1. Introduction

Malaria is one of the most widespread parasitic diseases worldwide, with approximately 3.2 billion people at risk of infection [1]. Indigenous Plasmodium vivax malaria was thought to have been eradicated from the Republic of Korea (ROK) in 1984 [2]. However, a resurgence occurred in 1993 with a case near the demilitarized zone (DMZ) in northern Gyeonggi-do, adjacent to the Democratic People’s ROK [3]. After its reemergence, P. vivax malaria again became endemic in the ROK, peaking in 2007 with 2192 cases—one of the highest incidence rates among countries with temperate climates [2]. Since 2010, due to effective control strategies, the annual number of reported cases has continuously decreased, ranging between approximately 400 and 800 [4]. Nonetheless, P. vivax malaria continues to pose public health challenges in the northern parts of the ROK near the DMZ, and its elimination remains elusive.

Vector surveillance has been conducted to support decision-making regarding the elimination goal; however, its practical utility in influencing disease control policy has not been clearly established. In the ROK, eight species of Anopheles mosquitoes have been identified [5], with Anopheles sinensis considered the primary vector of P. vivax malaria [6]. The numbers of malaria vector mosquitoes and human malaria cases typically begin to rise in early spring and decrease significantly in late autumn, displaying a corresponding trend [3, 7]. However, these trends do not establish a clear relationship between mosquito abundance and malaria incidence. For example, regions with a high density of malaria vector mosquitoes do not necessarily have high malaria incidences [4, 8]. Additionally, the onset of malaria cases often precedes the active period of malaria vectors [4, 8]. Considering that the association between malaria vector mosquitoes and the incidence of malaria is not well-defined, vector control program planning is challenging. Therefore, it is essential to quantitatively assess these associations to establish refined objectives for malaria management and conduct effective control programs.

If associations between vector mosquito abundance and malaria epidemics are well-established, monitoring mosquito populations could be crucial for effective disease control. Since 2009, the Division of Vectors and Parasitic Diseases at the Korea Disease Control and Prevention Agency (KDCA) has conducted surveillance of vector mosquitoes in malaria-risk areas, publishing information on mosquito abundance and P. vivax infection rates [8]. However, timely forecast data for malaria epidemics are limited due to an approximate 2-week delay in data collection and processing. Previous studies have suggested that the abundance of Anopheles mosquitoes can be estimated based on various meteorological factors. Studies conducted in the ROK have shown that meteorological factors, such as ambient temperature, humidity, and precipitation, can be used to estimate mosquito distribution [9, 10]. Similar studies conducted in China, another temperate zone, have demonstrated that meteorological factors significantly affect mosquito distribution [11]. Therefore, it is plausible to use these ecological factors to predict vector abundance in real-time, thereby, addressing the reporting delay.

In this study, we aimed to nowcast mosquito abundance at each monitoring site to address the current reporting delays. We applied a “nowcasting” approach, which estimates current mosquito abundance using historical data. Specifically, we developed machine learning models that incorporate mosquito occurrence data (e.g., from 2 weeks and earlier) and real-time meteorological factors to infer the current state of mosquito abundance. Additionally, we demonstrated the utility of nowcasting by showing how accurately vector mosquito abundance can predict the peak of malaria epidemics.

2. Methods

2.1. Study Design

We nowcasted mosquito abundance using machine learning models and verified whether the results could serve as reliable predictors of malaria epidemic peaks. Figure 1 illustrates the schematic diagram of the study process. To nowcast mosquito abundance, we utilized meteorological factors combined with mosquito abundance data from previous weeks as explanatory variables. We assessed the peak of malaria cases based on the effective reproduction number, R_t, an important epidemiological measure of the transmission potential of an infectious pathogen [12]. Then, we examined the association between R_t and mosquito abundance. The study area comprises parts of three provinces (the first administrative level, Si-Do): Incheon-si, Gyeonggi-do, and Gangwon-do, which are designated by the KDCA as high-risk regions for malaria (Figure 2A and Figure S1).

