Three machine-learning models to predict dengue vectors Breteau Index based on meteorology and biotope in Fujian China

doi:10.21203/rs.3.rs-4286413/v1

Background: Dengue fever's rising prevalence in China underscores the need for improved surveillance tools. The Breteau Index (BI), critical for tracking dengue transmission, currently lacks timely and precise predictions, hindering effective response strategies. This study proposes a machine learning-enhanced BI predictive model to refine dengue forecasting in Fujian China.

Methods: Data Collection: In this study, data about the Breteau Index (BI), meteorological conditions, and biotope characteristics from 2015 to 2022 were systematically gathered from the Siming District in Xiamen. Model Development: A range of machine learning algorithms was employed for BI prediction, including Support Vector Regression (SVR), Step-down Linear Regression, Random Forest (RF), Decision Trees (DT), Logistic Regression Models, Deep Neural Networks (DNN), LASSO Logistic Models, and Generalized Additive Models (GAM). Evaluation Metrics: The effectiveness and fit of the predictive models were quantitatively evaluated using the Akaike Information Criterion (AIC)/Bayesian Information Criterion (BIC) and R-squared values, alongside autocorrelation analyses of residuals to ensure statistical integrity. External Validation: Models were further validated with BI data from other Fujian cities to test their generalizability.

Results: This study highlights the Deep Neural Network (DNN) model's superiority over alternative forecasting methods for the Breteau Index (BI), with significant determinants being stagnant water sources like unused containers and aquatic plant bonsais. Meteorological conditions, including precipitation(7-day lag), wind speed(3, 5, 7-day lags), sunshine(2-day lag), barometric pressure(3, 4-day lags), humidity, and temperature(7-day lag), were identified as critical influencers of BI. The analysis underscores the necessity of prompt waterlogged area management post-rainfall to mitigate dengue vector breeding, aligning with findings on the ecological preferences of Aedes albopictus.

Conclusions: In conclusion, the validated Deep Neural Network (DNN) model efficiently predicts the Breteau Index (BI) in Xiamen, indicating its utility in Fujian. Key strategies include controlling waterlogged containers to reduce Aedes aegypti habitats and promptly addressing weather-related BI fluctuations within three days post-rainstorm to lower dengue transmission risk.

Machine learning

Breteau Index

Meteorology

Biotope

Globally, dengue remains a major public health concern and problem, a virus transmitted by arthropods[1]. This is an arbovirus infection that mainly affects tropical and subtropical populations. A wide variety of dengue viruses spread in the Americas, the Eastern Mediterranean, the Western Pacific, and Southeast Asia. According to The World Health Organization, it is underestimated that between 50 to 100 million individuals are infected with the virus, resulting in almost 10,000 deaths[2, 3]. Aedes albopictus, as an important vector of infectious diseases, can transmit at least 26 viruses, including yellow fever viruses, alphaviruses, and Bunyaviruses[4], with dengue, chikungunya and Zika viruses being the most representative ones[5, 6]. While there is no specific treatment or vaccine available for treating DF or DHF, vector control measures remain the standard tool for preventing and controlling these diseases.

For the determination of the transmission of the dengue virus in the immature stages of the mosquito population[7], conventional indices such as container indexes (CI), house indexes (HI), and Breteau indexes (BIs) are often used as statistical indices[8]. As one of the most widely used indices among these, the BI index, which measures the number of positive containers (containing larvae of Aedes albopictus) per 100 houses inspected, plays a significant role in larval surveillance since it is more accurate in predicting mosquito intensity in suspected areas even in areas with low Aedes infestation[9]. However, accurately obtaining these data is often time-consuming, labor-intensive, and even hard in some regions, making it difficult to respond quickly.

Numerous methods for the rapid prediction of Aedes aegypti population density currently exist, yet they exhibit varying degrees of shortcomings in accounting for complex ecosystem interactions, the uncertainties associated with climate change, and the generalization capabilities of models. In O. Mudele’s work, a Poisson GLM is used to model and extrapolate Ae. aegypti eggs count based on relevant environmental variables (precipitation, temperature, humidity, and vegetation condition) extracted from RS data[10]. D A. Focks also employs dynamic life table methods to predict the dynamic changes of Ae. aegypti across different life stages, yet fail to capture the stochasticity and complexity inherent in natural processes[11]. Similarly, the construction of stochastic population dynamics models by Otero M, which predict based on temperature variations, also struggles to account for the complexity of environmental changes, and the parameters involved carry uncertainties[12]. Given the intricate and nonlinear relationships between climate, habitat, and larval density, traditional methodologies falter in addressing complex "black box" issues.

In contrast, machine learning can handle multiple variables and their interactions, allowing for customized, real-time predictions and updates tailored to specific regional conditions. With the incorporation of local new data, machine learning models can learn and improve autonomously. RF has been applied to monitor the intensity and distribution of adult mosquitoes to identify variables that influence the distribution and abundance of three mosquito malaria vectors and their relative importance[13]. Salim et al used dengue outbreak viable based on historical data, combined with meteorology viable, and constructed a Support Vector Machine (SVM) model, which is of high accuracy to predict the strike of dengue while influenced by unbalanced data[14]. Another complementary study combined a high-dimensional multidisciplinary dataset with the ML techniques of SVM, Gradient Boosting Machine (GBM), and RF[15].

