Late HIV Presentation Remains High in Korea: 28-Year National Surveillance Trends and the Limits of Inference under Incomplete CD4 Capture, 1998–2025

Abstract

Background

Whether late human immunodeficiency virus (HIV) presentation in Korea has improved is difficult to judge because a CD4 count at diagnosis was captured for only a minority of notified cases until 2025. We examined 28-year trends in late presentation (CD4 <350 cells/µL) and advanced HIV disease (AHD; CD4 <200 cells/µL) among cases with a captured CD4 result, quantified how incomplete and time-varying capture limits inference about population-level trends, and characterized the burden in 2025, the year of highest CD4 capture (80.4%).

Materials and Methods

We analyzed national HIV/acquired immunodeficiency syndrome (AIDS) notification data from the Korea Disease Control and Prevention Agency (1998–2025). Annual proportions among captured cases were modeled with overdispersion-adjusted logistic regression and adjusted for mean age and for the annual uncaptured fraction; change-points were identified by joinpoint regression. Missing CD4 values were treated as incomplete capture; distribution-free Manski bounds were computed for every year, with a missing-not-at-random tipping-point analysis and a sensitivity analysis for recent infection.

Results

Among captured cases, late presentation showed no significant change (odds ratio [OR] per year, 1.008; 95% confidence interval [CI], 0.995–1.021; P=0.259) and the AHD proportion increased (OR per year, 1.027; 95% CI, 1.012–1.042; P<0.001), independent of age. These captured-only trends cannot be read as population-level trends: the uncaptured fraction was 0.40–0.83 before 2025, so the Manski bounds do not identify the direction of any trend, and adjusting for the uncaptured fraction (OR per 0.10 increase, 1.059 and 1.062) did not attenuate the year effect. The decrease in the AHD proportion after 2022 coincided with an abrupt increase in capture (uncaptured fraction, 0.83 to 0.20). In 2025, late presentation was 0.596 among captured cases (Manski bounds, 0.480–0.675) and AHD was 0.379 (0.305–0.501); both intervals contain the 2013–2015 Korea HIV/AIDS Cohort estimates (0.564 and 0.356). Reclassifying up to half of recent infections (27.3% of diagnoses) as not late left the 2025 late-presentation proportion at ≥0.460.

Conclusion

In 2025, approximately half to two-thirds of new HIV diagnoses in Korea were late and one-third to one-half had advanced disease, above contemporary European levels and far above the World Health Organization European target. Whether this burden has improved cannot be determined from surveillance data, because captured-only trends track the completeness of CD4 capture. Late-presentation indicators should be reported together with capture completeness, and cohort data with complete CD4 ascertainment are needed to establish the trend.

Graphical Abstract

Introduction

Antiretroviral therapy has transformed human immunodeficiency virus (HIV) into a manageable chronic condition, and the global age-standardized burden of HIV/acquired immunodeficiency syndrome (AIDS), measured in disability-adjusted life-years, decreased by nearly half between 2010 and 2021 [1]. These gains, however, are largely conditional on timely diagnosis and treatment. A European consensus defines late presentation as a CD4 count below 350 cells/µL at diagnosis, or an AIDS-defining event regardless of CD4 count, and advanced HIV disease (AHD) as a CD4 count below 200 cells/µL [2]; the two categories are clinically distinct, and patients presenting late—especially with AHD—carry a higher burden of opportunistic infection, respond less well to therapy, and have a several-fold higher risk of clinical progression and death [3]. Late diagnosis also has consequences beyond the individual, because onward transmission from people unaware of their infection is estimated to be several-fold higher than from those already diagnosed [4]. Across Asia, an estimated one-third to two-thirds of people are still diagnosed late [4,5].

In Korea, newly notified HIV cases increased for most of the past four decades before a modest recent decrease, yet the immunological stage at which people enter care has attracted less attention. In the Korea HIV/AIDS Cohort Study, the proportion diagnosed with a CD4 count ≤350 cells/µL increased to 75.9% in 2010–2012 and then decreased for the first time, to 56.4%, in 2013–2015 [6]. However, national estimates of late presentation have been constrained by a structural feature of the surveillance system: a CD4 result at diagnosis is available for only a subset of notified cases each year. Because CD4 and HIV-RNA are measured routinely in treating facilities, this gap reflects incomplete capture of existing laboratory results into case-based reporting rather than untested patients—a distinction that has not been formally addressed in prior Korean analyses and that limits any trend estimated from captured cases alone.

