key: cord-0868400-dzdxv7pc authors: Yip, S.; Che Him, N.; Jamil, N. I.; He, D.; Sahu, S. K. title: Spatio-temporal detection of dengue outbreaks for the Malaysian Central Region using meteorological drivers at mesoscale and synoptic scale date: 2021-09-23 journal: nan DOI: 10.1101/2021.09.22.21263997 sha: ce7d3c667f03c35fb5969ab9309e1db7cdb58541 doc_id: 868400 cord_uid: dzdxv7pc Background: The disease dengue is associated with both mesoscale and synoptic scale meteorology. Previous studies of this disease, especially for south-east Asia, have found very limited association between synoptic meteorological variables and the number of dengue hospitalisations. However, to tackle future severe outbreaks and to institute an early warning system, it will immensely beneficial to find and establish more clear association with rate of dengue hospitalisations and the most relevant meteorological variables. Objectives: A rigorous Bayesian modelling framework is provided to identify the most important meteorological covariates and their lagged effects for developing an early warning system for dengue outbreaks in the Central Region of Malaysia. Method: We obtain the dengue hospitalisations count data and the demographic information in the Central Region of Malaysia. Along with other mesoscale environmental measurements such as local temperature, precipitation and ozone concentration level, we also examine multiple synoptic scale Niño indices which are related to the phenomenon of El Niño Southern Oscillation and an unobserved meteorological variable derived from reanalysis data. A probabilistic early warning system is built based on a Bayesian spatio-temporal hierarchical model with a physically interpretable complex structure. Results: Our study finds a 46.87% of increase in dengue hospitalisations due to one degree increase in the sea surface temperature anomalies in the central equatorial Pacific region with a lag time of six weeks. We also discover the existence of a mild association between the rate of cases and a distant lagged cooling effect of 28 weeks related to a phenomenon called El Niño Modoki. The proposed model also observes significant amount of association with some other meteorological parameters with disease rates. These associations are assessed by using an optimal Bayesian spatio-temporal model that outperforms other candidate models in terms of estimated out-of-sample predictive accuracy and performance in correctly issuing the warnings early. Discussion: With the novel spatial dynamic early warning system presented, our results show that the synoptic meteorological drivers can enhance short-term detection of dengue outbreaks and these can also potentially be used to provide longer-term forecasts in the Indian Ocean and regional seas in the Maritime Continent, the impact on the winter 106 rainfall during conventional El Niño in boreal winter season over Peninsular Malaysia is 107 minimal but significant higher during El Niño Modoki. Tangang et al. (2017) show that, 108 during winter, a strong La Niña leads to a significant decrease in wet precipitation extremes 109 over the Peninsular Malaysia due to the anomalous cyclonic circulation over strong La Niña. trend of both Niño1+2 and Niño4 indices during El Niño (Fig. 3) . Taking out the effect of 175 Niño4, the partial correlation between DIR and Niño1+2 index is 0.0578 only although the 176 Niño1+2 and Niño4 are highly correlated (Fig. 4) . These indices, refer to distant regions Without explicit spatial and temporal dependent terms, consider a Poisson generalised linear model for the disease count Y kt defined in Section 2.2 of the form: where Y kt is the expected number of cases in the district k at time t, e kt is the population persion is quite high. Such a statistical property can be easily described by a person is more 234 likely to be infected by disease through close contacts. It appears that the Poisson distribu-235 tion is better suited to explain the "number of infected groups" rather than the total disease 236 8 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. individually specified conditional to other parameters and data. In this Section, we will go 243 through the key components of our candidate models. To overcome overdispersion, we use the negative binomial parametrisation in which introduces r as a universal control parameter for overdispersion (Gelman et al., 1995) . The mean and variance of the random variable are E[Y kt ] = µ kt and Var[Y kt ] = µ kt + µ 2 kt /r. As r goes to infinity, the distribution of Y kt converges to the Poisson distribution. The Besag-York-Mollié model (BYM; Besag, York and Mollié, 1991; Besag and Kooperberg, 1995) specifies the additive relationship of the overall risk level as an intercept, the fixed effect by the covariates, the pure random effect θ kt and the