Despite these advancements, an accurate machine-learning approach has yet to be developed and is presently constrained to meteorological factors, limiting its applicability to other regions. Consequently, additional methods, such as the implementation of Deep Neural Networks (DNN) to identify significant features, are deemed necessary[16]. This article concentrates on various methods such as the Logistics model, RF, DT, SVM, Logistic Models, and DNN for model construction. These methods will be compared to develop an optimal model to predict BI, which assists the government in preparing necessities and issuing warnings.

2.1 Data collection and preparation

2.1.1 Biotope data

The dataset, provided by the CDC, encompasses vector metrics from Siming District, Xiamen, between 2015 and 2022, including geographic coordinates and the BI. Biotope data details stagnant water locations, such as in bonsai, cisterns, tanks, basins, and discarded containers, among others. specific biotope variables are operationalized as follows:

tanks: cistern water jars, tanks, and basins with stagnant water
bonsai: bonsai aquatic plants with stagnant water
other: other water bodies with stagnant water
tires: used tires with stagnant water

2.1.2 Meteorological data

Even though the Aedes aegypti mosquito is commonly found in tropical and sub-tropical areas, its distribution is modulated by environmental factors[17]. Meteorological data includes the maximum and minimum temperatures (℃), mean temperature (℃), sunshine hour (h), precipitation within 20 hours, mean relative humidity, mean pressure (hPa), and mean wind speed. Specific meteorological variables are operationalized as follows:

pressureX: mean pressure lags before X days (hPa)
precipitationX: 20-20 hour precipitation lags before X days (mm)
windX: mean wind speed lags before X days (m/s)
sunshine_hourX: sunshine hour lags before X days
humidityX: mean relative humidity lags before X days (%)
tem meanX: mean temperature lags before X days (°C)
tem highX: maximum temperature lags before X days (°C)
tem high2: maximum temperature lags before two days (°C)
humidity: mean relative humidity(%)
wind: mean wind speed (m/s)

It is worth noting that considering the breeding cycle of Aedes aegypti ranged from 1.22 to 2.49 days[18], we selected the meteorological factors that lagged 7 days before the Breteau index as variables. A total of 78 variables were selected, including 10 habitat variables, 64 meteorological variables and 4 seasonal variables.

2.2 Statistical analyses

2.2.1 Modeling

The model construction process follows two steps in Fig 1. Step 1, the study initially explores the correlation between the BI and habitat characteristics through a comprehensive analysis of risk factors. In order to improve the generalization ability and robustness of the model, simplify the complexity of the model, and the quantity and quality of variables are taken into account, this study adopts machine learning to select the variables, finally selecting variables whose importance ≥ 0.01 as a threshold to be reintroduced into the regression model. Step 2, after selecting the appropriate variables, this article applies ordinary least squares, stepwise regression, and least squares to thoroughly analyze potential influential factors.

The integration of machine learning in the variable screening process serves a dual purpose: first, it streamlines the model, enhancing its generalization capabilities and mitigating the risks associated with overfitting. In addition, it addresses the challenges posed by multicollinearity, thereby improving the model's explanatory power in reflecting real-world scenarios. Notably, the exclusion of other spatial factors is deliberate, driven by the considerable variability of the Breteau Index within a confined range. Additionally, the absence of reliable data sources for generating high-precision maps is a key consideration, as incorporating such data could potentially lead to overfitting and undermine our ability to account for the influential role of core habitat factors.

2.2.2 Prediction

The BI is an important indicator for mosquito vector density surveillance, and WHO recommends a BI of ≤5 as the risk threshold for epidemics[16]. Based both on recent and historic studies[19, 20], BI is divided into three prediction models, including the binary classification model（1.BI≤20; 2. BI≤5; 2. BI＞5&BI≤10；3.BI＞10&BI≤20；4.BI＞20，and BI＞20）、multi classification model) and numerical model. To train and test the model performance, we obtained 710 sample data, of which 70% were randomly selected as training data and the remaining 30% were used as validation data, respectively. After using one-hot encoding to transform the value into binary and multi classification model, we construct a comprehensive model using RF-DT-SVM-Logistic-DNN model. For numerical prediction, this paper mainly uses Deep Neural Networks. In this article, we use Stata 17.0 to construct linear regression for impact factor analysis, and Python 3.9.16 to construct machine learning and neural network models. The packages used in this paper are Scikit-earn 1.2.1, and TensorFlow (GPU) 2.6.0.

2.3 Model performance evaluation

Model evaluation quality primarily focuses on the usage of R² and AIC/BIC. R² is conventionally used to assess the degree of correspondence of the regression model, while AIC/BIC is used to limit the number of parameters according to the degree of fitting of observed data, thus preventing overfitting. Furthermore, Binary Cross-Entropy (BCE) and Multiclass Cross-Entropy (MCE) serve as the preferred loss functions for the neural network model. Evaluating the training results of machine learning methods is usually done through performance indices, e.g. ‘precision’ and ‘recall’, which are derived from a confusion matrix[21].

