📄 public and global health

Lessons learned from real-time nowcasting: The 2024 dengue outbreak in Puerto Rico

This paper demonstrates that applying Nowcasting by Bayesian Smoothing (NobBS) to the 2024 dengue outbreak in Puerto Rico effectively mitigated reporting delays to accurately track transmission dynamics, while highlighting that joint parameter estimation improves performance for sparse data and stable reporting patterns are crucial for reliable real-time public health nowcasting.

Original authors: Tran, Q. M., Detmar, A. L., Liu, C. Y., Madewell, Z. J., Rodriguez, D. M., Aponte, J. T., Marzan-Rodriguez, M., Paz-Bailey, G., Adams, L., Holcomb, K., Johansson, M. A., Thayer, M.

Published 2026-07-23

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Tran, Q. M., Detmar, A. L., Liu, C. Y., Madewell, Z. J., Rodriguez, D. M., Aponte, J. T., Marzan-Rodriguez, M., Paz-Bailey, G., Adams, L., Holcomb, K., Johansson, M. A., Thayer, M.

Original paper dedicated to the public domain under CC0 1.0 (https://creativecommons.org/publicdomain/zero/1.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to watch a live sports game, but your TV signal is glitchy. Every time a player scores, the broadcast doesn't show it immediately; sometimes it takes a few minutes, sometimes an hour, and occasionally, the whole scoreboard freezes and dumps all the missed points at once. If you only looked at the scoreboard as it appeared in real-time, you might think the game was boring or that the team was losing, when in reality, they were dominating. This is exactly the problem public health officials face with infectious diseases like dengue fever. There is always a "lag" between when a person gets sick, when they see a doctor, when the lab confirms the virus, and when that information finally gets entered into the government's database. By the time officials see the numbers, the outbreak might have already moved on, making it hard to know if the danger is rising or falling. To fix this, scientists use a technique called "nowcasting." Think of it as a super-smart weather forecast for the present moment. Instead of just looking at the rain that has already fallen, a nowcast uses patterns from the past to guess how much rain is currently falling but hasn't been measured yet, giving a much clearer picture of the storm right now.

This paper tells the story of how scientists in Puerto Rico used this "present-moment weather forecast" to track a massive dengue outbreak in 2024. Dengue is a mosquito-borne virus that is common in the region, and in early 2024, the numbers started to climb. The researchers applied a sophisticated mathematical tool called "Nowcasting by Bayesian Smoothing" (NobBS) to the island's health data. Their main goal was to see if this method could spot the outbreak earlier than the standard reporting system and if it could handle the messy reality of real-world data, where reports sometimes get delayed or dumped in big batches.

The results were a mix of high-tech success and a few messy lessons. The NobBS model was a star player. It successfully predicted the rise of the outbreak about four weeks before the official reported numbers crossed the danger line. In fact, the model signaled that an epidemic was likely starting in late January, while the official count didn't hit the threshold until early February. This early warning gave health officials a crucial head start, helping them declare a public health emergency and mobilize resources like mosquito control and public awareness campaigns before the situation got out of hand. The model was generally very good at tracking the total number of cases, often predicting the final count weeks in advance.

However, the story isn't just about a perfect robot; it's about how the model handled human errors. The researchers found that the model worked best when the data was steady. But in October 2024, a strange thing happened: a "batch reporting" event where 119 cases from a single week were reported all at once after being delayed for up to three months. This glitch confused the model, causing it to miss the mark for a few weeks. It's like if your TV suddenly showed three months' worth of game highlights all in one minute; the scoreboard would look wildly wrong for a moment. The study also tested two different ways of running the model for smaller groups, like specific islands or different types of the virus. They found that when there were very few cases in a specific area, it was better to "pool" the data—sharing information between groups to make the estimates more stable—rather than trying to guess each area's numbers independently. Trying to guess alone in a low-data area was like trying to predict the weather in a tiny town with only one thermometer; it was too shaky.

The paper also looked back at past outbreaks to see how the model would have performed in different years. It did well in 2009 and 2010, but struggled in 2013. Why? Because in 2013, the time it took to report cases was all over the place, swinging wildly from two weeks to six weeks. The model suggests that while it can handle some bumps, it gets confused when the reporting system itself is erratic. The authors conclude that while nowcasting is a powerful tool that can save lives by providing early warnings, it relies heavily on having a stable reporting system. If the data stream is broken or inconsistent, even the smartest model will stumble. Ultimately, this study shows that with the right tools and a bit of patience for data glitches, we can see the "invisible" part of an outbreak and respond faster than ever before.

