Disease case data

Disease surveillance data between 2003 and 2019 was obtained from Thailand’s Ministry of Public Health disease surveillance system which records reported monthly disease case counts at the province level [17]. Bueng Khan province was separated from Nong Khai in 2011, therefore for consistency over the timeframe of the dataset, we merged disease case counts from Bueng Khan back to Nong Khai from 2011 onwards. We considered seven major infectious diseases circulating in Thailand, which include the common infectious diseases that are reported by Thailand’s Ministry of Public Health [17]. The diseases include Japanese encephalitis virus (JEV), malaria, dengue, pneumonia, influenza, leptospirosis and melioidosis. These diseases were chosen among the common infectious diseases in Thailand as the risk of transmission from these diseases are plausibly influenced by weather conditions [2–4,18,19]. There were no documented changes in case definitions for the diseases of interest between 2003 and 2019, and cases were reported based on clinical suspicion. We also chose a group of diseases which varied in transmission mode to investigate how the burdens of these diseases may be impacted by current and future extreme weather patterns.

Historical meteorological data

Meteorological data from 2003 to 2019 was obtained from ERA5, published by the European Centre for Medium-Range Weather Forecasts. ERA5 provided hourly estimates across a 30km grid, which we spatially averaged for each province in Thailand [20]. Air temperature at 2m, dewpoint temperature, total precipitation was collected. The hourly estimates were used to calculate daily minimum and maximum temperature and daily total precipitation.

Demographic data

Population size for each province from 2003 to 2019 was obtained from the Official Statistics Registration Systems of Thailand [21]. Similarly, as Bueng Khan was split from Nong Khai in 2011 to form a separate province, we merged the population from Bueng Khan back to Nong Khai from 2011 onwards, to allow for consistent analysis over the timeframe of the dataset. Annual population projections at 30-arc-seconds resolution under 4 climate change scenarios SSP126, SSP245, SSP370 and SSP585 from 2021 to 2100 were obtained, which were simulated based on the CMIP6 experimental design data and historical population from 2010 [22]. The projections were spatially aggregated to each population in Thailand. The annual projections were aggregated in 20-year periods to be consistent with the meteorological data projections. We obtained a second historical population dataset from WorldPop for each province from 2003 to 2019, which was the data used to train the CMIP6 population projection data [23]. This dataset was used exclusively to train the models in the sensitivity analysis that incorporated future population changes. For the main analysis, we used government census-based population data, which are more reliable for historical estimates but do not extend to future projections.

Extreme weather events

Extreme weather events were classified into extreme heat, extreme dry weather and extreme wet weather. An extreme heat day was defined as a daily maximum temperature above a threshold for at least 3 consecutive days. The threshold was set at the 90^th percentile of the national daily maximum temperature (34.9°C), with the reference period being 2003–2021. A single national threshold, rather than a region-specific threshold, was used to define extreme heat days to ensure comparability across regions. Furthermore, the 90^th percentile daily maximum temperature differed by only 1–2 °C across geographical regions, supporting the use of a national threshold. Daily precipitation was converted to monthly Standardised Precipitation Index-1 (SPI) values; which is commonly used as a drought index and can reflect precipitation anomalies [24]. Conversions used the R package SPEI, using 2003–2021 as the reference period [25]. We used SPI rather than Standardized Precipitation-Evapotranspiration Index (SPEI) due to the limited availability of projected potential evapotranspiration under future climate change scenarios. To test if SPEI and SPI categorised wet and dry months similarly, potential evapotranspiration was estimated for the historical period by using the Hargreaves equation [26] to calculate SPEI. SPEI-1 and SPI-1 categorised extreme dry, extreme wet and normal months identically in 81% of the data. Hence, the use of SPI rather than SPEI would not have significantly affected our results. An accumulation period of 1 month was chosen to account for short-term impacts. Months with SPI lower than -1.5 were characterised as extreme dry weather and months with SPI higher than 1.5 were characterised as extreme wet weather [27].