This reporting-completeness problem is not unique to Korea; back-calculation and undiagnosed-fraction analyses elsewhere have shown that the immunological profile of cases entering surveillance shapes conclusions about the size and stage of the epidemic [7,8]. When the completeness of CD4 capture itself changes over time, apparent trends in late presentation may reflect reporting artifacts as much as genuine clinical change. In Korea, the fraction of notified cases lacking a captured CD4 result changed abruptly in the most recent reporting year, creating a natural opportunity to examine how sensitive national late-presentation estimates are to reporting completeness.

A separate analysis of the same national series by our group estimates the size of the undiagnosed population by back-calculation (manuscript under review) and does not examine the trend in late presentation, which is the subject of the present study. We therefore analyzed 28 years of national HIV notification data with four objectives: (1) to describe trends in late presentation and AHD among cases with a captured CD4 result; (2) to test whether any such trend is explained by the changing age structure of new diagnoses; (3) to quantify how far incomplete and time-varying capture of CD4 results constrains inference about population-level trends, using year-by-year distribution-free bounds, capture-adjusted models, change-point analysis, and sensitivity analyses for the missing-data mechanism and for recent infection; and (4) to characterize the late-presentation burden in 2025, the first year in which a CD4 result was captured for four in five notified cases (80.4%), against prior Korean cohort estimates and international benchmarks.

Materials and Methods

1. Data source

We analyzed the annual national HIV/AIDS notification reports published by the Korea Disease Control and Prevention Agency (KDCA) for diagnoses reported between 1985 and 2025. From the published aggregate tables we abstracted the CD4+ T-cell count at diagnosis, sex, and age group. Because fewer than one in five diagnoses before the late 1990s had a recorded CD4 count and annual numbers were small, all CD4-based trend analyses were restricted to 1998–2025. The proportion of new diagnoses classified as recent infection by the KDCA recency-assay program (27.3% among Korean nationals in 2016–2023) was obtained from the published KDCA study [9]. The annual values underlying every analysis are provided in Supplementary Table 1.

2. Ethics statement

This study used aggregate, de-identified national surveillance tabulations that are publicly reported and contain no individually identifiable information. The study protocol was reviewed and approved by the Institutional Review Board of Kyungpook National University Hospital (IRB No. 2026-05-007), and the requirement for informed consent was waived owing to the retrospective use of published aggregate data.

3. Outcome definitions and surveillance capture

Late presentation was defined as a CD4 count below 350 cells/µL at diagnosis and AHD as a CD4 count below 200 cells/µL, following CD4-based approximations of the consensus European definitions [2,10]. The updated European definition also counts an AIDS-defining event regardless of CD4 count and excludes persons with evidence of recent infection [10]; neither element is recorded in the published aggregate tables, so neither was used, and the consequences are examined in a sensitivity analysis and in the Discussion. The two categories were kept strictly distinct throughout, AHD being the more severe subset.

A CD4 count was recorded for only a subset of notified cases each year. CD4 and HIV-RNA are measured routinely in treating facilities, so a missing value in the surveillance record reflects incomplete capture of an existing laboratory result into case-based reporting rather than an untested patient. We therefore distinguished two quantities: the captured-only proportion (events divided by cases with a CD4 result available in surveillance) and the lower-bound proportion (events divided by all notified cases), which treats every case without a captured CD4 result as a non-event and is therefore identical to the lower Manski bound described below. The fraction of notified cases without a captured CD4 result—the uncaptured fraction—was tracked each year as an index of surveillance reporting completeness.

4. Statistical analysis

We first described how late presentation and AHD changed over time among cases with a captured CD4 result. For each outcome we fitted a logistic regression with calendar year as a continuous predictor and summarized the trend as an odds ratio per year (OR/yr): the factor by which the odds of the outcome change with each additional year. Because the year-to-year proportions varied more than a binomial model assumes (overdispersion), standard errors were inflated by a quasi-binomial correction so that ordinary fluctuation would not be mistaken for a real trend. To check that the conclusion did not depend on one modeling choice, we repeated each trend with two alternative models that make different assumptions about how adjacent years relate to one another (generalized estimating equations and first-order autoregressive generalized least squares); the direction and magnitude of the estimates were broadly consistent across models, although statistical significance was not (Table 1). Because these models use only cases with a captured CD4 result, their estimates describe the captured series and are not, by themselves, estimates of the population-level trend among all notified cases.

Table 1. Overdispersion-adjusted temporal trends in late presentation and advanced HIV disease among cases with a captured CD4 result, Korea, 1998–2025.

