Dengue virus (DENV) remains one of the most pressing vector-borne public health threats in Brazil, causing recurrent epidemics that vary widely in magnitude and geography. Four antigenically distinct serotypes (DENV-1-4) co-circulate in Brazil with regional and temporal variation. Primary infection typically confers lifelong immunity to the infecting serotype but transient cross-protection to others; subsequent heterotypic infections elevate the risk of severe outcomes via antibody-dependent enhancement (1–3). The 2024 season marked Brazil’s largest recorded dengue epidemic, with more than six million probable cases and tens of thousands of hospitalizations, imposing substantial strain on health systems (4). Because epidemic intensity varies over short distances, national and even state summaries can obscure actionable hotspots relevant to vector control, clinical surge planning, and targeted communication.

During epidemic peaks, routine laboratory confirmation is limited and most SINAN notifications rely on clinical-epidemiologic criteria (5). In this operational reality, geospatial methods that use syndromic notifications and hospitalizations can identify places where transmission is unusually intense or persistent even when virologic typing is incomplete. Spatial epidemiology offers multiple approaches, but several common methods require strong assumptions: k-means demands a pre-specified number of clusters and favors spherical geometries, and scan statistics impose moving windows that may not align with municipal boundaries. Density-Based Clustering of Applications with Noise (DBSCAN) is attractive because it does not require pre-specifying the number of clusters, can recover irregular shapes, and explicitly labels “noise,” separating isolated outliers from coherent high-density zones (6, 7). Although DBSCAN has seen growing use in infectious-disease surveillance internationally (8, 9), national-scale applications at Brazil’s municipal resolution remain limited. We implement DBSCAN on a joint feature space combining municipal centroids with standardized case and hospitalization rates.

Here we present a national application of multivariate DBSCAN framework that integrates municipal geography (centroids) with standardized epidemiologic burden (case and hospitalization rate per 100,000) to delineate dengue hotspots in 2024. We selected parameters (eps, minPts) using k-distance diagnostics and sensitivity checks (Supplementary Figure S1; Supplementary Table S1). Because response planning requires both spatial and temporal perspectives, we reran DBSCAN monthly using a common parameter set to assess persistence, the extent to which municipalities and populations remain in clusters across consecutive months (Supplementary Figure S2; Supplementary Table S4). Given the established linkage between rainfall and Aedes aegypti dynamics, we paired municipal dengue indicators with CHIRPS precipitation at 0–3 lags to characterize short-lag climate associations and operational lead time (27, 28).

We address three practical questions: (i) Where are the municipal clusters of greatest burden when geography and epidemiology jointly determine membership? (ii) How stable are these patterns over the epidemic year? (iii) How do short-lag rainfall patterns relate to the observed spatial structure? To aid interpretation, we treat “Cluster 1” as a low-burden background, focus on higher-intensity clusters (ID >1), and preserve isolated high-incidence municipalities labeled as noise (ID = 0). The resulting products include a national map of multivariate clusters (Figure 1); cluster-level summaries (Tables 1, 2; Figure 2); zoomed composite for exemplar high-burden areas (Figure 3); an outlier panel that preserves isolated hotspots (Figure 4); a climate-aligned national panel (Figure 5); and diagnostics and robustness checks (Supplementary Figures S1–S4; Supplementary Tables S1–S5).

Using a multivariate DBSCAN with eps = 0.3 and minPts = 3, we identified 25 geographically coherent clusters along with a large low-burden background (Cluster 1) and a small northern regional cluster (Cluster 2). In total, 5,111 municipalities (92.6%) were assigned to a cluster and 408 (7.4%) labeled as noise. Global Moran’s I for municipal case rates was strongly positive (I = 0.598; z = 72.8; p < 0.001), confirming non-random spatial autocorrelation consistent with the clustering patterns (Supplementary Table S2; Supplementary Figure S3).

High-burden clusters were geographically concentrated and often compact. Minas Gerais contained several of the most intense clusters, including a four-municipality cluster (Cluster 8: Jequitibá, Pequi, São Francisco de Paula, and Catas Altas) with average case rates of 24,000 per 100,000 and hospitalization rates of 76.6 per 100,000. Similarly, in Paraná, Cluster 13 (Arapuã, Ariranha do Ivaí, Corbélia, Janiópolis, and Primeiro de Maio) presented average hospitalization burdens of 284.2 per 100,000, while other southern clusters (e.g., Cluster 15) reached hospitalization burdens above 600 per 100,000. Cluster 4, concentrated in southern Minas Gerais and containing municipalities such as Guaxupé and São Gonçalo do Rio Abaixo, also demonstrated high overall burden (Table 3). These clusters are visualized at the national scale (Figure 1), with zoomed-in composite maps (Figure 3) and full municipal composition listed in Table 2, 3.