These metrics cover the model fitting characteristics and avoid overfitting, and perform excellent interpretation and predictive effect. To compare the efficiency of different models and to predict the continuous variable model (BI), this paper adopts the cross-sectional comparison using R², while for the categorical variable model, accuracy, precision, and recall are used for comprehensive evaluation.

3.1 Variables description and selection

In Table 1, The BI's mean in Siming District between 2015 and 2022 is 9.27, with a standard deviation of 12.31. The highest BI recorded was 92, on 22nd July 2016, while the lowest was 0. Autocorrelation analysis was performed on a set of 78 variables (Additional file 1: Fig S1.). The analysis reveals a significant positive correlation among the mean temperature, daily minimum temperature, and daily maximum temperature. Conversely, there is a marked negative correlation between the mean pressure and mean temperature, daily minimum temperature, and daily maximum temperature separately.

Using the Decision tree method, 20 viables were identified in which there is a high weight of unused containers, stagnant water, and cistern water jars, tanks, and basins. Using the Random Forest method, 18 viables were identified in which there is a high weight of unused containers, stagnant water, and cistern water jars, tanks, and basins in Fig 2. The findings from assessing autocorrelation in the random forest screening variables, decision tree screening variables, and the combined random forest and decision tree screening variables have been illustrated in (Additional file 1: Fig S2.). The results indicate that there is an insignificant correlation between the variables, except for the barometric pressure and air temperature variables.

Table 1. Basic Information of BI and Meteorological Data

Variable	Mean	Std. dev.	Min	Max
BI	9.27	12.31	0.00	92.00
tem low (℃)	21.05	5.08	6.20	28.20
tem mean (℃)	23.58	5.17	8.30	31.10
tem high (℃)	27.70	5.55	11.70	36.80
sunshine hour (h)	6.01	3.98	0.00	14.80
precipitation (mm)	2.34	8.16	0.00	97.50
humidity (%)	76.45	12.85	27.00	99.00
pressure (hPa)	996.47	6.05	980.80	1013.60
wind (m/s)	2.77	0.99	0.90	7.20

3.2 Analysis of meteorological and biotope factors influencing the BI

As some meteorological conditions were missing from the variables screened by the random forest, insignificant meteorological factors may have been produced. To consider a more comprehensive set of influencing factors, a regression was carried out using random forests and decision trees to screen variables in parallel sets. This was done to investigate the influence of selected variables on the BI, with the results illustrated in Table 2.

Table 2. Analysis of the impact of meteorological and biotope factors on the Breteau Index

Variable	OLS	Stepwise OLS	GAM	Variable	OLS	Stepwise OLS	GAM
containers	15.760***	16.250***	3.881***	sunshine_hour	-0.025		-0.008
containers	-8.530	-9.070	-5.290	sunshine_hour	(-0.230)		(-0.270)
tanks	20.900***	21.170***	3.045**	precipitation3	0.053		-0.013
tanks	-8.740	-9.170	-3.040	precipitation3	-1.000		(-0.900)
bonsai	6.685**	6.703**	2.238**	sunshine_hour2	0.251	0.251*	0.043
bonsai	-3.12	-3.2	-3.11	sunshine_hour2	-1.760	-2.540	-1.120
pressure7 (hPa)	-0.062		-0.058	tem mean 3	0.374		0.051
pressure7 (hPa)	(-0.430)		(-1.430)	tem mean 3	-0.840		-0.450
precipitation7	0.103*	0.128***	0.002	pressure3	0.300		-0.129*
precipitation7	-2.480	-3.570	-0.180	pressure3	-1.310		(-2.120)
other	3.653		-0.904	humidity1	-0.020		-0.021
other	-0.430		(-0.330)	humidity1	(-0.380)		(-1.550)
wind3	-0.378	-0.495	0.301*	tem high3	-0.083		-0.034
wind3	(-0.810)	(-1.320)	-2.350	tem high3	(-0.250)		(-0.390)
wind5	1.071*	0.901*	-0.005	precipitation1	0.020		-0.013
wind5	-2.200	-2.200	(-0.040)	precipitation1	-0.610		(-1.570)
tires	30.020**	31.530***	4.669	wind2	-0.165		0.357**
tires	-3.110	-3.370	-1.370	wind2	(-0.320)		(-2.630)
wind	-0.004		-0.002	humidity5	0.023		0.006
wind	(-0.010)		(-0.020)	humidity5	-0.570		(-0.580)
sunshine_hour4	-0.145	-0.174	0.036	tem high2	0.005		-0.078
sunshine_hour4	(-1.210)	(-1.750)	-1.180	tem high2	-0.020		(-1.340)
wind1	0.290		0.209	pressure4	-0.334	-0.204*	0.117*
wind1	-0.570		-1.630	pressure4	(-1.550)	(-2.240)	-2.030
wind6	0.265		-0.017	tem_low7	0.455	0.706***	0.183**
wind6	-0.550		(-0.130)	tem_low7	-1.700	-3.470	-2.700
wind7	0.506	0.600	0.249*	tem high7	-0.353	-0.393*	-0.077
wind7	-1.060	-1.640	-1.990	tem high7	(-1.680)	(-2.120)	(-1.420)
humidity7	0.022		-0.006	humidity	0.0342		0.0250*
humidity7	-0.530		(-0.520)	humidity	-0.680		-2.040
constant	82.870	199.000*	69.770	AIC	5278.4	5251.2	731.4
constant	-0.560	-2.150	-1.640	BIC	5420.0	5315.1	872.9
N	710	710	710	R²	0.402	0.396
P<0.05, P<0.01, **P<0.001