2.5. Predicting the Peak of Malaria Cases Based on Mosquito Abundance Data

We defined the epidemic peak of malaria based on R_t and examined whether mosquito abundance data could effectively predict it. R_t measures the transmission potential and is defined as the average number of secondary cases produced by a single infected person within a population [12]. It is commonly used to assess surveillance systems and intervention programs for vector-borne diseases, including malaria [21]. We calculated R_t based on the aggregated weekly malaria cases at the provincial level using the backward-looking method [22]. The formula is as follows:

$$

$$

where R_t is the reproduction number, t is the number of days elapsed since the start of the epidemic, I_t is the number of cases on day t, W (s) is the current infectivity on day s after infection, a is the shape parameter, and θ is the scale parameter. To estimate current infectivity W (s) while calculating R_t, we utilized the serial interval and standard deviation for P. vivax malaria [23]. The serial interval is a key epidemiological metric that reflects the infectivity and transmission dynamics of an infectious disease. It represents the average time between the onset of symptoms in an initial case and the appearance of symptoms in a secondary case.

Then, we fitted a generalized linear model (GLM) for each province, with log-transformed weekly R_t as the response variable and log-transformed weekly vector mosquito abundances as the explanatory variable. For each year, we conducted the analysis using data from the periods when R_t was ≥1 (Figures S3, S4). Considering that the epidemic curve typically has one peak per year (Figure 3A), the moment when R_t falls below 1 can be considered the epidemic peak of that year. Thus, if the GLMs demonstrate a reliable fit, mosquito abundance may be able to predict when malaria cases will peak. The following equation is the mathematical expression of the model:

$$

where R_t represents the effective reproduction number, α₀ is the constant intercept, β_n is the regression coefficient, M_t-n denotes mosquito abundance from previous weeks, “week” refers to the week number to adjust for seasonality, and “year” represents the year to distinguish between years. The model’s explanatory power was evaluated using the significance of the regression coefficient (β₁) and R². The methods related to the calculation and application of R_t were detailed in a previous paper [24].

3. Results

Figure 2 presents the distribution of the cumulative number of malaria cases and mosquitoes collected between 2009 and 2022. During this period, 4392 malaria cases were reported in the study area, with the highest incidence per 100,000 population occurring in northern Gyeonggi-do. In total, 444,956 mosquitoes were collected, of which 157,219 were malaria vector mosquitoes. Notably, areas with high mosquito abundance differed from the hotspots for human malaria cases. While the pattern of malaria occurrence remained consistent throughout the study period, the distribution of mosquito abundances varied annually (Figures S5–S7).

Malaria cases and mosquito abundance showed clear seasonality, with an increase around April and nearly zero cases after November (Figures 3A,B and S8). R_t consistently peaked prior to the peak of malaria cases each year, suggesting that it could serve as an indicator for malaria epidemic peaks (Figure 3A). Additionally, the increase in malaria cases often precedes the increase in vector mosquito abundance (Figure S8). Therefore, instead of predicting the onset of malaria cases, it appears more effective to forecast the peak period of malaria as defined by R_t. Fluctuations in ambient temperature and precipitation appear to be correlated with the seasonality of malaria cases and mosquito abundance (Figure 3C). The time series data illustrating associations among the variables are described in detail in Figures S9–S15.

Machine learning models reliably nowcasted log-transformed vector mosquito abundance using meteorological factors combined with the log-transformed mosquito abundances from 2 weeks prior (Figure 4). The ensemble models showed the best performance in nowcasting mosquito abundance, with an RMSE of 0.90 and an R² value of 0.85. The RMSE and R² were 0.91 and 0.84 for the GBM model and 0.92 and 0.85 for the XGB model, respectively. No significant differences in predictive performance were evident between variable sets one and two (Tables S4–S6). The greatest predictive accuracy was observed when the moving average of mosquito abundance over the previous 3 weeks served as an input. A model is generally assumed to adequately explain the majority of the data when it has an R² > 0.7 [24]. Additionally, the range of vector mosquito abundance data extends from 0 to 2,398, with an average value of 23.68 (on the same scale as RMSE). Therefore, the models appear reliable for nowcasting vector mosquito abundance. The results of visualizing the predicted and observed values for each model are depicted in Figures S16–S18.