Meteorological data under climate change scenarios

Climate change projections were obtained from 3 general circulation models, MIROC6, CMCC-ESM2 and IPSL-CM6A-LR under 4 climate change scenarios, SSP126, SSP245, SSP370 and SSP585 [28–30]. We chose only 3 models as they covered a range of climate sensitivities and focus on different processes [28–30]. MIROC6, CMCC-ESM2 and IPSL-CM6A-LR represents low to moderate, moderate and high climate sensitivities respectively [24–26]. Projected monthly minimum temperature, maximum temperature and total precipitation were obtained across the periods of 2021–2040, 2041–2060, 2061–2080 and 2081–2100, at a 30-arc-second resolution. This means that the dataset provided the average projection of each calendar month over 20 years. MODAWEC, a daily weather generator, was used to disaggregate daily precipitation, daily maximum and daily minimum temperature from the respective monthly data [31]. The monthly number of wet days was also required as an input by MODAWEC, where a wet day is defined as a day with more than 1 mm of precipitation, which is consistent with standard definitions and climate studies. The future number of wet days were linearly interpolated based on historical daily precipitation data and number of wet days. Only the daily maximum temperature was used in subsequent analyses to determine the monthly number of extreme heat days.

Climate change scenarios

The Shared Socioeconomic Pathway (SSP) framework developed for the CMIP6 provides potential future scenarios based on socioeconomic factors and greenhouse gas concentrations [1]. In this study, the SSP126, SSP245, SSP370 and SSP585 scenarios were considered, which are summarised in Table 1 [32].

Download:

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

Table 1. Description of SSP scenarios.

https://doi.org/10.1371/journal.pntd.0013896.t001

Statistical model

Generalised additive models (GAMs) with a negative binomial distribution were used to model the associations between disease case counts and the covariates. The covariates include lagged monthly SPI, lagged number of monthly extreme heat days, population density and contemporaneous relative humidity. The negative binomial link function was chosen as it was the appropriate distribution to model the case counts which were overdispersed based on the Cameron and Trivedi test [33] (S1 Table). The models were trained on data from 2003 to 2019. GAM has been employed in other similar studies to project future disease incidence using meteorological variables and to study the association between disease incidence and these factors [34–36]. GAMs offer the advantage of flexibility by modelling predictors as smooth non-linear functions, as opposed to only linearly in generalised linear models (GLMs) [37]. This is particularly important given the evidence of non-linear associations between extreme weather and disease risk [35,36,38]. Spatial random effects were used to account for differences in transmission dynamics across provinces. For each province and disease, the following GAM specification was used:

(1)

where is the case count of disease at monthly time in province , is the intercept, is the logarithm of the province population at year which is used as an offset for the at-risk population. is the SPI up to lags, is the number of extreme heat days in a month up to lags, is the population density at year in province , is the contemporaneous relative humidity in province , and ,, and are the thin-plate splines for disease . Seasonal trends were accounted for using thin-plate splines on each calendar month for disease . Although the splines for calendar month were not explicitly specified as cyclic, the fitted monthly smooths naturally exhibited a cyclical pattern, appropriately capturing seasonality. The covariate lags and were chosen by the mean Akaike information criterion (AIC) across all models of a disease (S2 Table). Thin-plate splines were chosen as they produce a lower mean squared error than alternative splines [39]. A basis dimension of 10 was used for each spline which allowed for a balance of flexibility and parsimony. Spatial dependencies across provinces were accounted for using random effects for each province . Moran’s I was calculated across diseases and timepoints. As the observed spatial autocorrelation was very weak and largely statistically non-significant, we did not incorporate spatial autocorrelation in the model. Restricted maximum likelihood (REML) was used to estimate the splines rather than generalised cross-validation as the latter tends to favour more complex models and under-smooth terms [40].

For dengue and JEV, lags of up to 2–3 months for extreme heat and SPI reflect their effects on the extrinsic incubation period and vector reproduction, together with the delayed effects on vector abundance [41,42]. For malaria, the short heat lag of 1 month aligns with the effect of extreme heat on the extrinsic incubation period, while rainfall effects on vector habitats accumulate over a longer period of 3 months [43,44]. For pneumonia, leptospirosis and influenza, 3-month lags for both extreme heat and SPI capture delayed effects of temperature and moisture on transmission, exposure pathways and host susceptibility [45,46]. For pneumonia and influenza, the longer lags compared to the incubation periods may be the cumulative effects of extreme weather on respiratory immunity [45,47]. Melioidosis included an immediate heat effect and a 3-month SPI lag to capture delayed rainfall impacts on soil and bacterial exposure [18].

