Dengue has moved through Thailand for more than six decades, and now a team studying two decades of hemorrhagic fever records from all 77 provinces has asked a quiet but consequential question: what best tells us where and when the danger will rise? Their answer, drawn from machine learning applied to surveillance, climate and socioeconomic data, is that temperature speaks louder than rain, with the mosquito's own biological limits appearing as sharp thresholds in the numbers. The model read the pre-pandemic years with notable clarity, yet it faltered when human life itself was reorganized by
Temperature, Not Rainfall, Emerges as Dominant Predictor of Dengue Risk Across Thailand's Provinces
Dengue hemorrhagic fever causes plasma leakage, hemorrhage and organ impairment; the study analyzed two decades of DHF surveillance cases across Thailand.
Below roughly 21°C, dengue transmission essentially stops.
Mark
Why does it matter that temperature beat rainfall? Rainfall is what most forecasts lean on.
Mimi
Because forecasters have leaned on it heavily, and the study says its link to dengue incidence has been inconsistent. Here, seven of the top ten features were temperature, and rainfall barely registered.
Mark
What's the 21°C figure about?
Mimi
It's a threshold the model found for minimum temperature lagged one month. Below about 21°C, transmission essentially stops. There's also one near 32°C for maximum temperature lagged three months.
Luke
Careful, though. Those thresholds come from a statistical model. The authors say they match known mosquito biology, but the model didn't prove that mechanism.
Mark
And the 80% accuracy?
Mimi
That's the AUC of 0.80 in the pre-pandemic test period. It was 0.74 across the later period combined.
Luke
Worth saying that AUC isn't quite the same as being right 80% of the time. The author summary uses that shorthand, but it's a looser description of the metric.
Mark
Why did COVID break it?
Mimi
The model overpredicted risk from 2020 to 2022, and the AUC dropped to 0.62 in 2021. The authors point to control measures that climate data couldn't capture.
Luke
That's their inference. The excerpt doesn't show behavior or surveillance changes being measured directly, so it remains likely rather than demonstrated.
Mark
And the link to wealth?
Mimi
Higher Gross Provincial Product went with higher risk, which the authors tie to urban conditions and density.
El Pulso
- A disease that has circulated in Thailand since 1958 still threatens provinces every year, and forecasters need to know where the next surge will strike.
- Temperature, not rainfall, dominated the model, with seven of the ten strongest predictors tied to heat and transmission appearing to halt when minimum temperatures drop below about 21°C.
- The XGBoost classifier flagged high-risk areas roughly 80% of the time before the pandemic, but accuracy slid to an AUC of 0.62 in 2021 as lockdowns and behavior shifted the ground beneath it.
- Researchers are steering toward real-time tools that blend climate readings with behavioral and surveillance indicators, aiming to keep forecasts reliable through the next major disruption.
Dengue has moved through Thailand for more than six decades, and now a team studying two decades of hemorrhagic fever records from all 77 provinces has asked a quiet but consequential question: what best tells us where and when the danger will rise? Their answer, drawn from machine learning applied to surveillance, climate and socioeconomic data, is that temperature speaks louder than rain, with the mosquito's own biological limits appearing as sharp thresholds in the numbers. The model read the pre-pandemic years with notable clarity, yet it faltered when human life itself was reorganized by COVID-19, a reminder that nature's rhythms can be forecast only so far when societies change their own patterns.
Every year, dengue hemorrhagic fever finds its way into Thailand's provinces, leaving plasma leakage, hemorrhage and damaged organs in its wake. A research team led by Suttirat, working with Chadsuthi, Aekthong, Rocklöv, Bicout and Haddawy, set out to learn whether the disease's movements could be anticipated. They gathered monthly case counts from every province between 2003 and 2022 and paired them with 54 environmental, climatic and socioeconomic features, then tested four classifiers before settling on XGBoost. The paper, published in PLOS Neglected Tropical Diseases on July 28, 2026, reports what the model learned.