Indicator		OR per year	95% CI	P-value	φ
Late presentation (<350 cells/µL)		1.008	0.995–1.021	0.259	3.26
Advanced HIV disease (<200 cells/µL)		1.027	1.012–1.042	<0.001	4.45
Lower-bound series, AHD (all notified cases as denominator)		0.999	0.975–1.023	0.929	20.11
Capture-adjusted models (year + uncaptured fraction)
	Late presentation, year effect	1.008	0.996–1.020	0.193	2.89
	Late presentation, uncaptured fraction (per 0.10)	1.059	1.005–1.115	0.031	2.89
	AHD, year effect	1.028	1.014–1.042	<0.001	3.96
	AHD, uncaptured fraction (per 0.10)	1.062	1.002–1.125	0.042	3.96
AHD trend, sensitivity analyses
	Quasi-binomial	1.024	1.006–1.041	0.007	—
	GEE, exchangeable (robust SE)	1.034	1.012–1.056	0.002	—
	AR1 GLS (ρ=0.40)	1.022	0.997–1.049	0.089	—

ORs per calendar year from logistic regression with standard errors inflated by √φ (quasi-binomial). Unless stated otherwise, models use cases with a captured CD4 result as the denominator and describe the captured series only. The lower-bound series divides events by all notified cases, treating every case without a captured result as a non-event; it equals the lower Manski bound and rises by construction as capture improves. The capture-adjusted models add the annual uncaptured fraction (per 0.10 increase) to the year model, fitted to 1998–2025; both coefficients from each model are shown. The sensitivity analyses for the AHD year trend are each fitted to 2003–2025 (the period with mean-age data); the primary AHD estimate above is fitted to the full 1998–2025 period, which accounts for the small difference in the point estimate (OR/yr 1.027 vs. 1.024). The direction and magnitude of the three sensitivity estimates were consistent; the AR1 GLS estimate was not statistically significant.

P-values are two-sided.

HIV, human immunodeficiency virus; OR, odds ratio; CI, confidence interval; AHD, advanced HIV disease; GEE, generalized estimating equations; SE, standard error; AR1 GLS, first-order autoregressive generalized least squares; φ, Pearson dispersion statistic.

We next examined whether an increasing captured-only AHD trend could reflect an aging population of newly notified cases rather than a true increase. We added the annual mean age of newly notified cases to the model alongside calendar year: if aging were the explanation, the year effect should weaken once age is accounted for. Because CD4-by-age data are not published at the individual level, this is a group-level (ecological) adjustment, a limitation we revisit in the Discussion.

5. Accounting for incomplete capture

Because CD4 was captured for only some cases, we used four complementary analyses to establish what the data can and cannot show about the population-level trend. First, a bounding analysis (Manski bounds) makes no assumption about the uncaptured cases: for every year, we counted each case without a CD4 result first as not late and then as late, which brackets the true proportion among all notified cases between a lower and an upper bound. The width of this interval equals the uncaptured fraction, so the bounds are informative in proportion to capture completeness; we report them for every year rather than for 2025 alone. Second, a capture-adjusted model added the annual uncaptured fraction to the logistic model alongside calendar year: if uncaptured results belonged disproportionately to patients at an earlier stage, the uncaptured fraction should be associated with the captured-only proportion; whether the year effect changes once it is accounted for indicates how much of the captured-only trend a linear adjustment can explain. Like the age adjustment, this is an ecological adjustment, and its ability to recover the population-level trend is limited by the width of the bounds described above. Third, a tipping-point analysis assessed how different the uncaptured cases would have to be from the captured ones before the captured-only trend lost significance: the odds of AHD among uncaptured cases were shifted by a factor δ relative to captured cases, and the trend was re-estimated as δ was varied from 0.25 to 4.0, with δ=1 indicating that the two groups are alike (missing at random). Fourth, a joinpoint analysis located the year in which the captured-only AHD trend changed slope, with the number of change-points chosen by a permutation procedure and the change-point year bounded by bootstrap resampling; the joinpoint analysis describes the timing of a change in the captured series and does not identify its cause.

6. Sensitivity analysis for recent infection

CD4 counts fall transiently during acute and recent HIV infection, so a proportion of diagnoses with CD4 <350 cells/µL are recent infections rather than late presentations; the updated European definition excludes such cases [10], but the aggregate tables do not permit this. We therefore recalculated the 2025 proportions after removing from the numerator a fraction f of recent infections, taking the recent-infection proportion among newly diagnosed Korean nationals as 0.273 [9] and varying f (the fraction of recent infections with a transiently low CD4 count). For late presentation, f was varied from 0.10 to 0.50, a range that spans published estimates. For AHD, an empirical value is available: the KDCA recent-infection testing algorithm removes avidity-recent cases with a CD4 count <200 cells/µL, and these were 3.5% of avidity-recent cases (72 of 2,075) [9]; f was therefore set at 0.035, 0.05, and 0.10. This calculation assumes that the proportion of recent infections is the same among cases with a captured CD4 result as among all newly diagnosed Korean nationals, and that the 2016–2023 proportion applies to 2025. To examine the first assumption, the range of recent-infection proportions among captured cases that is compatible with the overall proportion and with 80.4% capture was also calculated, by assigning all or none of the uncaptured cases to recent infection. The same subtraction was applied to the Manski bounds.