Estimation of incidence rate ratio and population attributable fraction

The historical effect of SPI and extreme heat days on diseases case counts were measured by the incidence rate ratio (IRR) and population attributable fraction (PAF). The IRR gives the ratio of predicted cases with exposure to predicted cases with no exposure. Specifically, the IRR for SPI gives the ratio between incidence rates at varying values of SPI and the incidence rate at 0 SPI. An SPI value of 0 represents average precipitation conditions based on the long-term mean precipitation in that location. The IRR for extreme heat gives the ratio between incidence rates at non-zero days of extreme heat in a month and the incidence rate with no days of extreme heat. In computation of IRRs, we only varied one covariate (SPI or extreme heat days) at some level while holding all other covariates at their historically observed mean value. The IRR was calculated as follows for each disease:

(2)

where is the estimated incidence rate of the disease in province , is the covariate of interest and are the remaining set of covariates set at the province-specific means .

In the case of computing national-level IRRs, province-level incidence rates and covariates in (2) are replaced by national-level incidence rates and covariates respectively.

The PAF gives the proportion of cases in a population that is attributed to a particular risk factor. A PAF of 0 indicates that the risk factor has no impact on the occurrence of the disease, and a negative PAF indicates that the factor has a protective effect against the disease. The province-specific PAF was estimated with the ratio of predicted cases when varying either SPI or extreme heat days from zero, with other covariates held constant, against predicted cases when SPI and extreme heat days are set to zero. The province-specific PAF was calculated as follows for each disease:

(3)

where is the estimated incidence rate of the disease in province , is the covariate of interest and are the remaining set of covariates set at the province-specific means . Similar to IRR, national-level PAF is calculated by replacing province-level incidence rates and covariates in (3) by national-level incidence rates and covariates respectively. 95% confidence intervals were calculated for both IRR and PAF using the delta method, which derives the asymptotic distribution of the IRR and PAF assuming incidence rates are asymptotically Gaussian.

Future projection and estimation of excess risk due to extreme weather scenarios

We calibrated the statistical models to province and disease specific data from 2003–2019. Thereafter, future monthly disease cases for each province, time period and climate change scenario were projected, based on the MIROC6 climate projections of total precipitation, relative humidity and maximum temperature. We chose MIROC6 as the main CMIP6 model as it was specifically designed to capture complex Asian monsoon systems and ocean-atmosphere interactions, which are crucial for modelling Thailand’s climate. The time periods considered were 2021–2040, 2041–2060, 2061–2080, 2081–2100 and the climate change scenarios considered were SSP126, SSP245, SSP370, SSP585. Projections were compared to a reference period of 2003–2019. Province population counts were set at the historical mean values. Population was assumed to remain at the historical level so that the excess risk could be solely attributed to extreme weather events.

To minimise noise, excess risk was estimated only based on timepoints where the respective extreme weather events were projected to take place. The main extreme weather exposures considered were (1) extreme heat, extreme dry and extreme wet weather, (2) extreme heat, (3) extreme dry weather and (4) extreme wet weather. Additionally, we considered sub-scenarios to verify that the impact of one extreme weather event was not confounded by another. For example, given that extreme heat is associated with extreme dry weather. the excess risk attributed to extreme heat may be due to extreme dry weather instead [48]. Hence, we considered the following sub-scenarios – (1) extreme heat and no extreme dry weather, (2) extreme heat and no extreme wet weather, (3) extreme heat with no extreme dry and no extreme wet weather, (4) extreme wet weather and no extreme heat, and (5) extreme dry weather and no extreme heat. The presence of extreme heat refers to one or more extreme heat days in a month. We also estimated the projected disease case counts and excess risk based on 2 alternative climate change models, CMCC-ESM2 and IPSL-CM6A-LR, for a sensitivity analysis. Using the disease-specific GAM models, the projected province-level incidence of each disease at each future time period and climate change scenario, and the corresponding fitted historical incidence is estimated as follows:

(4)