Despite the strong spatial signal, a subset of municipalities with extreme case rates remained unclustered because no similar-burden neighbors existed within the DBSCAN neighborhood radius. For example, Iguatu and Boa Vista da Aparecida in Paraná, and Paiva and Monjolos in Minas Gerais each exceeded 30,000 cases per 100,000 but were algorithmically excluded as noise (Supplementary Table S3). In total, 50 municipalities surpassed 22,000 cases per 100,000, many of which appear in the ranked bar plot of the top 50 municipalities (Figure 4B; Table 3). Presenting these outliers separately ensures that single-municipality hotspots are not eclipsed by density-based methods.

Monthly DBSCAN runs revealed stability in the overall footprint of clustering across the year. The majority of municipalities were consistently assigned to clusters, and large proportions of the national population resided in municipalities persisting in clusters for ≥3 consecutive months (Supplementary Table S4). Small-multiple maps (Supplementary Figure S2) highlight waxing and waning seasonal dynamics, yet the reappearance of the same macro-areas across months suggests structural vulnerability layered on seasonal forcing.

At national scale, monthly total dengue cases rose sharply early in 2024, aligning the seasonal maximum of CHIRPS precipitation (Figure 5) (37, 38). Macroregional correlation analyses confirmed positive rainfall-dengue associations, with Spearman coefficients strenthening from lag 0 to lag 2-3 months (Supplementary Table S5), consistent with prior studies (29–31). The strongest associations were observed in the Southeast and Central-West, while the South displayed positive but more modest correlations, patterns consistent with Aedes aegypti biology (10–13). A case study from Jequitibá (Minas Gerais, Cluster 8) reinforced this relationship where case surges in February–March closely followed local peak precipitation, and mean annual temperatures remained within the optimal range for vectorial capacity (23–26 °C).

Socioeconomic comparisons revealed that clustered municipalities had lower GDP per capita than non-clustered ones. Among the 5,519 municipalities, those in clusters had a lower average (26,891 BRL) and median (19,367 BRL) GDP per capita compared to non-clustered municipalities (33,743 BRL and 28,687 BRL, respectively). Logistic regression confirmed a modest but statistically significant inverse association between GDP per capita and the likelihood of cluster membership (OR ≈ 0.999994, p < 0.001). While the effect size is small, the direction of association aligns with literature linking poverty, inadequate water infrastructure, and limited health system access to arboviral vulnerability (15–18).

We present a national application of multivariate DBSCAN that integrates municipal geography with epidemiological burden to delineate dengue hotspots in Brazil’s record 2024 epidemic year. This approach offers three operational benefits. First, it identifies compact municipal clusters of exceptionally high cases and hospitalization rates, which represent natural focal areas for targeted vector control and surge capacity planning. Second, it preserves algorithmic “noise,” ensuring that municipalities with extreme but isolated burdens, often overlooked in density-based clustering, remain visible and actionable. Third, monthly repetitions of the analysis reveal temporal stability, highlighting persistent hot zones where interventions should be sustained across consecutive months. The month-to-month persistence of the same municipal hotspots, despite seasonal waxing and waning, signals structural vulnerability, arguing for sustained, area-based vector control and pre-positioned clinical surge capacity rather than episodic, reactive campaigns.

The clustering patterns we identified, particularly in Minas Gerais, Paraná, and Bahia, are parts of the South and Central-West, mirror the strong autocorrelation measured Moran’s I and are consistent with known ecological and infrastructural dengue drivers (4, 10, 15, 16, 19). The rainfall-dengue correlations reinforce the established expectation of positive associations at one- to three-month lags, consistent with Aedes aegypti life cycles and with previous findings in Brazil and elsewhere (27, 28). Case studies like Jequitibá highlight the close coupling of rainfall and dengue incidence at the municipal level, underlining the importance of integrating environmental data into early warning systems.

Methodologically, DBSCAN demonstrated several strengths compared to traditional approaches. Unlike k-means or hierarchical clustering, DBSCAN does not assume spherical clusters or pre-specify cluster number, enabling detection of irregularly shaped and context-specific hotspots (6, 7). Its explicit handling of “noise” is particularly valuable in national applications, where isolated municipalities may experience extreme outbreaks despite lacking nearby peers.