spatial variation component φ k : where θ kt is a normally distributed unstructured error and φ k is the structured error modelled by an intrinsic conditionally autoregressive model (ICAR). It has a conditional specification that is normally distributed with a mean equal to the average of its neighbours (φ k∼j ) and its variance decreased as the number of neighbours d k increases: correlation parameter ranging from a full spatial neighbourhood dependent variation and pure residual randomness in which the terms φ k and θ kt combined to one entity φ kt : where the overdispersion parameter follows a Gamma distribution with hyperparameters a 253 and b, α is the autoregressive parameter to control temporal dependency between adjacent 254 time points, ω kt is the Gaussian distributed evolution error. Initial information is required 255 for this temporal structure, s is the scaling parameter controls the proportion of a spatial 256 and non-spatial variation, W is the neighbourhood information formulated as a connected We consider five models with different levels of complexity (Table 1 ). The regression part of the model x kt β is specified by the following setup: Temp k,t−3 + Rain k,t−10 + Ozone k,t−7 + omega t−15 + Niño12 t−28 + Niño4 t−6 + Niño4 t−10 + Capital k . preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. .0 (C) NB + spatial log λ kt = β 0 + x kt β + φ k 33389.0 ± 180.9 20.6 (D) NB + dynamic log λ kt = α log λ k,(t−1) + x kt β 29180.6 ± 181.6 1053.7 (E) NB + spatial + dynamic log λ kt = α log λ k,(t−1) + x kt β + φ k 29180.9 ± 182.5 1056.6 Although models A, B and C possess lower looic value, compared to dynamic models, they 286 preserve a considerable explanatory power. Taking a closer look at the coefficient estimates 287 11 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this this version posted September 23, 2021. ; https://doi.org/10.1101/2021.09.22.21263997 doi: medRxiv preprint of Model C, the coefficient estimates in the form of relative risk (RR) is shown in Table 288 2. The covariate lag (Niño4, 6) and lag (Niño4, 10) have a strong positive relationship with 289 the disease, for each degree increase of the indices, the RRs increase by 46.87% and 8.44% preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. appears to be a more preferable model after evaluating the overall performance measures. 338 We found that the most important RR comes from Niño4. It makes a longer-term pre- preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this this version posted September 23, 2021. ; This paper presents a Bayesian spatio-temporal modelling framework leading to a full imple-349 mentation of an EWS for dengue outbreaks from upstream data source to production. We preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this this version posted September 23, 2021. ; https://doi.org/10.1101/2021.09.22.21263997 doi: medRxiv preprint claim. Due to data limitations, the impact from spatio-temporal variations of virus serotype are 376 missing from the study. An anomalous upsurge happens twice in our study period, the first 377 one occurred in the 2013 summer is verified by microbiology evidence (Ng et al., 2015) . The preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. 24 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. 25 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Week 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 Year Week 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 Year 26 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this this version posted September 23, 2021. 27 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this this version posted September 23, 2021. ; https://doi.org/10.1101/2021.09.22.21263997 doi: medRxiv preprint All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this this version posted September 23, 2021. Week 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50 Year Bayesian measures 622 of model complexity and fit Climate and non-climate drivers of dengue epi-625 demics in southern coastal Ecuador Spatio-temporal analysis of the main dengue vector populations in 628 Formation of an air pollution index Characteristics of precipitation extremes in Malaysia Correlation 635 analysis of air pollutant index levels and dengue cases across five different zones in Selangor Practical Bayesian model evaluation using 638 leave-one-out cross-validation and WAIC Research note 640 repellency effects of an ozone-producing air purifier against medically important insect 641 vectors Bayesian forecasting and dynamic models A climate model for predicting 645 the abundance of Aedes mosquitoes in Hong Kong A simple, coherent framework 648 for partitioning uncertainty in climate predictions Figure 6 : Mean weekly DIR, temperature, rainfall, ground-level ozone concentration level from 2013 to 2019 by district.