Study design and data source

This study constitutes a retrospective time-series analysis of secondary, national-level data. The primary outcome was monthly dengue admission data spanning from January 2008 to October 2025. Data were collected from the publicly available archives and situation reports of the Institute of Epidemiology Disease Control and Research (IEDCR) and the Directorate General of Health Services (DGHS) [2,11]. This passive surveillance system primarily captures reported cases from all hospitals under DGHS (both public and private), with likely underreporting of milder community cases. Data included complete surveillance records through October 2025, representing actual reported cases rather than projections or forecasts. No monthly values were missing in the compiled national aggregates, so no imputation was required. This complete dataset enables confirmation of sustained transmission patterns and regime stability.

The reporting of this study follows the STROBE guidelines for observational studies.

Analytical framework: STL decomposition

To isolate and quantify the underlying patterns in the monthly dengue case data, we employed the STL (Seasonal-Trend decomposition using LOESS) method. STL is a robust and versatile filtering procedure for decomposing a time series into three additive components:

Where:

• Y_t is the observed dengue case count at time t.
• T_t is the Trend-cycle component, representing long-term progression (e.g., multi-year increases or decreases).
• S_t is the Seasonal component, representing regular, fixed-period fluctuations (e.g., annual cycles).
• R_t is the Remainder (or residual) component, containing the irregular noise or randomness not explained by the trend or seasonality.

An additive model was selected because visual inspection and component strength metrics showed relatively stable seasonal amplitude compared with the explosive trend component. The STL algorithm employs an iterative, two-loop process using LOESS (Locally Estimated Scatterplot Smoothing) regression and robustness weights to ensure the decomposition is not skewed by outliers. Implementation parameters in Python’s statsmodels library were: period = 12, seasonal = 13, and robust = True.

To provide objective metrics, we calculated the strength of the trend (F_T) and seasonality (F_S) using established formulae:

Trend strength:

Seasonality strength:

Where, Var(⋅) denotes the sample variance. Values close to 1 indicate the component dominates the remainder, signifying a strong pattern relative to the noise.

Statistical analysis: Structural break identification and regime delineation

1. 1. Structural Break Detection

Primary Break Detection: To formally test for fundamental shifts, we employed the Zivot-Andrews test [7], a robust procedure that endogenously identifies a single unknown breakpoint, thus avoiding the need for a priori specification of break dates. The test is based on an augmented regression model that incorporates dummy variables to capture a break in both the level (intercept) and the trend of the series:

Here:

• t is a linear time trend.
• DU_t (λ) is an intercept break dummy.
• DT_t (λ) is a trend break dummy.

The null hypothesis (H₀: α = 0) posits a unit root without a break, while the alternative (H₁: α < 0) indicates a trend-stationary process with a break at time T_B. The break date λ is selected to minimize the one-sided t-statistic for α, identifying the most likely point of structural change. Statistically significant breakpoints (p < 0.05) were subsequently interpreted as the start of new epidemiological phases.

Multi-Algorithm Consensus Approach: To robustly identify multiple regime shifts, we implemented a consensus-based multiple breakpoint detection framework combining three state-of-the-art change-point algorithms after the primary Zivot-Andrews test:

• PELT (Pruned Exact Linear Time) – optimal for known number of changes [12]
• Binary Segmentation – fast heuristic for multiple breaks [13]
• Window-based segmentation – sliding window comparison [14]

All algorithms used radial basis function (RBF) cost functions with min_size = 12 months. Breakpoints were ranked by consensus score (frequency across methods) and magnitude of mean change (pre- vs post-break). Breaks detected by ≥2 algorithms or showing large mean ratio changes were prioritized. The top-ranked breaks defined candidate epidemiological transition points.

1. 2. Transmission Regime Identification

Two complementary approaches were used to delineate distinct epidemiological regimes:

1. Unsupervised clustering of transmission features (K-means, k = 2–5) on log-transformed case counts, 12-month rolling mean, coefficient of variation, and outbreak intensity (z-score > 2). Optimal cluster number was determined via silhouette score.
2. Markov-switching regression model [15] with k = 3 regimes, allowing switching mean and variance:

is the unobserved regime at time t, governed by a first-order Markov chain. Smoothed regime probabilities were used to assign dominant regimes and detect transition points.