Several key biotope elements exert influence on the BI, including used tires with stagnant water; unused containers with stagnant water; cistern water jars, tanks, and basins with stagnant water; other water bodies with stagnant water; and bonsai aquatic plants with stagnant water. These habitat indicators are all habitats directly related to the breeding of Aedes aegypti larvae. Meanwhile, these meteorological factors influence the BI: the 20-20 hour precipitation lags before seven days; mean wind speed lags before three, five, and seven days; sunshine hour lags before two days; mean pressure lags before three and four days; mean humidity; maximum temperature lags before seven days; and mean temperature lags before seven days. In a nutshell, the most significant impacts stem from the following factors: used tires with stagnant water; unused containers with stagnant water; cistern water jars, tanks, and basins with stagnant water; mean temperature over seven days; and 20-20 hour precipitation lags before seven days.

3.3 Three Predictive models and evaluation

The study established a binary classification model to predict the BI rank. The Deep Neural Network model was found to be the most effective, with an area under the curve (AUC) of 0.96 (95% CI: 0.91-1.00). The random forest model had an AUC of 0.85 (95% CI: 0.77-0.93), while the support vector machine model had an AUC of 0.81 (95% CI: 0.72-0.90). Additionally, the logistics model had an AUC of 0.82 (95% CI: 0.73-0.91), but the decision tree model yielded poor results with an AUC of 0.58 (95% CI: 0.47-0.68) in Fig 3. Consequently, at a significance level of 0.05, no discernible difference was observed between the performance of the Deep Neural Network model and the decision tree model. Notably, distinctions in performance were detected between the support vector machine model, the logistic regression model, and the decision tree model. It is noteworthy that the Deep Neural Network model outperformed all other models in terms of performance.

The binary rank prediction models for the Breteau Index were all found to be effective. Out of these, the neural network model and random forest showed the best prediction performance, with a precious rate of 0.865 and 0.853, respectively. The support vector machine model, on the other hand, performed slightly worse, with a precious rate of 0.675. The BI multiclassification rank prediction model remains effective, showing little variance among the methods. Among them, the neural network model outperforms the support vector machine model with a precious rate of 0.532 and 0.302, respectively. As for numerical prediction models for the BI, the study employed a Deep Neural Network methodology to construct a BI forecasting model (R²=0.6797), which is highly proficient and suitable for quantitative predictions of the BI in Fig 4.

Table 3. Comprehensive quality evaluation of the Breteau Index binary prediction model

Method	Precision rate	Recall rate	F₁score
Random Forest	0.853	0.850	0.807
Decision Tree	0.741	0.801	0.767
Support Vector Machine	0.721	0.849	0.780
Logistics model	0.772	0.817	0.778
Deep Neural Network	0.865	0.869	0.867

Table 4. Comprehensive quality evaluation of BI multiclassification prediction models

method	Precision rate	Recall rate	F₁ score
Random Forest	0.500	0.577	0.481
Decision Tree	0.513	0.498	0.505
Support Vector Machine	0.302	0.549	0.390
Logistics model	0.491	0.549	0.469
Deep Neural Network	0.532	0.535	0.529

3.4 Extrapolation tests from other cities in Fujian Province

In addition to Xiamen, we gather pertinent data from various sources in other cities within Fujian Province to assess the generalizability of the aforementioned model. Upon examining the extrapolation outcomes for these cities, the Deep Neural Network model and the random forest model emerged as the most effective, exhibiting respective AUC values of 0.73 (95% CI: 0.7-1.00) and 0.78 (95% CI: 0.76-0.81). Statistically, these values did not indicate a significant difference. Conversely, the support vector machine model achieved an AUC of 0.73 (95% CI: 0.70-0.75), while the logistic model yielded an AUC of 0.70 (95% CI: 0.67-0.73). Regrettably, the decision tree model displayed suboptimal results with an AUC of 0.59 (95% CI: 0.56-0.62), as illustrated in Fig 5. Therefore, at a significance level of 0.05, the Deep Neural Network model and the random forest model significantly outperform the decision tree model and logistic regression model. Furthermore, the random forest model demonstrates a significant superiority over the support vector machine model, while any other model surpasses the decision tree model.