Analyses were performed using Python (version 3.12, Python Software Foundation, Beaverton, OR, USA; statsmodels, SciPy, NumPy), and two-sided P<0.05 was considered statistically significant.

Results

1. Trends among cases with a captured CD4 result

Among cases with a captured CD4 result, the proportion presenting late had no significant temporal trend after overdispersion adjustment (OR, 1.008; 95% confidence interval [CI], 0.995–1.021; φ=3.26, indicating substantial excess year-to-year variation; P=0.259). The captured-only AHD proportion increased over the long term (OR, 1.027; 95% CI, 1.012–1.042; φ=4.45; P<0.001) (Fig. 1A, Table 1). These estimates describe cases with a captured CD4 result, which were a minority of notified cases in every year before 2025; whether they reflect the population-level trend is examined below.

Figure 1. Late and advanced HIV presentation and surveillance capture of CD4 results, Korea, 1998–2025. (A) Proportion of cases with a captured CD4 result presenting late (CD4 <350 cells/µL) and with AHD (CD4 <200 cells/µL), shown as lines; the shaded bands are the year-by-year Manski bounds among all notified cases (lower bound: every uncaptured case counted as a non-event; upper bound: every uncaptured case counted as an event), whose width equals the uncaptured fraction. (B) Annual proportion of notified cases whose CD4 result was not captured into surveillance; CD4 is measured routinely in treating facilities, so this reflects reporting completeness rather than testing.

HIV, human immunodeficiency virus; AHD, advanced HIV disease.

2. The captured-only AHD trend is not explained by cohort aging

The mean age of newly notified cases remained essentially unchanged (37.7 years in 2003; 38.0 in 2025). Adding mean age to the model did not attenuate the year effect (year OR, 1.029; 95% CI, 1.006–1.052; P=0.013; mean age OR, 1.036; 95% CI, 0.936–1.146; P=0.490); had aging driven the captured-only AHD trend, the year coefficient would have moved toward the null upon adjustment. The within-year age distribution likewise showed no shift toward older ages across the study period (Fig. 2, Table 2).

Figure 2. Age distribution of newly notified HIV diagnoses by year, 2003–2025. Heat map of the within-year age distribution (each column sums to 100%); darker cells denote a higher percentage.

HIV, human immunodeficiency virus.

Table 2. Multivariable logistic model for advanced HIV disease at diagnosis among cases with a captured CD4 result, adjusting for the age structure of new diagnoses, 2003–2025.

Term	OR	95% CI	P-value
Calendar year (per year)	1.029	1.006–1.052	0.013
Mean age of new diagnoses (per year)	1.036	0.936–1.146	0.490

Standard errors overdispersion-adjusted. The year effect is essentially unchanged from the univariable estimate, indicating that the captured-only trend in advanced HIV disease is not explained by the age structure of new diagnoses.

HIV, human immunodeficiency virus; OR, odds ratio; CI, confidence interval.

3. Surveillance capture and the 2022 change-point

The uncaptured fraction decreased from 0.70 in 1998 to approximately 0.40 around 2010, increased to 0.76–0.83 in 2018–2023, and then decreased abruptly to 0.20 in 2025 (Fig. 1B). We are not aware of a published account of the 2025 change; we presume that it reflects public health centers obtaining clinic CD4 and HIV-RNA results more actively, for example through case managers and patients, but this presumption is untested and no policy document describing it has been identified. Permutation testing selected a single joinpoint in 2022 (0-vs-1 P=0.044; 1-vs-2 P=0.070). The captured-only AHD log-odds increased at an annual percent change of +4.24%/yr during 2003–2022 and reversed to −15.38%/yr during 2022–2025 (average annual percent change +1.32%/yr). The change-point year had a wide bootstrap interval (95% CI, 2008–2022), indicating that its timing is uncertain; the joinpoint analysis locates a change in slope in the captured-only series but does not, by itself, identify whether that change reflects surveillance capture or clinical change (Supplementary Fig. 1A).

