rules. Missing or unmapped values were excluded only from comparisons involving the affected variable, and each distribution was normalized over its comparison-specific denominator. All 1 million persona rows were aged 19 to 99 years, so the adult restriction excluded no persona rows.

For each benchmark comparison, we computed CHS weighted percentages and persona row percentages. Distributional consistency was summarized using percentage-point differences, standardized differences, total variation distance (TVD), and Jensen-Shannon divergence [20,21]. TVD was used as the main scalar summary because it quantifies the total amount of distributional mass that would need to be shifted for the 2 distributions to match [20]. For example, a TVD of 0.05 indicates that 5% of the distributional mass would need to be reassigned across categories for the CHS and persona distributions to match [20]. Prior synthetic data evaluation work has used TVD across univariate, bivariate, and 3-way empirical marginal distributions as a model-free measure of synthetic data fidelity, with holdout data used as a reference for sampling noise rather than a fixed universal cutoff [9,12,13]. Accordingly, we interpreted TVD descriptively and comparatively, focusing on whether discrepancies increased from marginal distributions to higher-dimensional profiles and whether they persisted after sparse-cell restrictions. We did not apply a universal TVD threshold to declare 2 empirical distributions as equivalent or nonequivalent [9,12,13].

CHS cell-level sampling uncertainty was estimated using a Taylor-linearized ratio estimator where the number of cells was computationally tractable, consistent with design-based analysis of complex health surveys [19]. For each benchmark cell, the weighted cell proportion was estimated as the ratio of the cell-specific weighted total to the total weighted denominator. The linearized residual for each analytic primary sampling unit was calculated as the cluster cell weight total minus the estimated cell proportion multiplied by the cluster total weight; stratum-level variance contributions were then summed across analytic strata and divided by the squared total weight. As a prespecified analytic handling rule, singleton strata were assigned the average variance contribution from nonsingleton strata for the same cell when needed. CIs were calculated on the percentage scale using the 95% normal critical value and truncated to the 0% to 100% range. Cell-level uncertainty estimates were not generated for very high-dimensional comparisons when the number of cell-by-stratum combinations exceeded the prespecified computational threshold of 8,000,000. Sparse cells were interpreted as diagnostic signals rather than definitive population truths.

To quantify sampling uncertainty in scalar distributional measures, we additionally used a 500-replicate stratified cluster bootstrap for the TVD and Jensen-Shannon divergence estimates reported in Tables 2 and 3 [22]. CHS analytic PSUs were resampled with replacement within analytic strata defined by survey year and kstrata; persona distributions were treated as fixed. The cell universe was fixed to the observed union of CHS and persona cells. We report the observed plug-in estimate, bootstrap mean, and 2.5th to 97.5th percentile interval separately. In sparse high-dimensional universes, resampled CHS data occupy fewer cells than the full sample, which can produce positive plug-in TVD bias. We did not replace the observed estimate with a bootstrap bias–corrected estimate because such corrections can overcorrect under sparse multinomial support; the observed full-sample TVD was retained as the descriptive point estimate. For the CHS split-half reference, PSUs were randomly divided into 2 approximately equal groups within each design stratum. Weighted distributions were estimated in each half and their TVD was calculated over 100 random splits; the median values and 2.5th to 97.5th percentiles were reported. Because each half used half the survey sample, this reference is a conservative upper bound for full-sample sampling noise.

We conducted 8 analyses. First, we compared marginal distributions for core harmonized variables. Second, we compared low-dimensional joint distributions, including age by sex, age by living-alone status, age by sex by living-alone status, and age by sex by education. Third, we evaluated core national, provincial, and municipal high-dimensional profiles excluding economic activity and occupation. Fourth, we compared policy-relevant vulnerable profiles that did not depend on economic activity or occupation recoding. Fifth, we performed sensitivity analyses for economic activity, occupation, occupation recoding, living-alone recoding, and economic activity–based vulnerable profiles. Sixth, we performed sparse-cell sensitivity analyses by restricting comparisons to cells with CHS sample sizes of at least 30, 50, or 100. Seventh, we compared persona regional profiles with KOSIS public statistics benchmarks using Pearson and Spearman correlations, including exact-match and conditional living-alone sensitivity analyses [18]. Eighth, we conducted additional uncertainty and robustness analyses using CHS temporal-stability diagnostics; within-stratum PSU split-half TVD references; product-of-CHS-margins independence baselines; and national, provincial, and municipal composition-standardized unmet health care analyses. These analyses were used as descriptive references for benchmark variability and fitness for use, not as formal equivalence tests.