Previously, unsupervised clustering (e.g., K-means) has been successfully applied to epidemic time-series to classify outbreak patterns (e.g., “-outbreak types”- during COVID-19) [16]. Separately, Markov-switching (regime-switching) models have long been used to detect epidemic vs non-epidemic phases in surveillance data, allowing for switching mean and variance across latent states (e.g., influenza surveillance) [17].

The Markov regime-switching model served as the primary method for defining final epidemiological regimes, as it accounts for temporal dependence and provides smoothed probabilistic state assignments. K-means clustering and structural break analysis provided complementary validation. This multi-method synthesis allowed ten structural breaks to be identified and coalesced into three main epidemiological regimes.

1. 3. Unit Root and Stationarity Testing with Structural Breaks

Series non-stationarity was assessed using:

• Zivot–Andrews test (Model C: break in intercept and trend) [7]
• Augmented Dickey–Fuller (ADF) [18]
• Phillips–Perron (PP) [19]
• DF-GLS tests [20]

A significant Zivot–Andrews test (p < 0.05) confirmed the presence of trend-stationarity with endogenous structural break(s).

1. 4. Epidemiological Metrics and Outbreak Definition

Key metrics included:

• Coefficient of variation (CV)
• Annual growth rate (linear trend coefficient × 12/ mean × 100)
• Outbreak frequency: proportion of months exceeding 2σ above 12-month centered rolling mean
• Seasonal ratio: peak vs trough monthly mean

Software and reproducibility

All analyses were implemented in Python 3.13 using:

• Pandas, numpy, matplotlib, seaborn
• Statsmodels (STL, Zivot–Andrews, MarkovRegression) [21].
• Ruptures (PELT, BinSeg, Window)
• Scikit-learn (K-means, silhouette score)
• Arch (Phillips–Perron, DF-GLS)

Ethics statement

This study analyzed exclusively aggregated, anonymized, and publicly available data from official government portals. No individual-level data was accessed. Following international guidelines (e.g., CIOMS) for epidemiological research using public surveillance data, this study was considered exempt from ethical board approval. The research was conducted with rigorous adherence to scientific integrity and the principles of the Declaration of Helsinki.

Conceptual framework

Structural breaks are points in time when the underlying parameters of dengue transmission shift sustained, usually due to events such as viral change, climatic variation, or major societal disruptions [22]. Transmission regimes are the sustained periods that follow these breakpoints and are defined by stable statistical properties such as mean incidence and seasonal pattern [23,24]. Breaks mark the transition, and regimes describe the new operating state of the system. These shifts in Bangladesh likely reflect combined effects of viral evolution, climate change, rapid urbanization, population mobility, and external shocks such as the COVID lockdown [25]. Distinguishing between breaks and regimes is essential for interpretation of dengue trends. Breaks provide early warning of fundamental change, while regimes offer the basis for long term planning and program design.

Seasonality and trend components

STL decomposition showed a highly dominant trend component (high trend strength) and moderate seasonality strength (0.335), indicating that long-term trend variability predominates over regular seasonal fluctuations in explaining overall series variance.

Nevertheless, the peak-to-trough seasonal ratio reached approximately 180, highlighting extremely intense amplification during annual monsoon epidemic cycles atop the elevated post-2023 baseline. The moderate seasonality strength value indicates the explosive trend dominates total variance, while the high seasonal ratio confirms that within each year, the monsoon-driven peak is massively amplified relative to the new, elevated baseline.

Overall series metrics (January 2008–October 2025): total cases 711,618, mean monthly ≈3,325, SD ≈ 10,727, CV ≈ 3.23, maximum monthly 73,023, minimum monthly 0. Linear trend coefficient ≈63.9 cases/month, estimated annual growth ≈23.1%, doubling time ≈3.0 years (see Supplementary epidemiological metrics).

Structural breaks and epidemiological phases

The Zivot-Andrews test confirmed a significant unit root with a structural break (Test Statistic: −8.277, p < 0.001). The multi-algorithm consensus approach identified ten structural breaks, with the most statistically significant breaks occurring in 2021−05, 2020−02, 2015−07, and 2011−05.

Structural breaks

Structural Breaks The consensus-based multi-algorithm approach (PELT, Binary Segmentation, Window-based segmentation) identified ten significant structural breaks in the monthly dengue admission time series (Supplementary Table S1 File). Breaks were ranked by consensus score (number of algorithms detecting the date) and magnitude of mean change. The highest-consensus break was in May 2021 (consensus score 3, detected by all three algorithms), with pre-break mean ≈819 cases/month increasing to post-break mean ≈10,750 (ratio ≈13.1, magnitude change ≈9,931).