Technical Summary: Lessons Learned from Real-Time Nowcasting of the 2024 Dengue Outbreak in Puerto Rico

Problem Statement
Timely situational awareness is critical for effective infectious disease outbreak response. However, case surveillance data is frequently compromised by reporting delays caused by lags in care-seeking, clinical confirmation, and laboratory reporting. These delays result in reported case counts that reflect past transmission levels rather than current reality, often creating a false perception of declining transmission when cases are plotted by symptom onset. In the context of the 2024 dengue outbreak in Puerto Rico (PR), these delays hindered the accurate assessment of transmission dynamics and the timely declaration of a public health emergency.

Methodology
The authors applied a Nowcasting by Bayesian Smoothing (NobBS) framework to dengue surveillance data from the Puerto Rico Passive Arboviral Disease Surveillance System (PADSS), covering data from 1986 to 2024.

Data: The analysis utilized probable (IgM positive) and confirmed (RT-PCR positive) dengue cases, aggregated by epidemiological week (EW). Both case types were modeled jointly with a single delay distribution.
Model Architecture: The NobBS model, reimplemented in Stan for computational efficiency, estimates the time-varying epidemiological signal ( $\alpha_t$ ), reporting delay probabilities ( $\beta_d$ ), and dispersion parameters. It generates nowcasts for horizons -2, -1, and 0 (two weeks prior, one week prior, and the current week).
Stratification and Parameterization: Nowcasts were generated at the island-wide level, as well as by eight health regions and four dengue virus (DENV) serotypes. To address data sparsity in subgroups, two modeling approaches were compared:
1. Independent Parameters: Unique parameters for transmission and reporting for each group.
2. Shared Parameters: Jointly estimated uncertainty parameters ( $r$ and $\phi$ ) across groups to pool information, stabilizing estimates for low-incidence groups.
Benchmarking and Evaluation: The NobBS models were compared against a simple baseline model (predicting the most recent observed value). Performance was evaluated using Weighted Interval Scores (WIS), decomposed into dispersion, absolute error, underprediction, and overprediction. Calibration was assessed via 50% and 95% predictive interval (PI) coverage. A "failed nowcast" was defined as a log-score below -9.2 for at least one horizon.
Retrospective Analysis: The framework was applied retrospectively to historical data (2009, 2010, 2013) to assess performance under varying incidence levels and reporting delay variabilities.

Key Results

Early Detection: In January 2024, island-wide nowcasts signaled that case numbers were approaching the epidemic threshold four weeks before reported cases did. The 25th percentile of the nowcast exceeded the threshold on January 27, while reported cases did not cross it until February 3. This early warning contributed to the declaration of a public health emergency in March.
Model Performance (2024): The NobBS model (WIS: 76.7) substantially outperformed the baseline model (WIS: 134.1), primarily by reducing dispersion and absolute error. However, the model exhibited a tendency toward overprediction.
Impact of Reporting Anomalies: Performance was significantly degraded by a batch reporting event in EW 47, where 119 cases were reported with delays exceeding four weeks. This event caused several "failed nowcasts" and narrowed prediction intervals below expected coverage targets (median 95% PI coverage was 0.88 vs. the target of 0.95).
Subgroup Analysis:
- Shared vs. Independent: For health regions and serotypes with low case counts, the shared parameter model generally achieved lower WIS and reduced uncertainty compared to the independent model. The independent model produced excessively wide intervals for rare events (e.g., a single DENV-4 case in EW 40), leading to failures under the strict log-score criteria.
- Trade-offs: While the shared model improved stability, it occasionally failed to capture local variations, resulting in overconfidence (tighter intervals) when local patterns deviated from the pooled trend.
Historical Context: The model performed well in low (2009) and high (2010) incidence years. However, performance degraded during the 2013 outbreak due to highly variable reporting delays (ranging from two to six weeks), demonstrating that model accuracy is limited by the stability of the reporting system.

Significance and Claims
The paper posits that real-time nowcasting is a reliable and valuable public health tool for dengue surveillance, provided specific conditions are met. The authors derive three primary lessons:

Reporting Stability is Critical: The accuracy of nowcasts is heavily dependent on consistent reporting patterns. Irregularities, such as batch reporting or erratic delays, can substantially reduce performance. Routine monitoring of reporting delays and anomaly detection are essential for model calibration.
Adaptive Modeling Strategies: For subgroups with low case counts, pooling parameters (shared models) improves statistical power and precision compared to independent models. However, adaptive strategies, including real-time adjustment of delay distributions or model switching, are necessary to handle complex or changing reporting conditions.
Actionable Integration: Predictive alerts must be integrated into decision-making processes through strong engagement with local health partners. The 2024 experience demonstrated that nowcasts can provide actionable evidence to support the declaration of epidemics and guide resource allocation, vector control, and public awareness campaigns.

The authors conclude that while the framework is feasible and effective, its utility is maximized in settings with continuous, reliable data flows and requires close collaboration between modelers and decision-makers to translate statistical outputs into operational interventions.

Technical Summary: Lessons Learned from Real-Time Nowcasting of the 2024 Dengue Outbreak in Puerto Rico

More like this