We interpreted these analyses in terms of practical surveillance readiness using evidence triangulation rather than fixed TVD cutoffs. A persona layer was interpreted as more suitable for descriptive planning when discrepancies were small in absolute percentage-point terms: TVD was low relative to lower-dimensional benchmark comparisons, persona-versus-pooled-CHS TVD did not exceed relevant CHS variability references, and KOSIS correlations supported regional ranking consistency. Policy-relevant vulnerable profiles were considered usable for scenario generation or dashboard prototyping only when benchmark differences and uncertainty diagnostics were displayed alongside persona estimates. A profile was interpreted as requiring calibration before inferential surveillance use when discrepancies increased with dimensionality, exceeded CHS variability references, or depended on variables with substantial harmonization uncertainty. Independence baselines were interpreted in relation to the strength of the underlying CHS dependence. Profiles whose discrepancy narrowed toward the CHS split-half reference after sparse-cell restriction were interpreted as weaker evidence of residual discrepancy than profiles that remained far above it.

All data processing and statistical analyses were conducted in Python (Python Software Foundation) using scripted workflows. CHS weighted percentages were calculated from person weights within each comparison-specific denominator, whereas persona percentages were calculated from synthetic row counts [19]. For aligned distributions, TVD and Jensen-Shannon divergence were calculated from normalized proportions, bootstrap intervals resampled CHS analytic PSUs within strata, sparse-cell restrictions used unweighted CHS cell counts, and KOSIS comparisons used Pearson and Spearman correlations across matched municipal-level units [18]. The analysis scripts are publicly archived on Zenodo [23].

The Institutional Review Board of Sunchon National University (1040173-202606-HR-032-02) approved this study for an exemption from review. This study used deidentified CHS benchmark estimates obtained through official data-use procedures, public KOSIS aggregate indicators, and synthetic persona rows [15-18]. The analysis did not involve participant recruitment, intervention, direct contact with individuals, or identifiable private information. The synthetic persona records do not represent observed individuals, although synthetic data still require explicit evaluation of utility, privacy, and intended-use fitness before downstream use [8,12,13]. Informed consent was not applicable to this secondary analysis of deidentified, aggregate, and synthetic data.

The persona dataset, CHS, and KOSIS played distinct benchmark roles (Table 4). Personas were evaluated as synthetic individual rows. CHS provided the main weighted survey benchmark. KOSIS provided seed-source–informed public statistics benchmarks for regional aging-related vulnerability.

Basic demographic margins showed close agreement between persona rows and CHS survey benchmarks (Table 2). TVD was <0.001 (bootstrap percentile interval 0.000-0.001) for sex, 0.006 (bootstrap percentile interval 0.005-0.006) for province, 0.008 (0.007-0.009) for living-alone status, and 0.013 (bootstrap percentile interval 0.012-0.014) for age group. Larger differences were observed for marital status (TVD 0.039, bootstrap percentile interval 0.038-0.040) and education (TVD 0.059, bootstrap percentile interval 0.058-0.061).

Economic activity and occupation were excluded from the core marginal table because both required interpretation of generated occupation text; their results are reported as sensitivity analyses in Table S5 in Multimedia Appendix 1.

Joint-distribution discrepancies increased with profile dimensionality (Table 3). TVD was 0.015 (bootstrap percentile interval 0.014-0.016) for age by sex, 0.020 (bootstrap percentile interval 0.019-0.022) for age by living-alone status, 0.023 (bootstrap percentile interval 0.022-0.024) for age by sex by living-alone status, and 0.063 (bootstrap percentile interval 0.062-0.064) for age by sex by education.

The largest core discrepancies appeared in higher-dimensional profiles. The provincial 5-way profile had observed TVD 0.080, a bootstrap mean of 0.080 (0.079-0.081), and a CHS split-half median of 0.025 (2.5th-97.5th percentile 0.023-0.028). The national 5-way profile had TVD 0.106 (bootstrap mean 0.106; 0.105-0.107), substantially above its split-half median of 0.009 (2.5th-97.5th percentile 0.008-0.011). The municipal 5-way profile had observed TVD 0.117, a bootstrap mean of 0.122 (0.120-0.123), and a split-half median of 0.082 (2.5th-97.5th percentile 0.080-0.084). The municipal bootstrap distribution lay above the full-sample estimate because resampling increased sparsity in the 14,550-cell universe. Relative to independence baselines, the gain was small for age by sex (0.015 vs 0.019), where CHS dependence was weak, but large for age by sex by education (0.063 vs 0.258), where dependence was stronger. Thus, the national discrepancy was clear, whereas the municipal comparison required interpretation alongside its sparse-cell and split-half diagnostics.