(5)

where is the estimated incidence rate of the disease in month of future time period and province under climate change scenario . include the covariates SPI, number of extreme heat days and relative humidity, and is the covariate of interest in month of future time period under climate change scenario in province . is the lagged monthly extreme heat days of up to lags, while is the lagged monthly extreme heat days of up to lags. The number of lags are shown in S2 Table. As climate projections only provide multi-year monthly averages rather than the continuous monthly time series, lagged covariates are treated as cyclical. For example, for disease projections in February, 3 months of lagged SPI would be taken from SPI in January, December and November. are the remaining covariates and are the mean of the covariates in month over the historical time period 2003–2019 in province . The fitted historical incidence is estimated based on , the mean of all covariates in month of the historical time period 2003–2019 in province . 95% prediction intervals were calculated for the excess risk estimates by using the 95% prediction intervals of the projected disease case counts.

Annual province-level ER attributable to extreme weather is calculated using the incidences obtained from (4) and (5), aggregated over months where extreme weather is projected to occur. For each month, the projected incidence under future weather conditions was compared with the expected incidence under historical average conditions for that same month. This ensures that excess risk reflects the impact of future extreme weather while holding all other covariates constant.

(6)

(7)

where are the months where extreme weather condition occurs in future time period under climate change scenario in province , which includes months with all extreme weather events, extreme dry weather, extreme wet weather and extreme heat respectively. The extreme weather conditions are explicitly defined in (7). This restricts our estimation of ER to include only incidence in months which extreme weather takes place.

Annual national-level ER is calculated by modifying (6) such that incidences are aggregated over all provinces. This represents the percentage change in projected nationwide incidence under extreme weather conditions compared to the expected nationwide incidence under historical average weather.

Model evaluation

For each disease, the Akaike information criterion (AIC) of models with different variable lags were compared, where the model with the lowest AIC was selected for the rest of the analysis. Model fit was evaluated using AIC as it balances model fit with complexity and is more suitable for long-term projections than time series cross-validation approaches, which are intended for short-term forecasts. The in-sample fit of the models were assessed by AIC and root mean squared error (RMSE). For each disease, the AIC and RMSE of the GAM model were compared against corresponding generalised linear models (GLMs) with the same covariates and lags, as chosen by the optimal AIC previously, without the smooth function.

As part of the sensitivity analysis, we also developed models which included future population as an offset and future population density as a predictor instead of historical population. We compared the projections of the models with future population against those using historical population for each disease. Additionally, we trained the models using three different sets of global climate model (GCM) projections, MIROC6, CMCC-ESM2 and IPSL-CM6A-LR, and compared the national-level excess risks attributable to extreme weather derived from each dataset across time and climate change scenarios. This ensured that our results were not heavily influenced by the choice of GCM used for our main results, which was MIROC6. Furthermore, we utilised the delta method to compute the confidence intervals of our IRR and PAF estimates, allowing us to appropriately characterise the uncertainty of the association between extreme weather and disease incidence rate.

Future climate change scenarios

In this study, we considered 4 of the CMIP6 climate change scenarios - SSP126, SSP245, SSP370, SSP585, which take into account a wide range of socioeconomic trends and radiative forcings (Table 1). Temperatures are expected to increase across SSPs, where SSP585 has the largest increase in extreme heat days from 2021 to 2100 across all provinces (Fig 1). A similar trend is observed on a provincial scale (S1 Fig). The national level precipitation remains largely the same across SSPs over time, with similar minimum and maximum monthly SPI values among all the SSPs (Fig 1). From 2021 to 2100, extreme SPI values are only observed at the province level (S1 Fig).

Download:

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

Fig 1. Historical and projected monthly SPI and extreme heat days. (A) Monthly SPI and (B) monthly extreme heat days from 2003 to 2021 and during the periods 2021–2040, 2041–2060, 2061–2080, 2081–2100 with climate change scenarios SSP126, SSP245, SSP370 and SSP585 in Thailand. Climate projections are based on the MIROC6 GCM.

https://doi.org/10.1371/journal.pntd.0013896.g001

Association of diseases with extreme weather

We developed generalised additive models for province-level disease cases and trained them on extreme weather indicators from 2003–2021 to obtain the relationship between extreme weather and disease case counts. We expressed the relationship between extreme weather and disease incidence in incidence rate ratios (IRR) and population attributable fractions (PAF). IRR can be interpreted as the ratio change in incidence rates given exposure to extreme weather in the same month compared to the baseline (0 heat days and 0 SPI respectively), while PAF can be interpreted as the proportion of cases in a population that could be avoided in the absence of anomalous SPI and extreme heat in the same month.

