Identifying leptospirosis hotspots in Selangor: uncovering climatic connections using remote sensing and developing a predictive model

Background

Leptospirosis is a globally significant zoonotic disease caused by the genus Leptospira pathogenic bacteria (Chacko et al., 2021). Disease outbreaks are closely linked to heavy rainfall, flooding, and hot, humid climates (Lau et al., 2010). It disproportionately affects impoverished communities in developing countries, particularly those in slums, agriculture, or water-based recreation activities (Jittimanee & Wongbutdee, 2019; Torgerson et al., 2015). Transmission to humans primarily occurs when broken skin or mucous membranes encounter water or soil contaminated by urine from infected animal reservoirs (Chacko et al., 2021; Karpagam & Ganesh, 2020). While many cases exhibit mild, flu-like symptoms and may not require treatment, severe leptospirosis can manifest as disease complications such as pneumonia, kidney failure, and pulmonary haemorrhage or can even be fatal (Karpagam & Ganesh, 2020).

Leptospirosis is endemic in tropical and subtropical areas of South Asia, Southeast Asia, and South America. While considered a neglected disease in developed countries like the United States and Europe due to less favourable environmental conditions, outbreaks can occur via travel to endemic countries or engaging in water-borne activities without adequate protection (Chacko et al., 2021; Karpagam & Ganesh, 2020). The disease also flourishes in settings with poor sanitation, as these conditions support large rodent populations, a significant disease reservoir in the Southeast Asia (SEA) region (Garba et al., 2018). Warm tropical climates with substantial rainfall common to Southeast Asia (SEA) create environments favourable for Leptospira growth, promoting transmission to humans (Nozmi et al., 2018), with Leptospira interrogans and L. borgpetersenii identified as the key pathogenic species in the region (Cosson et al., 2014).

Malaysia’s consistently hot and humid climate with intermittent heavy rainfall patterns creates a conducive environment for the disease, influencing leptospirosis outbreaks (Holt, Davis & Leirs, 2006; Lopez et al., 2019). The state of Selangor exhibits particularly high leptospirosis incidence, peaking in 2013 at 24.68 cases per 100,000 population (Tan et al., 2016). Even though the Selangor State Health Department reports the disease reduction trend in the years following, i.e., 0.63 cases per 100,000 population in 2019, the actual incidence may be underestimated due to clinical overlap with other endemic diseases like dengue fever and malaria (Benacer et al., 2016). Furthermore, climate change may further worsen leptospirosis risk through increased flood intensity and weather fluctuations, promoting vector population growth and Leptospira spread (Abdul Rahman, 2018).

Geographic information systems (GIS) revolutionised epidemiological research by analysing and layering various spatial data to look at disease patterns (Goodchild, 2005). Spatio-temporal analysis of disease patterns further enhances understanding by incorporating the time dimension, facilitating understanding disease trends and the creation of predictive models for disease management (Byun, Lee & Hwang, 2021; Convertino et al., 2021). Additionally, remote sensing complements GIS in measuring the earth’s surface environmental properties, such as rainfall patterns and surface temperature, by detecting electromagnetic waves (DeMers, 2009). These data could then be integrated into epidemiological studies to examine disease dynamics and their environmental correlations (Dukiya, 2021; Tran, Kassie & Herbreteau, 2016). Spatial data is essential for understanding disease patterns and identifying risk factors in health research. By analysing point data (e.g., individual disease cases) and aggregate data (e.g., regional disease rates), researchers can uncover valuable insights into the geographic distribution of diseases and potential contributing factors (Lin & Wen, 2022).

Spatial autocorrelation statistics (Moran’s I) measure the degree of similarity between observed values at different spatial locations, with positive autocorrelation indicating an area with high values surrounded by neighbouring values with a higher incidence than other areas (Lin & Wen, 2022). This analysis has broad applications in identifying spatial associations of disease incidence and distribution. Studies that demonstrate the clustering of diseases at particular regions in respective researched areas include kala-azar cases in India (Bhunia et al., 2013), COVID-19 in Vietnam (Thi-Bich-Thuy & Thi-Hien, 2023), and leptospirosis in Thailand (Chadsuthi et al., 2022). Meanwhile, the Getis-Ord Gi* statistics identify hotspot areas by measuring the intensity of high values. Hotspot analysis reveals spatial clusters and provides visual insights into disease trends. Identifying hotspots enables targeted resource allocation, interventions, and public awareness campaigns to alleviate environmental risk factors associated with infectious diseases. Additionally, analysis of climatic variables and their association with leptospirosis hotspots could address the issues of climate change’s impact on leptospirosis vulnerability, with the associated secondary chronic disease complications and economic loss from uncontrolled leptospirosis outbreaks.