The binary extrapolation prediction models for the Breteau Index were all found to be effective. Out of these, the neural network model and the Logistic model showed the best prediction performance, with a precious rate of 0.866 and 0.855, respectively. The support vector machine model, on the other hand, performed slightly worse, with a precious rate of 0.787 in Table 5. Meanwhile, the BI multiclassification extrapolation prediction model continues to demonstrate effectiveness. In comparison, the precision rates for the random forest and logistic models are 0.532 and 0.521, respectively, indicating their relatively superior performance. Surprisingly, the Deep Neural Network model does not exhibit optimal performance in this context, potentially attributed to differences among datasets, as highlighted in Table 6. What’s more, the binary rank prediction model (R²=0.26) exhibits significant deviations from the performance on the original dataset. The numerical prediction model requires further refinement, as indicated in Fig 6.

Table 5. Comprehensive quality evaluation of the Breteau Index binary extrapolation prediction model

Method	Precious rate	Recall rate	F₁score
Random Forest	0.837	0.886	0.840
Decision Tree	0.821	0.821	0.821
Support Vector Machine	0.787	0.887	0.834
Logistics model	0.855	0.857	0.856
Deep Neural Network	0.866	0.889	0.872

Table 6. Comprehensive quality evaluation of BI multiclassification extrapolation prediction models

method	Precious rate	Recall rate	F₁ score
Random Forest	0.532	0.519	0.525
Decision Tree	0.528	0.613	0.518
Support Vector Machine	0.369	0.608	0.459
Logistics model	0.521	0.600	0.545
Deep Neural Network	0.445	0.536	0.454

Aedes albopictus, which has a wide global distribution and has spread from South-East Asia to all continents of the world (except Antarctica), transmits a wide range of diseases[22]. As a subtropical oceanic monsoon climate province, Fujian is ideally suited for breeding Aedes albopictus, which facilitates the spread of dengue viruses. Situated on the southeastern coast of China, Xiamen Fujian serves as a pivotal hub city near high-risk Southeast Asia regions. As Xiamen resumes interactions with the global community post the COVID-19 pandemic, there arises an urgent need for more stringent measures to forestall the importation, transmission, and outbreak of dengue fever. Given the absence of effective vaccines, the pivotal strategy lies in the meticulous control of arbovirus vectors. This study compares different machine learning models to develop an optimal model to predict the density of mosquito vectors in Xiamen and further validated using BI from other cities in Fujian.

Meteorological variations as a major influence on mosquito vector transmission and dengue fever epidemics[23]. Good climatic conditions can shorten the development time of mosquitoes and the replication time of viruses in mosquitoes, and increase the activity level of adult mosquitoes, from the external environment to change the development cycle of larval mosquitoes to increase adult mosquito densities, a certain amount of start-up and action time is needed, which leads to the lag effect The emergence of stress[24]. Vindhya et al. conducted a time series analysis on Sri Lanka's Bray Chart index with a cycle of weeks, but its lag analysis is in the unit of months, which greatly exceeds the intergenerational interval of Aedes aegypti, and lacks the analysis of meteorological factors and the change factors of Bray Chart index in a short period[25]. Muhammad Irfan Akram et al. carried out correlation analysis and regression analysis on Pakistan meteorological factors and BI[26]. But considering that Aedes aegypti takes a certain time from egg laying to hatching, the time lag effect was taken into account in the study. We find that BI is influenced by meteorological factors like precipitation lags before seven days, wind speed lags before three, five, and seven days, sunshine hour lags before two days, pressure lags before three and four days, humidity and temperature lags before seven days. Among these meteorological factors, the lag effect of precipitation and temperature and humidity on the pupal stage of Aedes aegypti larvae is more pronounced than that of sunshine hours and air pressure, and the lag effect of wind speed lasts longer. This has significant implications for the work of local disease control agencies, such as the need to promptly clean up stagnant water factors within at least a week after a rainy day.

Regardless of the location of the outdoor containers or the general surrounding area of the localities, Lim et al. found that the larval productivity was higher in outdoor containers. Both indoor and outdoor artificial containers were found to be equally breeding sites for Ae aegypti[27]. In addition, Qiang G measured the water-holding capacity of the drainage system and the density of mosquito larvae. Catch basins and stormwater manhole chambers in residential neighborhoods were found to be dominated by Aedes albopictus. This study primarily manifested through three ecological indicators: the quantity of unused containers with stagnant water, the quantity of cisterns, tanks, and pots with stagnant water, and the quantity of aquatic plants with stagnant water in bonsai. These indicators exhibit a closer association with the populace's daily living environment than other ecological markers. Aedes aegypti, requiring a wet habitat for oviposition and reproduction, The larval population was mostly established in the areas having exposed shady conditions and less distance from human habitation. They find essential breeding grounds in containers of stagnant water common in households and aquatic plant pots[28]. Conversely, waterlogging situations such as wigwam nullahs and waste tires, distant from the majority of the population's living environment, do not exhibit a significant impact. This underscores the need to enhance the control and management of small, waterlogged containers in homes and communities to curb the breeding sites of Aedes aegypti mosquitoes in the ongoing fight against dengue fever. This finding will be useful in promoting awareness in the Xiamen Ministry of Health vector control personnel and residents on dengue vectors breeding habitats and the need for their eradication[29].