In dengue, JEV, influenza, pneumonia and melioidosis, the number of extreme heat days in a month was positively associated with disease risk (Fig 2). On the other hand, increasing days of extreme heat days in a month was negatively associated with malaria and leptospirosis risk (Fig 2). The pattern was also consistent when the association of disease incidence against extreme weather was broken down on a province-specific basis, particularly in highly populated provinces like Bangkok, Nakhon Ratchasima and Ubon Ratchathani (S2–S10 Figs). Province-level PAFs demonstrated the same trend, with positive values signifying that a proportion of cases in the population can be attributed to extreme heat, extreme dry weather and extreme wet weather respectively (S11–S17 Figs).

Download:

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

Fig 2. National level incidence rate ratio (IRR) and population attributable fraction (PAF) over days of extreme heat and standardised precipitation index (SPI) for dengue, JEV, influenza, malaria, pneumonia, leptospirosis and melioidosis.

Bars represent 95% confidence intervals and red bars represent statistically significant points.

https://doi.org/10.1371/journal.pntd.0013896.g002

Extreme wet weather (i.e., SPI 1.5) was historically associated with lower incidence of dengue and higher incidence of JEV, pneumonia and melioidosis. At SPI levels of 3, dengue IRR was 0.66 (95% CI: 0.62, 0.69), while JEV IRR was 1.29 (95% CI: 1.19, 1.40). Extreme dry weather (i.e., SPI -1.5) was historically associated with higher incidence of JEV and lower incidence of dengue and pneumonia (Fig 2). At SPI levels of3, dengue IRR was 0.88 (95% CI: 0.83, 0.92), while JEV IRR was 1.21 (95% CI: 1.12, 1.30). Province-level IRRs and PAFs were concordant with these trends (S2-S24 Figs). Lag-exposure curves for SPI and extreme heat days can be found in S25 Fig.

Overall change in disease risk during periods of extreme weather

Using the calibrated model, we projected the excess risk of infectious diseases in future climate change scenarios based on projections of extreme heat and rainfall from 2021 to 2100. Excess risk describes the changes in incidence rates with reference to the historical baseline from 2003 to 2020.

Pneumonia and melioidosis risks are expected to increase in the future during periods of extreme weather, particularly in the highest carbon emission scenario, SSP585 respectively (Figs 3D, 2L). Influenza risks during periods of extreme weather are highest in the moderate scenario SSP245, specifically in the first half of the century (Fig 3B). In other scenarios, there is conversely a decrease in influenza risk during periods of extreme weather (Figs 3A, 2C, 2D). A decrease in JEV risk from 2061–2080 is observed across all four climate change scenarios, except for SSP585, where there is an expected increased risk—though the excess risk still follows the same trend as other scenarios (Fig 3E-3H).

Download:

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

Fig 3. Excess risk of pneumonia, influenza, JEV, malaria, dengue, melioidosis and leptospirosis during periods of extreme weather, across 2021 to 2100, in climate change scenarios SSP126, SSP245, SSP370 and SSP585.

Excess risk represents the percentage change in annual disease case counts from the historical baseline from 2003 to 2020.

https://doi.org/10.1371/journal.pntd.0013896.g003

An increase in malaria risk is expected during periods of extreme weather across all climate change scenarios and time periods (Fig 3I-3L). The exception is in SSP245, which sees a decrease in 2021–2040 and 2081–2100 (Fig 3J). Conversely, a decrease in dengue risk is expected during periods of extreme weather, across all climate change scenarios and time periods, apart from SSP245 (Fig 3E-3H). Under SSP245, an increase in dengue risk is expected from 2021–2040, which is followed by an expected steep decrease in the rest of the century (Fig 3F). Similar to dengue, leptospirosis risk is expected to decrease during periods of extreme weather, across all climate change scenarios and time periods.