4. Year-by-year bounds and the capture-adjusted trend

Because the width of the Manski interval equals the uncaptured fraction, the bounds were at least 0.40 wide in every year before 2025 and 0.76–0.83 wide in 2018–2023 (Fig. 1A). In no year before 2025 did the bounds identify the direction of change relative to any other year; the bounds therefore constrain the level of late presentation only in 2025 and do not establish a population-level trend in either direction. The lower-bound series (events divided by all notified cases) increased from 0.101 in 2022 to 0.305 in 2025 as capture improved, which is expected by construction because every newly captured event moves from the uncaptured to the captured category; its long-term year trend was flat (OR, 0.999; 95% CI, 0.975–1.023; P=0.929) and showed no 2022 change-point (Supplementary Fig. 2). We do not interpret this series as an estimate of the population-level trend.

In the capture-adjusted models, the uncaptured fraction was independently associated with the captured-only proportions: OR per 0.10 increase 1.059 (95% CI, 1.005–1.115; P=0.031) for late presentation and 1.062 (95% CI, 1.002–1.125; P=0.042) for AHD, indicating that the captured-only estimates depend on the completeness of capture. The year effect, however, did not attenuate (late presentation OR, per year 1.008, 95% CI, 0.996–1.020; AHD 1.028, 95% CI, 1.014–1.042) (Table 1); the same held when the uncaptured fraction was entered on the logit scale (AHD year OR, 1.028) and was only slightly attenuated when the fit was restricted to 2003–2025 (AHD year OR, 1.022; 95% CI, 1.006–1.038; late presentation 1.003, 95% CI, 0.989–1.016). The captured-only AHD proportion and the uncaptured fraction were weakly correlated over the full period (r=0.24) but strongly correlated from 2011 onward (r=0.68; late presentation r=0.73), the period in which the uncaptured fraction increased from 0.40 to 0.83 and then decreased to 0.20. A linear ecological adjustment therefore does not remove the captured-only rise, and the population-level trend remains unidentified within the bounds.

5. 2025: the year of highest CD4 capture

With CD4 results captured for 80.4% of notified cases in 2025, the captured-only estimates (late 0.596; AHD, 0.379) were lower than the 2018–2024 captured-only averages (late 0.725; AHD, 0.519), suggesting the earlier figures were inflated because results were more completely captured for patients presenting with advanced disease (Fig. 3). Consistent with this, the proportion of newly captured cases with a high CD4 count (≥500 cells/µL) increased markedly in 2025 (from 0.151 of captured cases in 2024 to 0.196 in 2025), as expected if previously uncaptured results belonged disproportionately to earlier-stage patients. When the influx was quantified directly, the approximately 355 cases captured in 2025 beyond the 2024 count had an estimated AHD proportion of about 0.36, below the 0.42 of cases already captured in 2024—consistent with the newly captured results belonging disproportionately to earlier-stage patients and diluting the captured-only estimate (Supplementary Fig. 2).

Figure 3. Smoothed trends in late and advanced HIV presentation among cases with a captured CD4 result, Korea, 1998–2025. Annual proportions among cases with a captured CD4 result are shown for late presentation (CD4 <350 cells/µL) and AHD-CD4 <200 cells/µL). Solid lines are natural cubic spline logistic fits and shaded bands the 95% CI of the fit; open circles are the observed annual proportions. With CD4 results captured for 80.4% of notified cases in 2025, the captured-only proportions (late 0.596; AHD 0.379) were lower than the 2018–2024 captured-only averages (late 0.725; AHD 0.519), a difference of 13 and 14 percentage points, respectively; the 2025 Manski bounds among all notified cases were 0.480–0.675 (late) and 0.305–0.501 (AHD).

HIV, human immunodeficiency virus; AHD, advanced HIV disease; CI, confidence interval.

Manski bounds for 2025 remained clinically substantial (late 0.480–0.675; AHD 0.305–0.501), so no assumption about the uncaptured 20% yields a low late-presentation burden (Fig. 1A). Thus, even under the most favorable assumption, nearly half of 2025 diagnoses were late and approximately one-third had advanced disease. Both 2025 intervals contain the estimates from the Korea HIV/AIDS Cohort Study for 2013–2015 (CD4 ≤350 cells/µL, 57 of 101, 0.564, 95% CI, 0.467–0.657; CD4 ≤200 cells/µL, 36 of 101, 0.356, 95% CI, 0.270–0.454) [6], so the 2025 surveillance data neither demonstrate an improvement nor a worsening relative to that cohort period.