As a result of machine learning methods, complex data connections can be managed, variables can be autonomously selected, and nonlinear relationships and high-dimensional data can be handled in a wide range of applications beyond the capabilities of traditional linear regression methods[30]. The study tackles the forecasting of the Breteau Index (BI) for dengue transmission, a complex issue modulated by diverse factors like meteorological conditions, habitats, and human activities. Traditional linear models fall short in capturing the full spectrum of patterns due to the intricate relationships among these factors and data's high dimensionality. Deep Neural Networks (DNN), with their multilayered structure, are adept at processing complex nonlinear relationships and extracting feature representations, making them well-suited for high-dimensional data and intricate pattern detection. For dengue transmission forecasting, DNNs' ability to autonomously discern abstract features and navigate through the data's nonlinear intricacies significantly boosts predictive accuracy and model generalization, validating their application for this multifaceted problem. In comparison with shallow neural networks (NN), DNN have better feature expression and are capable of fitting complex mappings[31].

Ding F 's approach is limited by significant data collection costs, prevalent missing data, and sensitivity to clustering quality[8], Marston C’s study employs single Random Forests and Google Earth Engine for mosquito vector analysis, yet faces challenges with data collection costs and model sensitivity to clustering and temporal parameters[14]. This study based on comprehensive biotope and meteorological data, compared other models such as RF, DT, SVR, and Logistics models, DNN model surpassed the efficiency of the alternative methods compared models in terms of all the parameters including precision, recall, and f1-score, attained a high level of a precision of 93.95%, a recall of 95.86%, and an f1-score of 96.22% respectively. Through the use of the back-propagation algorithm for dynamic parameter adjustment, the neural network model developed in this study significantly enhances prediction accuracy. Results indicate that neural networks are superior at predicting binary and multi-classification levels of BI, making them a viable option for large-scale vector risk prediction tasks.

Using Earth Observation data inputs as proxies for environmental variables, Mudele O implemented a recurrent neural network to forecast Aedes aegypti mosquito counts locally while the generalization ability of the model is constrained[32]. This study extrapolate the model, the neural network model and the logical model showed the best prediction performance, with a precious rate of 0.866 and 0.855, respectively. The proposed model has the potential to be applied beyond Siming district, since it can be applied to other regions and extended to incorporate climatic and biotope changes. As a result, early warning systems for disease outbreaks could assist in proactive planning and preparedness.

According to the binary classification model, the neural network model and random forest had the best prediction performance with accuracies of 0.865 and 0.853, respectively. In contrast, DT (AUC=0.58) and SVM (precision=0.675) perform slightly worse. With certain datasets, DT is prone to overfitting anyway, producing inaccurate and variable outputs in comparison to other models[33]. A decision tree is constructed to determine the non-linear relationship between the features and to ultimately output the conditional distribution expectation[34]. In complex feature landscapes, decision trees risk overfitting, diminishing efficacy. SVMs, operational in high-dimensional spaces, face performance declines due to escalated computational demands in feature-rich scenarios. Disproportions in class distributions can bias decision trees and SVMs towards majority classes, undermining minority class accuracy. Critical hyperparameters, including decision tree depth and SVM kernel type, necessitate careful calibration to optimize performance. Furthermore, the effectiveness of decision trees and SVMs hinges on judicious feature selection; suboptimal or redundant features can impair model outcomes.

However, there are several limitations associated with the meteorology-based and biotope-based DNN models used for BI prediction. First, it is concluded that DNN is the best model for Breteau prediction, which provides commendable predictive results for forecasting the BI. However, in the multiclassification extrapolation prediction model for BI, the Deep Neural Network model does not perform optimally, which may be caused by differences in the datasets. In order to avoid serious overfitting, it is necessary to modify the adjustments of each model before applying the default settings[35]. Second, although meteorological and biotope data were integrated and analyzed in this study, other sources of information about relevant risk indicators, such as migrants, may subsequently be incorporated. Furthermore, the binary rank prediction model (R²=0.26) shows significant deviations from its original performance. Variations in measurement data and methodologies across different cities or variations in data sources may account for this variance.

This study introduces a Deep Neural Network (DNN) model for predicting the Breteau Index (BI), offering an approach to forecasting dengue fever. A comparative analysis with various machine learning algorithms, including Support Vector Regression (SVR), Random Forest (RF), Decision Trees (DT), Logistic Models, and the Least Absolute Shrinkage and Selection Operator (LASSO), revealed the superior predictive performance of the DNN model. Additionally, the study examines the impact of meteorological and habitat factors on the BI, providing a scientific basis for targeted measures by disease control authorities. The reliability and generalizability of the model were further affirmed through validation with BI data from other cities in Fujian, enhancing the credibility of the research findings.