In Figure 1, the multiplication sign “x” denotes cross-classified joint distributions. The municipal-level profile comprises municipal area, age, sex, living-alone status, and education. The provincial profile comprises province, age, sex, living-alone status, and education. The national high-dimensional profile comprises age, sex, marital status, living-alone status, and education. Economic activity and occupation are excluded from this core figure and evaluated in sensitivity analyses. Bars show persona-versus-CHS TVD; diamonds and horizontal lines show the CHS split-half median and 2.5th to 97.5th percentile range across 100 within-stratum primary sampling unit splits. “K” denotes the number of union cells. Municipal-level areas correspond to Korean si/gun/gu units, including cities, counties, districts, and wards.

Policy-relevant vulnerable profiles showed moderate national-level agreement (Table 5). Adults aged 65 years or older who were widowed women represented 5.10% of the CHS weighted benchmark and 6.02% of persona rows, a difference of 0.93 percentage points. Adults aged 65 years or older with low education and living alone represented 3.59% of CHS and 3.30% of persona rows, a difference of –0.29 percentage points. Differences were –0.13 percentage points for women aged 75 years or older living alone and 0.01 percentage points for adults aged 19 to 29 years living alone. Economic activity–based vulnerable profiles were not treated as core findings and are reported separately in Table S6 in Multimedia Appendix 1.

In Table 5, profiles are cross-classified sociodemographic groups selected for public health surveillance relevance, including older adults, women, living-alone status, and low education. Economic activity–based vulnerable profiles are reported in Table S6 in Multimedia Appendix 1. Differences were calculated from unrounded percentages.

Economic activity and occupation harmonization were evaluated in Table S5 in Multimedia Appendix 1, and vulnerable profiles depending on economic activity were reported in Table S6 in Multimedia Appendix 1. Economic activity had marginal TVD 0.017 (95% bootstrap percentile interval 0.016-0.018), and the profile adding economic activity had TVD 0.137 (95% bootstrap percentile interval 0.136-0.138). Occupation had marginal TVD 0.271 (95% bootstrap percentile interval 0.271-0.272), which decreased to 0.149 under the expanded keyword mapping; the occupation-containing high-dimensional profile had TVD 0.368 (95% bootstrap percentile interval 0.367-0.369). Economic activity and occupation were therefore retained as sensitivity evidence rather than core benchmark domains. Living-alone recoding sensitivity was evaluated separately in Table S9 in Multimedia Appendix 1. The living-alone results were robust to alternative treatment of the single ambiguous family-type label: marginal TVD ranged from 0.008 to 0.014, national core TVD from 0.105 to 0.106, and municipal core TVD was 0.117 under all 3 rules.

As a regional downstream compositional sensitivity analysis, we applied national CHS profile–specific medical-need and unmet-need probabilities to CHS and persona profile distributions (Table S7 in Multimedia Appendix 1). Using a primary minimum cell count of 30, the mean absolute difference between CHS- and persona-composition standardized rates was 0.076 percentage points across 17 provinces and 0.109 percentage points across 252 municipalities. Municipal standardized rates had Pearson r=0.567 and Spearman ρ=0.683 across the 2 compositions. However, composition-standardized rates did not reproduce observed municipal CHS unmet-need rates (persona composition: mean absolute difference 1.635 percentage points; Pearson r=0.004). Because CHS- and persona-composition rates differed by only 0.109 percentage points on average, CHS-composition rates also departed from observed municipal rates by at least 1.5 percentage points on average. Therefore, the lack of agreement with observed local rates reflects the limits of standardization with national profile-specific probabilities rather than persona composition alone. This analysis evaluated compositional sensitivity, not health outcome prediction.

Sparse-cell sensitivity differed materially between national and municipal profiles (Table S8 in Multimedia Appendix 1). The national core TVD remained 0.105 after restriction to CHS cells with n≥100, far above its split-half median of 0.009 (2.5th-97.5th percentile 0.008-0.011). In contrast, municipal TVD decreased from 0.117 across all cells to 0.105 at n≥30 and 0.091 at n≥100, with 73.0% of CHS mass and 69.2% of persona mass retained at n≥100. The all-cell municipal TVD exceeded its split-half median of 0.082 (2.5th-97.5th percentile 0.080-0.084), which is itself a conservative upper bound for full-sample sampling noise. Because the split-half reference was estimated across all cells rather than within each restricted cell universe, the restricted TVD values are not directly comparable with it. The municipal discrepancy was therefore smaller and less clearly separated from CHS sampling noise than the national discrepancy, but it was not explained by sparse cells alone. For the occupation-containing sensitivity profile, TVD decreased from 0.368 to 0.232 at n≥100; this restriction retained 96.9% of CHS mass but only 70.2% of persona mass, indicating that 29.8% of persona rows occupied combinations that were rare or absent in CHS.