Machine learning (ML) is a field within computer science that enables computers to learn and adapt without being explicitly programmed. It utilises algorithms to analyse data, identify patterns, and train models that automate decision-making processes (Kufel et al., 2023). Some examples of ML algorithms include artificial neural networks (ANN), support vector machine (SVM), decision tree, and Random Forest. Successful development of a hotspot predictive model relies on integrating pre-processing techniques to ensure high-quality data. Techniques like normalisation and sampling adjustments are essential in ML to address biases and data imbalances (Abas et al., 2024). Underreporting or mishandling geographic of confirmed cases can be handled by oversampling, where minority classes are augmented, and weights are assigned to equalise data for feature target class, ensuring a quality data input for building a robust and reliable predictive model (Abas et al., 2024).

Objectives

This research aims to explore the spatio-temporal distribution of leptospirosis hotspot areas and their association with climatic factors to develop a predictive hotspot area using GIS and remote sensing methods. The study hypothesis posits a significant correlation between leptospirosis hotspot areas and specific climatic factors. Therefore, this study could (1) describe the characteristics of leptospirosis cases in Selangor from 2011 to 2019 through descriptive statistics; (2) determine the spatiotemporal distribution map of leptospirosis cases; (3) determine leptospirosis hotspot maps at the district level in Selangor; (4) describe Selangor’s climatic characteristics using remote sensing imagery data (monthly rainfall and monthly land surface temperature); (5) determine the association between climatic factors and leptospirosis hotspot areas; and (6) to validate the leptospirosis hotspot area predictive model (machine learning algorithm). This life-threatening condition demands proactive hotspot prediction and outbreak management. The methodology used in this study can be adapted for future research on leptospirosis in other regions, contributing to a better understanding of disease patterns and risk factors.

Study design and data collection

The study was conducted in Selangor, situated in the centre of Peninsular Malaysia, bordering the Perak, Pahang, and Negeri Sembilan. It is a retrospective ecological observational study with GIS and remote sensing mapping and analysis concerning leptospirosis in Selangor using secondary data over nine years from January 2011 to December 2019. This study obtained data from various available spatial data sets, including the leptospirosis case reports, satellite images, river hydrometric levels, and topographical data of Selangor. All of Selangor’s subdistrict polygon areas represent the sampling unit with a sampling size of fifty-five subdistrict polygon areas. The study utilises the universal sampling method of all notified laboratory-confirmed leptospirosis cases. Researchers examined the data to determine the sociodemographic characteristics of cases and coordinates of possible infection sources. Before the study commenced, ethical approval (NMRR ID-22-01548-C0Z IIR) was obtained from the Medical Research Committee on the Ministry of Health, the Director General of the Ministry of Health, Malaysia, and the Ethic Committee for Research Involving Human Subjects, Universiti Putra Malaysia and registered with the National Medical Research Registry.

Operational definition of variables

Dependent variable: A leptospirosis hotspot area in this study refers to a spatially or geographically concentrated area in Selangor with a statistically significant clustering of reported cases in a subdistrict compared to surrounding subdistrict areas identified using the Getis-Ord Gi* statistics in the ArcGIS Pro software (Esri, 2021). Some researchers suggest including modifiers such as “transmission hotspots,” “emergence hotspots” or “burden hotspots” to further explain how hotspots are defined. The burden hotspot, which is an area with elevate disease incidence or a geographic cluster of cases, describes hotspot areas in this study (Lessler et al., 2017). The Getis-Ord Gi* analysis examines a hotspot according to the high aggregated values (cases) for a subdistrict surrounded by subdistrict(s) with high values. In this research, the confidence levels corresponding to the binned p-values are labelled ‘Gi_bin.’ The confidence levels involve a range of values from −3 to 3. Within this range, values from −3 to −1 are categorised as “cold spot” areas, values from −1 to 1 are considered “not significant,” and values from 1 to 3 are identified as “hot spot” areas. Independent variables: (1) the rainfall images in millimetres (mm) captured by the satellite will be analysed to obtain the average monthly rainfall, and (2) the LST is the earth’s surface temperature captured by the satellite in degrees Celsius. Figure 1 shows the conceptual framework of this study (NASA, 2022). The framework outlines the geospatial processes of satellite images and their association with leptospirosis hotspot areas in Selangor from 2011 to 2019.

Data processing and analysis

Processing leptospirosis data: The data cases were cleaned and geocoded according to their longitude and latitudes of possible infection locations. Next, the coordinates were imported to the ArcGIS Pro software and plotted as point-shape files on the layered base map to determine the spatial distribution of the leptospirosis case. Processing satellite data: Images obtained were processed to obtain the average monthly data for rainfall and LST for each subdistrict using the clipping and zonal statistic tools in the ArcGIS Pro software. The images were resampled to smaller pixels to be projected and overlayed on all subdistrict polygon shapefile boundaries. The output obtained in numerical values was exported as an Excel table for further processing and analysis. The descriptive statistics of leptospirosis cases for each month from 2011 to 2019 were analysed using Statistical Package for the Social Sciences (SPSS) software version 27.0. Descriptive data was presented using the appropriate frequency and percentage tables.