BI: Breteau Index; SVR: support vector regression; RF: Random Forest; DT: Decision Tree; DNN: Deep Neural Networks; LASSO: least absolute shrinkage and selection operator; OLS: ordinary liner regression; Stepwise OLS: stepwise regression; GAM: generalized additive mode; CDCs: Chinese Center for Disease Control and Prevention; DF: dengue fever; DHF: dengue hemorrhagic fever; GLM: Generalized Linear Mode; ML: machine learning; SVM: Support Vector Machine; GBM: Gradient Boosting Machine; CI: container indexes; HI: house indexes; WHO: World Health Organization; AIC: Akaike information criterion; BIC: Bayesian Information Criterions; BCE: Binary Cross-Entropy; MCE: Multiclass Cross-Entropy; AUC: area under the curve; NN: neural networks; kNN: k-Nearest Neighbor; SMO: Sequential Minimal Optimization

Acknowledgements

We thank the staff members at the Centers for Disease Control and Prevention for their great assistance in coordinating data collection.

Authors' contributions

ZMY and HL conceived and designed the study. ZMY and HL collected and prepared the data. HL analyzed the data, helped by TMC and JR. QCW and ZMY drafted the manuscript. TMC and HJW reviewed the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by the Science and Technology Department Natural Science Foundation of Fujian Province(2020D033)

Availability of data and materials

The datasets used and/or analyzed during the current study are available from the corresponding author on special request.

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Author details

¹Xiamen Center for Disease Control and Prevention, 361000 Fujian, China. ²School of Public Health, Xiamen University, 361101 Fujian, China.²Xiamen Meteorological Service Center, 361100 Fujian, China.

Messina JP, Brady OJ, Golding N, Kraemer MUG, Wint GRW, Ray SE, et al. The current and future global distribution and population at risk of dengue. Nat Microbiol. 2019;4 9:1508-1515.
Bhatt S, Gething PW, Brady OJ, Messina JP, Farlow AW, Moyes CL, et al. The global distribution and burden of dengue. Nature. 2013;496 7446:504-507.
Stanaway JD, Shepard DS, Undurraga EA, Halasa YA, Coffeng LE, Brady OJ, et al. The global burden of dengue: an analysis from the Global Burden of Disease Study 2013. Lancet Infectious Diseases. 2016;16 6:712-723.
Paupy C, Delatte H, Bagny L, Corbel V, Fontenille D. Aedes albopictus，an arbovirus vector: from the darkness to the light. Microbes Infect. 2009;11 14-15:1177-1185.
Paupy C, Ollomo B, Kamgang B, Moutailler S, Rousset D, Demanou M, et al. Comparative role of Aedes albopictus and Aedes aegypti in the emergence of Dengue and Chikungunya in central Africa. Vector-Borne Zoonot. 2010;10 3:259-266.
Musso D, Nilles EJ, Cao-Lormeau VM. Rapid spread of emerging Zika virus in the Pacific area. Clin Microbiol Infec. 2014;20 10:O595-O596.
Aburas HM: ABURAS Index: A Statistically Developed Index for Dengue-Transmitting Vector Population Prediction. In: PROCEEDINGS OF WORLD ACADEMY OF SCIENCE, ENGINEERING AND TECHNOLOGY, VOL 23, vol. 232007: 151-154.
Bowman LR, Runge-Ranzinger S, McCall PJ. Assessing the Relationship between Vector Indices and Dengue Transmission: A Systematic Review of the Evidence. Plos Neglected Tropical Diseases. 2014;8 5: e2848.
Sanchez L, Cortinas J, Pelaez O, Gutierrez H, Concepción D, Van der Stuyft P. Breteau Index threshold levels indicating risk for dengue transmission in areas with low infestation. Trop Med Int Health. 2010;15 2:173-175.
Mudele O, Andreo V, Porcasi X, Scavuzzo CM, Lopez L, Gamba P: Predicting Aedes Aegypti Eggs Count Using Remote Sensing Data and a Generalized Linear Model. In: 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS2021: 84-87.
Focks DA, Haile DG, Daniels E, Mount GA. Dynamic Life Table Model for Aedes-Aegypti (Diptera, Culcidae) - Analysis of the Literature and Model Development. Journal of Medical Entomology. 1993;30 6:1003-1017.
Otero M, Solari HG, Schweigmann N. A stochastic population dynamics model for Aedes Aegypti: Formulation and application to a city with temperate climate. Bulletin of Mathematical Biology. 2006;68 8:1945-1974.
Marston C, Rowland C, O'Neil A, Irish S, Wat'senga F, Martín-Gallego P, et al. Developing the Role of Earth Observation in Spatio-Temporal Mosquito Modelling to Identify Malaria Hot-Spots. Remote Sens-Basel. 2023;15 1: 43.
Salim NAM, Wah YB, Reeves C, Smith M, Yaacob WFW, Mudin RN, et al. Prediction of dengue outbreak in Selangor Malaysia using machine learning techniques. Scientific Reports. 2021;11 1: 939.
Kuno G. Review of the factors modulating dengue transmission. Epidemiol Rev. 1995;17 2:321-335.