Change in disease risk by individual extreme weather events

We stratified the projected change in disease risk by the type of extreme event – extreme heat, extreme dry weather and extreme wet weather. A decrease in dengue is expected during periods of extreme heat under all climate change scenarios other than SSP245 (Fig 4A). In SSP245, a large increase of 26.2% (95% CI: 0.90%, 51.7%) is expected from 2021–2040, which then declines and increases slightly again in 2081–2100 (Fig 4A). The decrease in dengue risk during extreme heat generally contradicts the historical association between extreme heat and increased dengue risk. We found that changes in dengue risk were primarily driven by future relative humidity. When relative humidity was held constant between the baseline and projection scenarios, dengue excess risk was small and statistically insignificant across all SSPs and time points (S26 Fig). This suggests that the projected decrease in dengue risk was likely driven by lower levels of relative humidity, which have been historically associated with reduced dengue risk (S27 Fig). Among other diseases, relative humidity was not found to drive any changes in excess risk during periods of extreme heat. A large decrease in dengue is expected in periods of extreme dry and wet weather across all climate change scenarios (Figs 4B, 3C). Extreme dry and wet weather was historically associated with decreases in dengue risk (Fig 2). The expected decrease in dengue risk during extreme dry weather ranges from -52.1% (95% CI: -63.2%, -41.0%) in 2081–2100 under SSP370 to -25.4% (95% CI: -41.0%, -9.77%) in 2021–2040 under SSP245 (Fig 4B). Dengue risk during extreme wet weather is expected to decrease the most by 52.9% (95% CI: 43.0%, 62.8%) in 2041–2060 under SSP370 and decreases the least by 35.0% (95% CI: 21.1%, 48.7%) in 2021–2040 under SSP245 (Fig 4C). For other diseases, this effect was more modest and statistically insignificant (Fig 4).

Download:

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

Fig 4. Excess risk (ER) of JEV, malaria and dengue during periods of extreme heat, extreme dry weather and extreme wet weather, across 2021 to 2100, in climate change scenarios SSP126, SSP245, SSP370 and SSP585.

Excess risk represents the percentage change in annual disease case counts from the historical baseline from 2003 to 2020.

https://doi.org/10.1371/journal.pntd.0013896.g004

We expect JEV risk to decrease during periods of extreme dry and extreme wet weather (Fig 4E, 4F). The greatest decline is expected during periods of extreme dry weather in 2081–2100 under SSP245 (ER: -18.2%, 95% CI: -43.4%, 6.96%) (Fig 4E). Conversely, malaria risk is expected to increase during periods of extreme dry weather and wet weather, but more so during extreme dry weather (Fig 4H, 4I). This is given that there was a historical association between extreme dry weather and increased malaria risk (Fig 2).

Influenza risk is expected to increase during periods of extreme heat only under SSP245, which peaks in 2021–2040 (ER: 9.74%, 95% CI: -11.8%, 31.3%) (Fig 5A). Historically, having over 8 days of extreme heat in a month is associated with increased influenza risk (Fig 2). During periods of extreme dry weather, influenza risk is expected to decrease, particularly in SSP245 and SSP370 (Fig 5B). Under SSP245, influenza risk is expected to decrease during periods of extreme dry weather to the greatest extent from 2061–2080 (ER: -18.9%, 95% CI: -35.4%, -2.33%) (Fig 5B). The risk of influenza is expected to decrease more during periods of extreme wet weather than during periods of extreme dry weather (Fig 5B, 5C). The greatest expected decrease during periods of extreme wet weather can be seen under SSP370, which decreases by -37.8% (95% CI: -50.8%, -24.8%) from 2061–2080 (Fig 5C). This relationship is also seen historically, where extreme wet weather has a negative association with influenza risk (Fig 2).

Download:

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

Fig 5. Excess risk (ER) of influenza, pneumonia, leptospirosis and melioidosis during periods of extreme heat, extreme dry weather and extreme wet weather, across 2021 to 2100, in climate change scenarios SSP126, SSP245, SSP370 and SSP585.

Excess risk represents the percentage change in annual disease case counts from the historical baseline from 2003 to 2020.

https://doi.org/10.1371/journal.pntd.0013896.g005

The risk of melioidosis is projected to increase on extreme heat days, especially under higher carbon emission scenarios like SSP370 and SSP585 (Fig 5J). SSP585 shows the most pronounced increase over time, although the confidence intervals remain wide (Fig 5J). By 2100, melioidosis risk increases by 27.8% (95% CI: -39.4%, 94.9%) during periods of extreme heat (Fig 5J). A large increase in melioidosis is expected during periods of extreme wet weather, although confidence intervals are wide (Fig 5L). Broad confidence intervals could be explained by low monthly case counts across provinces. The greatest increase is expected under SSP585, from 2041–2060 (ER: 25.4%, 95% CI: -32.1%, 82.8%) (Fig 5L).