6. Sensitivity analysis for recent infection

Removing from the numerator a fraction of recent infections with transiently low CD4 counts reduced the 2025 late-presentation proportion from 0.596 to 0.569, 0.528, and 0.460 for f=0.10, 0.25, and 0.50, respectively, with corresponding Manski bounds of 0.453–0.648, 0.412–0.607, and 0.344–0.539; the AHD proportion decreased from 0.379 to 0.369, 0.365, and 0.352 for f=0.035, 0.05, and 0.10, with bounds of 0.295–0.491, 0.291–0.487, and 0.278–0.474 (Table 3). Under every assumption examined, at least 34% of 2025 diagnoses were late by the CD4 criterion and at least 28% had advanced disease. Because CD4 results were captured for 80.4% of cases in 2025, the proportion of recent infections among captured cases is arithmetically constrained to 0.096–0.339 whatever the proportion among uncaptured cases; across this range the adjusted late-presentation proportion at f=0.50 lies between 0.427 and 0.548.

Table 3. Sensitivity of the 2025 late-presentation and advanced HIV disease proportions to the misclassification of recent infections with transiently low CD4 counts.

Outcome and assumed fraction of recent infections with a low CD4 count (f)		Captured-only proportion	Manski bounds (all notified cases)
Late presentation (CD4 <350 cells/µL)
	No adjustment (f=0)	0.596	0.480–0.675
	f=0.10	0.569	0.453–0.648
	f=0.25	0.528	0.412–0.607
	f=0.50	0.460	0.344–0.539
Advanced HIV disease (CD4 <200 cells/µL)
	No adjustment (f=0)	0.379	0.305–0.501
	f=0.035 (observed in the KDCA recency program)	0.369	0.295–0.491
	f=0.05	0.365	0.291–0.487
	f=0.10	0.352	0.278–0.474

The recent-infection proportion among new diagnoses was taken as 0.273, the value reported by the KDCA recency-assay program for Korean nationals in 2016–2023 [9], and was assumed to be the same among captured and uncaptured cases and to apply to 2025. For advanced HIV disease, f = 0.035 is the observed proportion of avidity-recent cases with CD4 <200 cells/µL in that program (72 of 2,075). Adjusted proportion = unadjusted proportion − 0.273 × f, i.e., recent infections with a transiently low CD4 count are removed from the numerator but remain in the denominator as diagnoses. The same subtraction is applied to both Manski bounds. f, assumed fraction of recent infections with CD4 <350 cells/µL (late presentation) or <200 cells/µL (advanced HIV disease) at diagnosis.

HIV, human immunodeficiency virus; KDCA, Korea Disease Control and Prevention Agency.

7. Robustness to the missing-data assumption and to model choice

In the missing-not-at-random tipping-point analysis, the captured-only AHD trend remained significant across the entire range of δ from 0.25 to 4.0; the trend lost significance only below δ ≈ 0.24, a scenario in which uncaptured patients had approximately one-quarter the AHD odds of captured patients (Supplementary Fig. 1B). Fitted over 2003–2025 (the years with mean-age data), the sensitivity models agreed in direction and magnitude with the primary 1998–2025 estimate—quasi-binomial OR/yr 1.024 (P=0.007) and generalized estimating equations OR/yr 1.034 (P=0.002) were significant, whereas the more conservative first-order autoregressive generalized least squares model gave OR/yr 1.022 (95% CI, 0.997–1.049; P=0.089), which was not statistically significant (Table 1).

Discussion

In this 28-year analysis of national HIV notification data, the central finding concerns the level of late presentation rather than its trend. In 2025, the first year in which a CD4 result was captured for four in five notified cases, approximately half to two-thirds of new diagnoses were late and one-third to one-half had advanced HIV disease, whatever is assumed about the remaining 20%. Among cases with a captured CD4 result, late presentation showed no change over 28 years and the AHD proportion increased, independent of the age structure of new diagnoses; but these captured-only trends cannot be read as population-level trends, because a CD4 result was captured for a minority of cases in every year before 2025 and the completeness of capture itself varied more than threefold over the period. The apparent decrease in AHD after 2022 coincided with an abrupt improvement in capture, and the increase before 2022 coincided with a deterioration in capture.

The sharp decrease in cases lacking a recorded CD4 count in 2025 is unlikely to reflect increased CD4 testing at the clinical level. CD4 and HIV-RNA are measured routinely in treating facilities; the change more plausibly reflects improved ascertainment of these laboratory results into the national surveillance record. The missing-CD4 fraction therefore represents incomplete reporting of existing results rather than absent testing. We presume, but cannot document, that the 2025 improvement followed public health centers obtaining clinic results more actively through case managers and patients; the KDCA press release and annual report for 2025 describe no such change, so the mechanism remains unestablished.