Ding FY, Fu JY, Jiang D, Hao MM, Lin G. Mapping the spatial distribution of Aedes aegypti and Aedes albopictus. Acta Tropica. 2018;178:155-162.
Marinho RA, Beserra EB, Bezerra-Gusmao MA, Porto VD, Olinda RA, dos Santos CAC. Effects of temperature on the life cycle, expansion, and dispersion of (Diptera: Culicidae) in three cities in Paraiba, Brazil. J Vector Ecol. 2016;41 1:1-10.
Arévalo-Cortés A, Granada Y, Torres D, Triana-Chavez O. Differential Hatching, Development, Oviposition, and Longevity Patterns among Colombian aedes aegypti Populations. Insects. 2022;13 6: 536.
Hanna JN, Ritchie SA, Phillips DA, Serafin IL, Hills SL, van den Hurk AF, et al. An epidemic of dengue 3 in far north Queensland, 1997-1999. Med J Australia. 2001;174 4:178-182.
Macdonald WW. Aedes aegypti in Malaya. II. Larval and adult biology. Annals of tropical medicine and parasitology. 1956;50 4:399-414.
Buckland ST, Elston DA. Empirical-Models for the Spatial-Distribution of Wildlife. J Appl Ecol. 1993;30 3:478-495.
Benedict MQ, Levine RS, Hawley WA, Lounibos LP. Spread of the tiger: Global risk of invasion by the mosquito. Vector-Borne Zoonot. 2007;7 1:76-85.
Schnell MJ, Lee SH, Nam KW, Jeong JY, Yoo SJ, Koh Y-S, et al. The Effects of Climate Change and Globalization on Mosquito Vectors: Evidence from Jeju Island, South Korea on the Potential for Asian Tiger Mosquito (Aedes albopictus) Influxes and Survival from Vietnam Rather Than Japan. PLoS ONE. 2013;8 7: e68512.
Rocklöv J, Dubrow R. Climate change: an enduring challenge for vector-borne disease prevention and control. Nat Immunol. 2020;21 5: 479-483.
Aryaprema VS, Xue RD. Breteau index as a promising early warning signal for dengue fever outbreaks in the Colombo District, Sri Lanka. Acta Tropica. 2019;199.
Akram MI, Akram W, Qayyoum MA, Rana AA, Yasin M, Saddiq B. Vector indices and metrological factors associated with dengue fever outbreak in Punjab, Pakistan. Environ Dev Sustain. 2023;25 9:9839-9850.
Nordin O NTG, Ahmad N W, et al. Identification of Aedes aegypti (L.) and Aedes albopictus (Skuse) breeding habitats in dengue endemic sites in Kuala Lumpur Federal Territory and Selangor State, Malaysia. Southeast Asian J Trop Med Publ Hlth, . 2017; 48:786-798.
Gao Q, Wang F, Lv X, Cao H, Su F, Zhou J, et al. Aedes albopictus: production in urban stormwater catch basins and manhole chambers of downtown Shanghai, China. PLOS ONE. 2018;13 8.
Hafeez F, Naeem-Ullah U, Akram W, Arshad M, Iftikhar A, Naeem A, et al. Habitat characterization of Aedes albopictus. Int J Trop Insect Sc. 2022;42 2:1555-1560.
Tarca AL, Carey VJ, Chen XW, Romero R, Draghici S. Machine learning and its applications to biology. Plos Comput Biol. 2007;3 6:953-963.
Wang L, Tang J, Liao Q. A Study on Radar Target Detection Based on Deep Neural Networks. IEEE Sensors Letters. 2019;3 3:1-4.
Mudele O, Frery A, Zanandrez L, Eiras A, Gamba P. Dengue Vector Population Forecasting Using Multisource Earth Observation Products and Recurrent Neural Networks. Ieee J-Stars. 2021;14:4390-4404.
Douglass MJJ. Hands-on Machine Learning with Scikit-Learn, Keras, and Tensorflow, 2nd edition. PHYSICAL AND ENGINEERING SCIENCES IN MEDICINE. 2020
Kotsiantis SB. Decision trees: a recent overview. Artif Intell Rev. 2013;39 4:261-83.
Rodda GH, Jarnevich CS, Reed RN. Challenges in Identifying Sites Climatically Matched to the Native Ranges of Animal Invaders. Plos One. 2011;6 2: e14670.

No competing interests reported.

Additionalfile1FigS1.Allvariableautocorrelationanalysis..tif
Additional file 1: Fig S1. All variable autocorrelation analysis.
Additionalfile2TableS2.18variablesautocorrelationanalysisafterscreeningvariables.a.18variablesautocorrelationanalysisofDecisionTreeafterscreeningvariables.tif
Additional file 2: Table S2. 18 variables autocorrelation analysis after screening variables.
Additionalfile2TableS2.18variablesautocorrelationanalysisafterscreeningvariables.b.18variablesautocorrelationanalysisofDecisionTreeafterscreeningvariables.tif
Additional file 2: Table S2. 18 variables autocorrelation analysis after screening variables.
Additionalfile2TableS2.18variablesautocorrelationanalysisafterscreeningvariables.c.AutocorrelationanalysisofRandomForestDecisionTreeafterscreeningvariables.tif
Additional file 2: Table S2. 18 variables autocorrelation analysis after screening variables.
FigS1.docx
FigS2.docx

Three machine-learning models to predict dengue vectors Breteau Index based on meteorology and biotope in Fujian China

Status:

Version 1

Abstract

Figures

1. Background

2. Methods

3. Results

4. Discussion

5. Conclusion

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1