Trends of dengue excess risk during extreme weather in SSP245

We used the calibrated model to estimate province-level changes in incidence rate during periods of extreme weather compared to the baseline of 2003–2019, which included periods of extreme heat, extreme dry weather and extreme wet weather. Province-level excess risks of all diseases during periods of extreme weather across climate change scenarios can be found in S28-S34 Figs, where it can be seen that influenza and dengue have the most statistically significant province-level excess risks. We conducted a finer analysis on the change in dengue risk during periods of extreme weather in SSP245, which showed a significant increase in the first half of the century (Fig 3F). Moreover, SSP245 is generally considered the most likely scenario [49]. We focused on the excess risks of five provinces which had the highest dengue burden from 2003 to 2019 (Fig 6A).

Download:

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

Fig 6. (A) Excess risk of the five provinces with the historically highest incidence of dengue during periods of extreme weather, under SSP245 from 2021–2100. (B) Province-level excess risks of dengue under SSP245 from 2021–2100. Asterisks indicate a statistically significant excess risk value. Excess risk represents the percentage change in annual disease case counts from the historical baseline from 2003 to 2020. Map created using GADM data (https://gadm.org/index.html, freely available for academic use). The map outlines and administrative boundaries are used with permission for academic publishing.

https://doi.org/10.1371/journal.pntd.0013896.g006

From 2021–2040, dengue risk is expected to increase in most provinces during periods of extreme weather (Fig 6B). Provinces with historically high dengue incidence, such as Bangkok, are expected to have an increased risk of dengue by 42.2% (95% CI: 13.6%, 70.7%) during periods of extreme weather (Fig 6A). On the other hand, there were also provinces which are expected to experience a decrease in dengue risk during extreme weather in this time, such as Chiang Mai (ER: -24.6%, 95% CI: -40.1%, -9.00%) (Fig 6A). From 2041–2060, a general decrease in excess risk is seen, particularly in the Southern region (Fig 6A, 6B). In contrast, provinces in central Thailand are expected to have increase in dengue risk (Fig 6B).

From 2061–2080, dengue risk during extreme weather are expected to decrease and reach its lowest point in most provinces (Fig 6A, 6B). A notable decrease in risk during periods of extreme weather is expected in Chiang Rai (ER: -34.2%, 95% CI: -47.2%, -21.1%) (Fig 6A). From 2081–2100, dengue risk remains moderately low among most provinces during periods of extreme weather (Fig 6A, 6B).

The increased dengue in central Thailand from 2021–2060 is likely to be driven by moderate levels of extreme heat in that period (Figs 4A-4C, S1, S35). Moreover, in the southern provinces with lower levels of extreme heat, no increase is expected in this period (S1 Fig). In the second half of the century, higher levels of extreme dry and extreme wet weather is expected to drive decreases in dengue across most of the provinces (Figs 4A-4C, S1). This aligns with our observation that extreme wet or dry weather was historically associated with a lower incidence rate of dengue (Fig 2).

Trends of influenza excess risk during extreme weather in SSP245

Similarly, we narrowed our analysis to the change in influenza risk during periods of extreme weather in SSP245, which is expected to experience significant changes at the province level over the century (Fig 7A, 7B).

Download:

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

Fig 7. (A) Excess risk of the five provinces with the historically highest incidence of influenza during periods of extreme weather, under SSP245 from 2021–2100. (B) Province-level excess risks of influenza under SSP245 from 2021–2100. Asterisks indicate a statistically significant excess risk value. Excess risk represents the percentage change in annual disease case counts from the historical baseline from 2003 to 2020. Map created using GADM data (https://gadm.org/index.html, freely available for academic use). The map outlines and administrative boundaries are used with permission for academic publishing.