Other prominent breaks included:

• July 2015 (score 2, ratio ≈82.6, magnitude ≈5,620)
• February 2020 (score 2, ratio ≈9.4, magnitude ≈7,536)
• June 2023 (score 1, ratio ≈14.0, magnitude ≈15,660)
• September 2024 (score 1, ratio ≈4.3, magnitude ≈8,946)

The full list of ten breaks, including dates, consensus scores, detecting methods, pre- and post-break means, ratios, and magnitude changes, is provided in Supplementary Table S1 File. K-means clustering applied to regime features produced an optimal solution of three clusters (silhouette score 0.867, the highest among 2–5 clusters evaluated; see Supplementary Table S2 File). This clustering solution was supported and refined by Markov regime-switching models.

The ten structural breaks were synthesized into three distinct epidemiological regimes (Supplementary Table S4 File) (Table 1):

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Identified transmission regimes in dengue admissions (2008-2025) based on markov regime-switching model.

https://doi.org/10.1371/journal.pone.0343246.t001

The visual progression of these regimes, with their corresponding changes in mean and variance, is presented in. The most critical shift occurred in mid-2019, initiating the Hyperendemic Regime, which dramatically and sustainedly elevated the expected scale of outbreaks, as evidenced by the historic peak of 73,023 monthly admissions in September 2023. The Hyperendemic Regime has sustained a mean of 26,127 monthly cases—a 556-fold increase over the Low Baseline Regime—and has persisted through October 2025 (although expected persistence being 24 months). This aligns with spatial studies confirming nationwide dissemination during this period (Hossain et al., 2023). The Markov Regime-Switching and clustering models confirm that, after the extreme 2023–2024 peak, Bangladesh has remained locked in a dramatically elevated transmission state through October 2025.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Epidemiological metrics and seasonality analysis in Bangladesh (2008-2025).

Fig 1 integrates four complementary panels that together demonstrate a profound structural shift in dengue transmission in Bangladesh. The top-left panel shows that monthly admitted cases stayed low and stable until 2018, rose sharply with the 2019 escalation, peaked explosively in 2023, and then settled into a new but still highly elevated baseline. The top-right panel reveals that yearly outbreak patterns have shifted from rare, isolated events before 2019 to near-continuous outbreak conditions thereafter, with 2023 exhibiting unprecedented sustained transmission. The bottom-left panel shows that seasonality remains strongly monsoon-driven and highly consistent, though now sitting atop a far higher baseline. Finally, the bottom-right panel uses clustering to identify three clear transmission regimes (from unsupervised k means clustering)—low-endemic, transitional escalation, and hyperendemic—indicating that the system no longer returns to its pre-2019 state. Taken together, these panels provide compelling visual evidence of a lasting structural shift toward sustained hyperendemic dengue transmission.

https://doi.org/10.1371/journal.pone.0343246.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Epidemiological timeline of monthly dengue admissions in Bangladesh (2008-2025, October).

This figure presents a unified epidemiological timeline across four panels, clearly demonstrating the structural transformation of dengue transmission in Bangladesh. The top panel overlays monthly cases with data-driven regime classifications (from unsupervised k means clustering) and detected breakpoints, showing three major phases: a long low-endemic era, a sharp escalation beginning around 2019–2021, and a sustained hyperendemic period through 2024–2025, all marked by abrupt, algorithm-identified shifts. The second panel displays outbreak intensity using Z-scores, revealing that months exceeding the + 2σ threshold were almost absent before 2019 but became persistent and extreme from 2022 onward, confirming that outbreaks transitioned from rare anomalies to the default state. The third panel compares seasonal patterns across eras, showing that the monsoon-driven peak remains intact but is now magnified by an order of magnitude in the hyperendemic era. Finally, the bottom panel charts cumulative dengue cases, which remained nearly flat for a decade before accelerating sharply in alignment with the identified structural breaks, forming a clear “hockey-stick” trajectory. Together, these panels provide strong, convergent evidence of a sustained shift to high, structurally altered dengue transmission.

https://doi.org/10.1371/journal.pone.0343246.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Markov regime switching analysis.