The predictive performance of GLMs was generally fair, and vector mosquito abundance appears to be a significant indicator for malaria epidemic peaks. The models forecasting R_t 2 weeks in advance using mosquito abundance showed a significant association between log-transformed R_t and log-transformed mosquito abundance (Figure 5 and Table S7). To examine the associations by year, while adjusting for the effect of the week (seasonality), we used partial residuals of log-transformed R_t. In all regions, the coefficient for the mosquito abundance variable was statistically significant and exhibited an R_t > 0.7, which is generally considered fair. Gangwon-do showed the lowest predictive performance, likely due to the presence of fewer malaria cases compared with other regions. The results of fitting GLMs for each region, using various combinations of forecast periods and lengths of moving averages, are presented in Tables S8–S11.

4. Discussion

Using mosquito surveillance data and human malaria cases collected between 2009 and 2022 in the ROK, our study provides valuable insights for developing malaria management policies in malaria-endemic regions. First, we proposed a reliable model for nowcasting mosquito abundance, driven by various meteorological factors. Considering the reporting lag that prevents access to real-time mosquito data, the estimated mosquito abundance from our model offers crucial components for establishing timely vector control strategies. Second, we suggested that mosquito abundance could be an indicator of malaria epidemic peaks. Thus, our findings are expected to contribute to establishing an effective malaria control system by accessing real-time vector surveillance data and utilizing it to identify regions and periods at high risk for malaria outbreaks.

Vector mosquito abundance has been identified as an important determinant of malaria epidemic peaks based on the results of GLMs, which use R_t as the response variable and vector mosquito abundance from previous weeks as the main explanatory variable. Furthermore, a higher number of malaria cases may create more opportunities for disease transmission, sustaining an elevated R_t and accelerating the spread of the epidemic [12]. However, the size and timing of the epidemic peak cannot be determined solely by the number of malaria cases. Vector mosquito abundance, as a key driver of malaria transmission, also plays a critical role in shaping both the timing and magnitude of the epidemic. Notably, the association between vector mosquito abundance and R_t may vary depending on the number of cases. For instance, when the number of cases is extremely low or extremely high, the association can differ from patterns observed within a moderate range of cases. When the number of infected individuals is small, high mosquito abundance can sustain or even steeply increase R_t, facilitating local transmission and potentially leading to a gradual rise in cases. In this scenario, mosquito abundance strongly influences the spread of malaria, amplifying transmission from the limited existing cases. Conversely, when the number of infected individuals is very high, the role of mosquito abundance becomes more context dependent. If mosquito abundance is sufficient to sustain transmission, R_t may remain above 1, driving a rapid increase in cases until the epidemic peak is reached. However, if mosquito abundance is insufficient relative to the number of infected individuals, transmission efficiency may decline due to increased competition among mosquitoes for infected hosts. These dynamics contribute to the eventual peak and subsequent decline of the epidemic.

Our model could forecast the timing and location of surges in malaria vector mosquitoes. The results can be used to enhance the timing of insecticide fogging distribution for controlling adult mosquitoes. These conventional interventions remain effective, particularly during epidemic seasons [25]. Intensive insecticide application in residential areas or livestock farms can lead to short-term reductions in mosquito abundance, thereby, preventing potential malaria outbreaks [26]. The existing vector control programs assume a consistent growth curve in vector mosquito abundance and malaria cases over the years [26]. However, our descriptive analysis (Figures S3, S4), and a previous study [27] revealed that the peak times of mosquito abundance and malaria cases were not consistent across years. In this regard, our models can be utilized in a practical manner by health authorities to design optimal control measures for reducing mosquito abundance.

Reduction of mosquito surveillance reporting lag is essential for efforts to expand our model into a larval control program. Control strategies that combine larval source management are more effective than those primarily targeting adult mosquitoes [28]. This effectiveness arises because mosquitoes cluster in water-containing habitats during their larval and pupal stages [29], making it easier to identify and target these habitats for control measures. However, constraints exist when applying our results to conduct control programs focused on the larval stage. Considering that the development period from larva to adult mosquitoes is approximately 1–2 weeks [29], at least a 2-week forecast of adult mosquito abundance is required to optimize the timing of larval control. Although the model can predict mosquito abundance 2 weeks in advance, the mosquito surveillance data have a reporting delay of 2 weeks. Thus, the most recent mosquito abundance data are those collected 2 weeks prior. By utilizing cutting-edge technologies such as smart mosquito traps [30] for real-time mosquito abundance monitoring, our model can be used to forecast adult mosquito abundance 2 weeks later. This would allow for the implementation of control measures at the appropriate time to prevent the larval stage from metamorphosing into adult mosquitoes.