https://doi.org/10.1371/journal.pntd.0013896.g007

From 2021–2040, an increase in influenza is expected in most provinces (Fig 7B). The heterogenous increase across provinces is likely to be driven by moderate levels of extreme heat and extreme dry weather in the area (Figs 5A-5C, S1). Influenza risk is expected to peak in 2041– 2060, where most provinces is expected to have an increased risk during periods of extreme weather, particularly in Northeastern and Central Thailand. Nakhon Ratchasima, a province with historically high influenza burden, is expected to have an excess risk of 36.8% (95% CI: 9.83%, 63.8%) during periods of extreme weather (Fig 7A). This is likely driven by increasing levels of extreme heat in those regions, which was historically associated with increased influenza risk (Figs 2, S1, S36). The lowest levels of influenza excess risk during extreme weather across the five highlighted provinces are seen from 2041–2060, particularly in Southern Thailand (Fig 7B). From 2081– 2100, influenza risks across the five highlighted provinces will reach a generally moderate level where about half of provinces are expected to experience an increase in influenza risk during extreme weather periods while the other half are expected to experience a decrease (Fig 7B). The lower levels of influenza, particularly in Southern Thailand, are likely to be driven by extreme wet weather which was historically associated with lower influenza risk (Figs 2, 5A–5C, S1).

Model evaluation

We compared the AIC of the GAM models to their corresponding GLM models, which used the same covariates without the smooth function. The AIC of each GAM model was lower than its corresponding GLM model in all 7 models (S3 Table). This showed that the incorporation of smooth functions to model predictors improved model fit in almost all cases.

We evaluated an alternative GAM model which included future population as an offset and future population density as a predictor instead of current population. Predictions were heavily skewed by future population projections (S37–S43 Figs). This was most likely due to models being unable to handle population projections which exceed current observations. Moreover, we wanted to observe the change in risk solely from extreme weather and not population levels. Hence, in this study, we assumed that population would remain at current levels based on data from 2003 to 2021.

Sensitivity analysis

We estimated the excess risks attributed to extreme weather events based on 2 other global climate models (GCMs), CMCC-ESM2 and IPSL-CM6A-LR to capture the uncertainty inherent in the climate projections and test the robustness of our findings. We also estimated excess risks based on a multi-model GCM ensemble, calculated as the mean of MIROC6, CMCC-ESM2 and IPSL-CM6A-LR outputs. Across all the diseases studied, the general trends of excess risk attributable to each extreme weather scenario based on the alternative GCMs were similar to the main projections in this study (S44-S50 Figs). We evaluated the fit of GAM models without including relative humidity as a predictor across provinces and diseases to determine if inclusion of relative humidity would enhance model fit. Including relative humidity improved the fit of all 7 models, given the AIC of all models with relative humidity was lower than the AIC of models without relative humidity (S4 Table). Moreover, we evaluated the fit of GAM models using a cubic regression spline instead of thin plate regression spline across provinces and diseases and found that the difference in spline used did not change model fit in most models (S5 Table).

We also estimated the excess risk attributable to extreme weather sub-scenarios while controlling for the presence or absence of other concurrent extreme weather conditions. Certain extreme weather events, such as extreme heat and extreme dry weather, may tend to take place simultaneously [48]. Thus, we explored if the excess risks atrributable to one type of extreme weather might actually reflect combined effects.

We first compared excess risk attributable to extreme heat against sub-scenarios combining extreme heat and extreme dry weather, extreme wet weather or normal conditions. Across diseases, the excess risks attributable to extreme heat were similar to excess risks attributable to extreme heat with no extreme dry or wet weather (S51 Fig). This suggests that excess risks attributable to extreme heat estimated were not influenced by other extreme weather conditions.

Additionally, we compared the excess risk attributable to extreme dry weather with those under sub-scenarios involving extreme dry weather with or without extreme heat respectively. Excess risks attributable to extreme dry weather resembled the excess risks attributable to extreme dry weather with no extreme heat (S52 Fig). This indicates that the excess risks attributable to extreme dry weather were not confounded by the effects of extreme heat.

Lastly, we compared the excess risk attributable to extreme wet weather with those under sub-scenarios involving extreme wet weather with or without extreme heat respectively. The excess risks were likewise consistent, suggesting that the excess risks attributable to extreme wet weather were similarly not confounded by the effects of extreme heat (S53 Fig).