Two observations suggest that the uncaptured results did not belong to a random subset of patients with respect to disease stage. When capture improved in 2025, the share of captured cases with CD4 ≥500 cells/µL increased markedly and the captured-only late and AHD proportions decreased, and the cases newly captured in 2025 had a lower estimated AHD proportion than those already captured in 2024. If, as this suggests, results were captured more completely for patients presenting with advanced disease, the captured-only proportion would be inflated most in the years in which capture was least complete. Such a mechanism would operate in both directions: it may account for part of the decrease after 2022, when capture improved, and it could in principle contribute to the increase between 2010 and 2022, when the uncaptured fraction increased from approximately 0.40 to 0.83. The capture-adjusted model supports the first part of this reading but not the second: the uncaptured fraction was independently associated with the captured-only proportions, but adjusting for it left the year effect unchanged, so the pre-2022 rise in the captured series is not explained by a linear adjustment for capture. We therefore neither attribute that rise to capture nor read it as a population-level increase; the year-by-year bounds leave the population-level trend unidentified, and the captured-only rise remains an observation about the captured series. The wide bootstrap interval of the 2022 joinpoint indicates that the timing of the change is uncertain, but it carries no information about its cause; we therefore do not attribute the reversal to surveillance capture on the basis of the joinpoint analysis, and we draw that attribution only from the internal checks above and from the coincidence of the reversal with the presumed change in reporting practice described above.

The bounds make the limits of inference explicit. Distribution-free bounds hold regardless of why results were uncaptured [11], but their width equals the uncaptured fraction, and before 2025 that fraction was never below 0.40 and exceeded 0.75 in 2018–2023. The year-by-year bounds therefore cannot identify a population-level trend in either direction, and they anchor our conclusions only for 2025. The lower-bound series, which treats every uncaptured case as a non-event, necessarily rises as capture improves and is not an estimate of the population-level trend; we report it only to make explicit that the captured-only rise and the post-2022 reversal are features of the captured series. Pre-2025 captured-only proportions were subject to reporting-completeness bias and should not be interpreted as unbiased temporal trends.

These findings can be interpreted in the context of prior work. The Korea HIV/AIDS Cohort Study, in which CD4 at diagnosis is ascertained clinically, reported that the proportion with CD4 ≤350 cells/µL at diagnosis increased from 55% before 2000 to 75.9% in 2010–2012 and then decreased for the first time, to 56.4%, in 2013–2015, with CD4 ≤200 cells/µL following the same course (51.4% to 35.6%) [6]; cohort data on CD4 at diagnosis have not been reported by diagnosis year beyond 2016, and prior Korean and other Asian estimates were generally derived from patients with an available CD4 value without addressing how incomplete and time-varying capture could bias the trend itself [5,6]. The 2013–2015 cohort estimates (0.564 and 0.356; n=101) lie inside the 2025 surveillance bounds, and their own CIs (0.467–0.657 and 0.270–0.454) overlap the bounds almost entirely, so the surveillance data neither confirm nor refute a continued decline after 2015. The cohort comparison is also indicative only: the analyzed cohort patients were 1–15% of national diagnoses in each period and 3.2% in 2013–2015, the cohort cutoffs are inclusive (≤350, ≤200 cells/µL), and the 2013–2015 cohort sample was younger (median age 34 years) than in earlier periods (39–42 years), whereas the age structure of national diagnoses was stable. A further reason for caution is that the proportion of newly diagnosed Korean nationals classified as recent infection increased to 27.3% in 2016–2023, from 20.5% in 2008–2015 [9]; because CD4 counts fall transiently during recent infection, some diagnoses with CD4 <350 cells/µL in recent years are recent infections rather than late presentations, a distinction the updated European definition draws [10] but the aggregate tables cannot. The sensitivity analysis shows that this misclassification could lower the 2025 late-presentation proportion by up to about 14 percentage points under the most generous assumption, but not to a level that would change the conclusion that the burden remains high; for advanced disease the effect is small, because the KDCA recent-infection testing algorithm itself found CD4 <200 cells/µL in only 3.5% of avidity-recent cases [9]. Published cohort data after 2015 are consistent with the 2013–2015 decrease but are not directly comparable and do not themselves show a continuing trend: among cohort enrollees diagnosed in 2015–2016, about 20% had CD4 <200 cells/µL at diagnosis (median 259 cells/µL, interquartile range, 190–363), with no consistent change in the median CD4 count by diagnosis year over 2006–2016 [12]; and among all enrollees in 2006–2023, 17.4% had CD4 <200 cells/µL and 38.0% had CD4 <350 cells/µL at enrollment [13]—values below the 2025 surveillance bounds (0.305–0.501 and 0.480–0.675). The cohort itself reports that patients treated at participating hospitals who did not enroll had a worse prognosis than those who did, and that cohort survival (90.4%) exceeded that of all notified cases (83.4%) [13]; whether the gap between cohort and surveillance values reflects genuine improvement or this selection cannot be resolved from either source alone. Whether late presentation has improved since 2015 is therefore an open question that cohort or multicenter data with complete CD4 ascertainment, ideally with a recent-infection marker, are best placed to answer; the Korea HIV/AIDS Cohort Study would be an appropriate setting for such an analysis.