Our sensitivity analyses confirmed overall cluster geography was robust across parameter ranges (Supplementary Table S1). However, DBSCAN’s reliance on density continuity limits its ability to capture isolated hotspots, underscoring the need for complementary approaches such as Getis-Ord Gi* or Kulldorff’s scan statistics to ensure comprehensive hotspot detection.

Importantly, incorporating hospitalization rates into the clustering process allowed us to capture both transmission intensity and disease severity. Clusters in Paraná and southern Brazil reached hospitalization rates exceeding 200–600 per 100,000, highlighting areas of potential health system overload. These findings resonate with reports of higher severity among older populations and those with comorbidities, particularly during the southern epidemic wave (3, 4).

We acknowledge that case notifications in SINAN, particularly during epidemic peaks, are not often accompanied by systematic laboratory confirmation. Nonetheless, these clinically reported cases represent the operational data stream available for epidemic management in Brazil and are the same signals upon which national response planning relies. By demonstrating that robust spatial clusters emerge even under these constraints, our analysis underscores the utility of geospatial clustering as a pragmatic surveillance tool that complements but does not replace virologic confirmation. Moreover, the congruence of our results with known ecological drivers and hospitalization patterns affirms that signal-to-noise-ratios in the surveillance system are sufficient to identify meaningful hotspots.

The socioeconomic analyses further underscore the role of structural vulnerability in shaping dengue risk. Clustered municipalities tended to be less economically advantaged, with lower GDP per capita. While the observed effect sizes were modest, this directionality aligns with evidence that poverty, water storage practices, sanitation gaps, and housing conditions amplify arboviral exposure (15–18, 23). Future studies should integrate richer structural indicators, such as sanitation coverage, water intermittency, urban density, and health systems access, to evaluate multivariable predictors of cluster membership and persistence.

From a policy perspective, our findings emphasize that dengue control in Brazil cannot rely on aggregate national metrics alone. The identification of small, localized but high-burden clusters highlight the need for municipal and regional-level targeting of vector control, diagnostic distribution, and hospital surge planning. Moreover, the observed rainfall-dengue lagged correlations support the integration of climate data into predictive modeling and early warning systems, an especially urgent need as climate variability increases (20–22, 24).

Finally, DBSCAN remains underutilized in Latin American public health surveillance despite its adaptability, scalability, and compatibility with open-source workflows. Prior studies in Southeast Asia and the Caribbean have applied DBSCAN successfully to arboviral clustering (8, 9), but national applications in Brazil remain rare. Our analysis demonstrates its feasibility and value at the municipal scale, offering a flexible geospatial tool that can complement existing surveillance systems.

Looking forward, DBSCAN-based clustering, paired with Earth observation, climate predictors, and sociodemographic indicators, can underpin predictive analytics and decentralized epidemic intelligence, improving equity through more precise, and timely interventions.

Several limitations should be noted. First, although DBSCAN effectively delineates coherent high-burden clusters, its reliance on local density continuity inevitably excludes single municipalities with exceptionally high rates when they lack comparable neighbors. We partially mitigate this limitation by preserving these outliers in separate panels, but complementary methods such as Kulldorff’s spatial scan or Getis-Ord Gi* could further capture such isolated hotspots. Second, our reliance on monthly CHIRPS precipitation (37, 38) improves upon single-station data by providing uniform national coverage, but may still may obscure localized microclimatic variability (28) compared with INMET station data (26). Third, the use of routine surveillance data, often unconfirmed by laboratory diagnostics, introduces potential for misclassification; however, this reflects the operational reality of epidemic response and underscores the importance of methods that can extract robust signals from imperfect data. Finally, GDP per capita is a crude proxy of socioeconomic vulnerability and should be complemented in future work with more granular indicators of water, sanitation, housing, and health system capacity. Despite these limitations, the methodological transparency, reproducibility, and national coverage of our analysis position DBSCAN clustering as a valuable addition to the toolkit for epidemic intelligence in Brazil.

In Brazil’s unprecedented 2024 dengue year, multivariate DBSCAN uncovered compact municipal clusters of high burden and preserved isolated outliers that demand targeted action. Monthly clustering showed persistence of risk in the same macro-areas across seasons, while rainfall correlations at short lags confirmed expected climate-epidemic coupling. The method is transparent, scalable, and immediately useful for prioritizing vector control, diagnostics, and hospital surge planning at municipal scale. As Brazil advances decentralized surveillance and climate-aware preparedness, density-based geospatial clustering can help bridge the gap between national statistics and neighborhood-level action.