This figure uses a Markov regime-switching framework to identify dengue transmission phases, and both panels show a clear, escalating shift toward hyperendemicity. In the top panel, the model classifies monthly dengue cases into three distinct regimes—an extremely low baseline through mid-2019, a sharply elevated intermediate phase through ~2022, and a dramatic hyperendemic regime thereafter—mirroring the same structural jumps detected by all other analytical methods. The bottom panel, which plots smoothed regime probabilities, reveals how these transitions unfolded: the probability of the old low-endemic regime stayed near 1.0 until 2018, collapsed abruptly around 2019 as the intermediate phase took over, and then gave way to the hyperendemic regime, whose probability rises to nearly 1.0 from 2022 onward and remains dominant even during the 2023 peak. Together, the two panels show that the transition was not gradual drift but a sequence of decisive, model-verified shifts culminating in a stable hyperendemic state.

https://doi.org/10.1371/journal.pone.0343246.g003

The progression through these regimes shows an alarming acceleration: the Low Baseline Regime persisted for approximately 15 years, the Intermediate Regime overlapped throughout much of the study period, while the Hyperendemic Regime has established a new, sustained baseline.

Addressing surveillance bias

While improvements in diagnostic capacity and reporting over the 18-year period may contribute to observed trends, the scale of the surge—particularly the 556-fold increase from the low baseline to the hyperendemic regime—argues strongly against artifact alone. Parallel evidence from hospital burden, mortality trends, and media reports during the 2023–2024 peak supports the reality of a true epidemiological escalation beyond improved detection [33]. Furthermore, structural breaks represent sudden changes in the data-generating process; gradual improvements in surveillance would be unlikely to produce such sharp, statistically identifiable discontinuities.

The implications for public health strategy are substantial. Dengue transmission in Bangladesh now operates under a new hyperendemic regime with a dramatically elevated baseline (mean ≈ 26,127 monthly cases). Preparedness strategies must therefore assume this consistently higher and more volatile baseline, rather than relying on assumptions drawn from earlier, lower-intensity phases.

Public health implications

The identification of these structural breaks and the confirmed sustained hyperendemic regime has profound implications for public health practice:

• Forecasting and Surveillance: Models that assume a single, stable state are inherently unreliable. Future forecasting efforts must be regime-specific and incorporate regime-shift detection algorithms to provide accurate early warnings. The historical data from the pre-2023 regimes is no longer a valid benchmark for predicting outbreak magnitude in the current hyperendemic regime. The sustained transmission through October 2025 provides the new baseline for all future projections. The detection of a break within the hyperendemic phase in September 2024 underscores the need for continuous, real-time surveillance to detect further escalations.
• Preparedness and Resource Allocation: The Hyperendemic Regime (2023–2025) provides the new baseline for preparedness. Health system capacity, including hospital beds, ICU facilities, and medical supplies, must be permanently scaled to the reality of sustained, high-level transmission with a new baseline of approximately 26,127 mean monthly cases. The confirmation that this elevated state has persisted for more than 2 years means budgetary planning must reflect this sustained increase in case load, not just seasonal surge capacity.
• Intervention Strategy: Vector control and public health messaging strategies must be fundamentally re-evaluated for efficacy in a sustained high-transmission regime. Approaches designed for a lower-intensity endemic state are completely insufficient and need to be replaced with more aggressive, sustained, and potentially novel interventions tailored to this new reality. The intense seasonality (seasonal ratio ≈ 180) suggests that targeted, high-intensity vector control in the months leading up to the August peak could yield significant reductions in disease burden, but must be implemented at a scale appropriate for the new hyperendemic baseline. The 2024 break indicates that even current control measures may be inadequate to contain the intensifying transmission. Interventions must now be year-round, with pre-monsoon intensification (March–May) critical to mitigating the explosive August peak.

Strengths and limitations

As an ecological study using aggregated national surveillance data, causal inference on drivers (e.g., serotype changes, climate, urbanization) is limited to temporal alignment with published literature. The univariate time-series approach does not directly incorporate covariate data. Passive surveillance likely results in substantial underreporting (especially of mild or non-hospitalized cases), and diagnostic/reporting improvements may contribute to apparent trends — though the scale of the surge argues strongly against artifact alone. Future research integrating this breakpoint framework with hierarchical models incorporating spatially-explicit climate data, serotype prevalence, and human mobility data is needed to attribute causal weights to these drivers.

In summary, Bangladesh has entered a sustained hyperendemic dengue era (ongoing as of October 2025). Public health planning, resource allocation, and intervention strategies must be fundamentally adapted to this new reality to mitigate morbidity, mortality, and healthcare system strain.