Compared with international benchmarks, the 2025 burden was high. Across 23 European Union/European Economic Area countries, 49.6% of new diagnoses in 2022–2024 were late, and the regional target of below 20% by 2025 was not met [14,15]; the Korean 2025 interval (0.480–0.675 by the CD4 criterion) lies at or above that European level and far above the target of below 20% by 2025 and below 10% by 2030 [15]. Across 31 European countries over a decade of surveillance, late presentation persisted and even increased in older adults [16]; globally the proportion presenting with AHD has remained close to 30% for years [17], and adoption of the World Health Organization (WHO) advanced-HIV-disease package of care into national guidelines has been incomplete [18]. The definitions are not identical—the European figures include AIDS-defining events and exclude recent infection—so the comparison is approximate, but under no interpretation does Korea approach the target.

Our contribution is less a new descriptive statistic than a reframing of the surveillance signal. We treat the missing-CD4 fraction as an object of analysis rather than as a nuisance, report bounds for every year so that their informativeness is visible, adjust the captured-only trend for capture completeness, and use the abrupt 2025 change in capture as a natural experiment. That experiment suggests that captured-only proportions in earlier years were inflated relative to the level among all notified cases—implying that some previously published late-presentation figures may overstate the burden among all notified cases while still leaving that burden high. The approach is generalizable: any surveillance system in which laboratory results are incompletely and inconsistently captured is vulnerable to the same artifact [7,8], and the practical recommendation that follows is that late-presentation indicators be reported together with the completeness of CD4 capture, so that apparent trends are not misread.

These findings have clinical and public health implications. Persistent late and advanced presentation has direct clinical consequences—greater opportunistic infection, poorer immune recovery, and higher mortality—and a public health consequence through onward transmission from undiagnosed infection. The actionable inference is that earlier-diagnosis strategies remain necessary: indicator-condition-guided testing in hospital settings identifies undiagnosed infection that routine practice misses [19], and HIV self-testing substantially increases testing uptake and reaches first-time and hard-to-engage testers in Asian and clinical settings [20,21].

This study has several limitations. This analysis used published aggregate tables, so individual-level linkage of CD4 with age, transmission route, or recent-infection status was not possible; the age adjustment and the capture adjustment are ecological and cannot exclude within-year confounding. The missing-data mechanism is not directly verifiable; the internal checks indicate that it was not random with respect to disease stage, which is why bounds and sensitivity analyses are reported rather than a single point estimate, and why no population-level trend is claimed. Because AIDS-defining events and recent-infection status were not captured in the aggregate tables, late presentation was operationalized from CD4 alone, and the recent-infection proportion used in the sensitivity analysis was not available for the CD4-captured subset and was assumed equal to that of all newly diagnosed Korean nationals; omitting AIDS-defining events understates the burden, whereas counting recent infections with transiently low CD4 counts overstates it, and the net effect cannot be determined from these data. The cohort comparisons involve small, clinic-based samples (101 patients in 2013–2015), inclusive cutoffs, and CD4 at enrollment rather than at diagnosis in the 2006–2023 series, and are indicative only. The joinpoint reversal after 2022 has a wide CI and overlaps the period of changing capture, so it should not be interpreted as a discrete clinical event. Finally, the analysis is descriptive and cannot establish the causes of persistent delay; linked clinical and behavioral data would be needed to identify them.

In 2025, the first year in which a CD4 result was captured for four in five notified cases, approximately half to two-thirds of new HIV diagnoses in Korea were late and one-third to one-half had advanced disease, a burden above contemporary European levels and far above the WHO European target. Over the preceding 28 years, trends among cases with a captured CD4 result tracked the completeness of capture, so whether late presentation has improved cannot be determined from surveillance data alone. Interpreting national late-presentation trends requires explicit attention to the completeness of CD4 ascertainment; the persistent burden documented in 2025 reinforces earlier diagnosis as a continuing clinical and public health priority, and cohort data with complete CD4 ascertainment are needed to establish whether it is improving.

Supplementary Materials

Supplementary Table 1

Annual counts and indicators, Korea, 1998–2025

Supplementary Figure 1

Robustness of the advanced HIV disease (AHD) trend among cases with a captured CD4 result.

Supplementary Figure 2

Advanced HIV disease (AHD) by denominator, Korea, 1998–2025.