We demonstrated that mosquito abundance can serve as a predictor of malaria epidemic peaks. In our statistical model, the R_t of malaria in the human population was effectively predicted by mosquito abundance from 2 weeks earlier. This finding suggests that mosquito abundance plays a crucial role in the secondary attack rates of human infections transmitted by infected mosquitoes. Considering that determination of the optimal timing for malaria control measure implementation is crucial for the efficacy [31] and cost-effectiveness [32] of these measures, our findings could be integrated into an early warning system to improve malaria control strategies by offering reliable information concerning mosquito abundance and its influence on malaria epidemics. Despite the high R² values of our models (generally around 0.8), more advanced modeling is required to use the models for practical forecasting of malaria epidemics based on mosquito abundance. For improved predictive capabilities, we strongly recommend the inclusion of additional predictors such as land use [33], temperature, humidity, precipitation [9], and extreme weather events [34]. Through such processes, the enhanced model could contribute to building malaria early warning systems in malaria-endemic regions [35].

Several limitations of this study warrant consideration. First, our nowcasting models for mosquito abundance incorporated the 3-week moving average of mosquito abundance as a predictive factor. We employed this moving average to mitigate potential errors in the mosquito counting process. However, this approach may have overestimated the performances of our nowcasting models because we used mosquito abundance data from 2 weeks prior as a predictor, which coincides with the period covered by the moving average. Second, the history of mosquito control implementation was not considered in our nowcasting of mosquito abundance. Incorporation of parameters that reflect these control measures would enhance the model’s predictive capacity. Third, our descriptive analysis revealed considerable variation in mosquito abundance across collection sites. Certain sites displayed excess zero values, potentially facilitating easier prediction of abundances at these locations compared with other locations. Consequently, this may lead to an overestimation of the models’ performances. Finally, it is crucial to acknowledge that the efficiency of mosquito collection differs between black light traps and LED traps. LED traps proved more efficient in collecting a larger and more varied mosquito population compared with black light traps [36], potentially introducing systematic bias into our analysis. Therefore, to prevent incorrect conclusions drawn from trapping methods, further assessments that consider host attraction and trapping techniques are necessary.

5. Conclusion

We have introduced a reliable model for nowcasting mosquito abundance, which could be incorporated into a malaria early warning system. Our models, which nowcast mosquito abundance based on various meteorological factors, provide evidence for the effective application of vector control measures. Furthermore, we suggest that mosquito abundance could predict malaria epidemic peaks. These results offer valuable insights for the development of malaria management policies in the ROK, where ongoing control activities have maintained P. vivax malaria at a steady-state level, but progress toward elimination has been slow. Therefore, our findings may be implemented in regions where malaria eradication presents a significant and immediate challenge.

Disclosure

The funders had no role in study design or conduct, data collection, management, analysis, or interpretation, manuscript preparation, review, or approval, or the decision to submit the article for publication.

Conflicts of Interest

The authors declare no conflicts of interest.

Author Contributions

Taehee Chang conceived the study, analyzed the data, and wrote the manuscript. Saebom Choi conceived the study, analyzed the data, and was involved in the manuscript’s review and revision process. Hojong Jun contributed to the conception of the study and assisted in data interpretation. Jong-Yil Chai assisted in data interpretation and was involved in the manuscript’s review and revision process. Sehyeon Kim assisted in data interpretation and was involved in the manuscript’s review and revision process. Sang Hoon Song and Joon-Sup Yeom were involved in the manuscript’s review and revision process. Sung-il Cho managed the dataset and played a role in data interpretation. Kyung-Duk Min was responsible for the conception of the study, assisted in data interpretation, and was involved in the manuscript’s review and revision process. All authors provided input on the study’s design and data analysis and made significant contributions to the discussion and interpretation of the results.

Funding

This research was supported by a National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (Grant NRF-2021R1C1C2012611). This study was also supported by the Korea Disease Control and Prevention Agency (Grant 6300-6331-311).

Supporting Information

Additional supporting information can be found online in the Supporting Information section.