Temperature emerged as the dominant signal. Seven of the top ten predictors were heat measures, and the model placed sharp thresholds near 21°C for minimum temperature and 32°C for maximum temperature, below which transmission largely ceases. The authors note these values match known limits on Aedes aegypti mosquitoes. Rainfall, a long-favored forecasting variable, added little. Economic output also surprised: provinces with higher Gross Provincial Product faced greater risk, a pattern the authors tie to dense urban settings where mosquitoes breed year-round and the virus spreads easily.
Before the pandemic, the model identified high-risk areas about 80% of the time. Then COVID-19 arrived, and accuracy fell to an AUC of 0.62 in 2021, with predictions systematically overshooting actual cases from 2020 to 2022. The authors believe that mobility restrictions and other pandemic controls reshaped dengue dynamics beyond what climate data could capture. Their conclusion is measured: environmental models perform well in stable times but weaken during major social disruption, so future real-time forecasts should incorporate behavioral and surveillance indicators alongside climate. The work was supported by the Rockefeller Foundation and Mahidol University.
The dataset behind this study is large: monthly dengue hemorrhagic fever counts from each of Thailand's 77 provinces, running from 2003 through 2022. Researchers led by Suttirat, with colleagues including Chadsuthi, Aekthong, Rocklöv, Bicout and Haddawy, combined those surveillance records with 54 environmental, climatic and socioeconomic features and asked which of them best predicts where and when dengue risk runs high. The paper appeared in PLOS Neglected Tropical Diseases on July 28, 2026.
The backdrop is a disease that has been in Thailand since at least 1958, when the country recorded its first dengue hemorrhagic fever epidemic. Globally, dengue is estimated to cause 390 million infections a year across more than 100 countries, and incidence has risen 30-fold over five decades. Southeast Asia carries roughly 70% of the burden, and Thailand reports some of the region's highest rates. The most severe forms of the disease can cause plasma leakage, hemorrhage and organ impairment.
The team tested four types of classifier: logistic regression, support vector machines, random forests and XGBoost. They chose XGBoost based on six performance metrics, and used a method called SHAP to see how much each feature contributed to its predictions. They focused on hemorrhagic fever cases rather than milder dengue fever, which is systematically under-reported, or dengue shock syndrome, which is too rare at province level. Of 18,240 province-months, 59.0% had no shock syndrome cases at all. The median hemorrhagic fever incidence, 2.78 per 100,000, gave a balanced split between high-risk and low-risk months.
The model was trained on 2003 to 2016 and tested on later years, divided into pre-COVID (2017–2019), COVID (2020–2021) and post-COVID (2022) periods. Before the pandemic it scored an AUC of 0.80, which the authors describe as identifying high-risk areas about 80% of the time. Across the combined during- and post-pandemic period, the AUC was 0.74.
Temperature dominated. Seven of the ten most important features were temperature measures. The model found non-linear thresholds near 21°C for minimum temperature lagged one month, and near 32°C for maximum temperature lagged three months. Below roughly 21°C, transmission essentially stops. The authors say these values align with known biological limits on transmission by Aedes aegypti mosquitoes. Rainfall, long emphasized in dengue forecasting, contributed little to the predictions.
Economic development pointed the other way from older assumptions. Higher Gross Provincial Product was associated with greater dengue risk. The authors link this to predominantly urban transmission: urbanization can give mosquitoes year-round breeding conditions, and dense populations help the virus spread. They also note that dengue now affects rural and urban populations across economic strata, so it is no longer simply a disease of poverty.
The pandemic exposed the model's limits. Performance fell to an AUC of 0.62 in 2021, and the model systematically predicted more risk than actually occurred from 2020 to 2022. The authors infer that factors outside the model, likely COVID-19 control measures such as mobility restrictions, dominated dengue dynamics.
Their conclusion is that environmental prediction works under stable conditions but degrades during major societal disruption. For real-time use, they recommend adding indicators of behavior and surveillance coverage alongside the climate data. The study was funded by the Rockefeller Foundation and Mahidol University, and the authors declare no competing interests.
Citas Notables
Environmentally driven prediction is reliable under stationary conditions but degrades substantially during periods of major societal disruption.— Study authors, PLOS Neglected Tropical Diseases
Quieres la nota completa?
Lee el original en PLOS (Public Library of Science) ↗
Technology