The temporal-stability analysis compared persona-versus-pooled-CHS TVD with year-specific CHS-versus-pooled-CHS TVD for the same benchmark layers. For age group, the persona TVD was 0.013, below the maximum CHS year-versus-pooled TVD of 0.020. For living-alone status, the persona TVD was 0.008, also below the maximum CHS yearly TVD of 0.012. These results support the interpretation that the age and living-alone margins were close relative to ordinary CHS year-to-year variation.

In contrast, discrepancies for education and higher-dimensional profiles exceeded the CHS temporal-stability reference (Table S4 in Multimedia Appendix 1). Education had persona-versus-pooled CHS TVD 0.059 compared with maximum CHS year-versus-pooled TVD 0.016. Age by sex by education had persona TVD 0.063 compared with maximum yearly TVD 0.037. The national core profile had persona TVD 0.106 compared with maximum yearly TVD 0.045. The municipal profile had persona TVD 0.117 compared with maximum yearly TVD 0.091; this smaller difference should be interpreted with the municipal split-half and sparse-cell results.

CHS 95% CI diagnostics were intentionally interpreted cautiously because the pooled CHS benchmark included more than 1 million observations and therefore produced narrow CIs. For several margins and vulnerable profiles, persona percentages fell outside the pooled CHS 95% CIs despite absolute differences of less than 1 percentage point. For example, the maximum age-group difference was 0.99 percentage points. These diagnostics show that statistical distinguishability from a large survey benchmark does not by itself imply public health invalidity; substantive interpretation should consider absolute percentage-point differences, TVD, temporal CHS stability, sparse cells, and intended surveillance use.

KOSIS public statistics comparisons showed strong regional ranking consistency for age structure but weaker consistency for the stricter conditional living-alone comparison (Figure 2 and Table S3 in Multimedia Appendix 1). Across 252 regions, the correlation between KOSIS aged-population share and persona aged-adult-row share was 0.985 by Pearson correlation and 0.980 by Spearman correlation. The composite proxy comparing KOSIS older-adult living-alone household share with persona aged-65-plus living-alone adult-row share was Pearson correlation of 0.926 and Spearman correlation of 0.933. However, the stricter conditional comparison of persona living-alone share among rows aged 65 years or older with the KOSIS older-adult living-alone household indicator was weaker (Pearson r=0.547; Spearman ρ=0.576). Exact municipal-match sensitivity produced similar conclusions: aged-population ranking remained strong (Pearson r=0.989; Spearman ρ=0.991), whereas conditional older-adult living-alone consistency remained modest (Pearson r=0.530; Spearman ρ=0.550).

Because KOSIS indicators used total population or household denominators while personas represented adult synthetic rows, absolute-level differences were not interpreted as direct validation. The KOSIS results indicate that the persona data captured regional age-structure rankings well, but they provide weaker support for living-arrangement structure among older adults after conditioning on age.

In Figure 2, points represent matched municipal-level areas, corresponding to Korean si/gun/gu units; fitted lines show linear trends. The left panel compares the KOSIS percentage of residents aged 65 years or older with the corresponding persona share among adult rows. The right panel compares the KOSIS older-adult one-person-household percentage with the corresponding persona share of rows that were both aged 65 years or older and living alone among all adult rows. Because these KOSIS indicators differ from persona row denominators and because KOSIS was among the public statistical sources used to construct the persona data, this figure is interpreted as a seed-source–informed public statistics benchmark check rather than independent validation; exact-match and conditional living-alone sensitivity results are reported in Table S3 in Multimedia Appendix 1.

The validation layers support a tiered interpretation for surveillance use. The persona data appear suitable for descriptive surveillance planning when the use case depends on basic demographic margins and regional age-structure rankings. Policy-relevant vulnerable profiles may be useful for scenario generation and dashboard prototyping if benchmark differences and uncertainty diagnostics are reported alongside persona estimates. The national high-dimensional discrepancy indicates a need for calibration. At the municipal level, the discrepancy exceeded the all-cell CHS split-half reference but narrowed after sparse-cell restriction, so its separation from CHS sampling noise was less clear. Detailed local profiles should therefore be used as calibration diagnostics rather than direct evidence for resource allocation, subgroup inference, or local health outcomes.