The spatio-temporal analysis of patterns and distribution of leptospirosis hotspot areas in Selangor was analysed using Moran’s I and Getis-Ord Gi spatial statistical tools in ArcGIS Pro software. Moran’s I analysis measures the global spatial autocorrelation between locations based on their characteristics. The tool calculates Moran’s index value with a z-score and p-value to evaluate the significance of the index. The value usually ranges from −1 to +, and a statistically significant p-value and positive z-score denote that the data is more spatially clustered. The pattern derived from the analysis indicates whether the cases are clustered, dispersed, or randomly distributed. Meanwhile, the Getis-Ord Gi* statistics perform hotspot analysis of the aggregated plotted cases in a subdistrict, indicating whether it is a hot or cold spot based on the clustering and z-scores derived. It involves comparing the aggregated values (total cases for a particular subdistrict polygon area in a specific month and year) with the surrounding values. The method computes a z-score and a p-value to assess the spatial statistical significance of the local clusters (Esri, 2021).

To assess the relationship between leptospirosis hotspot areas and climatic variables, the Spearmann’s correlation analysis was employed. Positive significant correlation at p-value <0.05 between the variables examined were examined. Subsequently, the predictive models for leptospirosis hotspot areas were developed using Python’s machine learning capabilities within the Jupyter Notebook environment (Anaconda Navigator). Three well-established machine learning algorithms were selected: support vector machine (SVM), Random Forest (RF), and light gradient boosting machine (LGBM). These algorithms have been successfully applied in various leptospirosis studies to predict disease outbreaks (Ahangarcani et al., 2019; Douchet et al., 2022; Jayaramu et al., 2023; Mohammadinia et al., 2019). The best model was determined by evaluating their performance metrics in identifying actual positive hotspot areas (Abas et al., 2024).

Characteristics of leptospirosis cases in Selangor

The characteristics of leptospirosis cases are shown in Table 1. Between 2011 and 2019, Selangor reported 1,045 confirmed leptospirosis cases, primarily affecting males (73%) with an average age of 31. Malays were the most common ethnic group (67%), followed by Indians (11%) and Chinese (5%). Foreigners, mainly from Indonesia and Bangladesh, accounted for 14% of cases. While most patients recovered, 5% unfortunately died. Hulu Langat, Hulu Selangor, and Petaling districts had the highest number of cases during this period. The overall incidence showed a general decline with fluctuations in some districts. A peak occurred in early 2011–2012, followed by a decline until 2014. Another peak emerged in 2014 before a steady decrease until 2019.

A high-density pattern could be observed in Ampang, Klang, Kajang, and Damansara subdistricts in the Hulu Langat, Petaling, Gombak, and Klang districts, with cumulative cases ranging from 58 to 109. Cases were more concentrated in Selangor’s central and northeastern regions, while other subdistricts show a more scattered distribution. Throughout 108 months from 2011 to 2019, case distribution was discovered in 103 months, mostly showing a random distribution pattern with the clustering of cases occurring in certain months. Case clustering was observed in Selangor’s central regions in the earlier months of 2011 before shifting toward the western coastal areas of the state. Cases were mostly randomly distributed throughout the state in 2012, while clustering was observed in the third quarter of 2013 and 2014. The distribution of cases declined from 2015 to 2019, with yearly cases ranging from 41 to 78 cases. Even though case plotting shows a random pattern along the years, clustering of cases was observed primarily in months of the first and fourth quarters of the years.

Spatio-temporal analysis

Moran’s I analysis confirmed spatial clustering of cases in 20 out of 103 months, with a positive z-value (0.067 to 0.370) and statistically significant p-values. Meanwhile, a dispersed distribution of cases was observed in June 2018, giving a negative Moran’s I index z-value. The Getis-Ord Gi* analysis depicts statistically significant hotspots where subdistricts with similar high cases surround subdistricts with higher leptospirosis cases.

The spatial relationship was conceptualised at a fixed distance band, in which each feature is analysed within the context of neighbouring features and receives a weightage according to the specified threshold distance. The Euclidean distance method, which is the straight line distance of subdistrict polygons containing aggregated leptospirosis cases, was used in the analysis. The hotspot maps areas throughout the months were mostly scattered in subdistricts throughout Selangor, with clustering mainly observed in the central regions in Selangor. The supplementary figures provide further insights into monthly hotspot areas following Getis-Ord Gi* analysis.

This study employed statistical methods to assess the link between climate factors (monthly rainfall and land surface temperature) and areas identified as leptospirosis hotspot areas. Figure 2 shows satellite data processing steps for a processed rainfall image for a particular month. The satellite image was resampled to appropriate pixels to fit the subdistrict boundaries before zonal statistics were performed to extract monthly average rainfall and LST values for each subdistrict. The similar process was repeated for all months throughout the study period.

The analysis in Table 2 revealed a statistically significant but weak, negative correlation only between monthly land surface temperature and leptospirosis hotspots (correlation coefficient, r =  − 0.086, p-value < 0.001). Conversely, no significant correlations were found for monthly rainfall (r =  − 0.006, p-value = 0.627). These weak associations in the initial analysis suggest that a more sophisticated approach might be necessary to capture the complexities underlying these relationships. To address the complexity of these relationships, we employed machine learning techniques, as supported by literature.

Predictive model

After the hyperparameters of the aforementioned algorithms were optimised, a split dataset was used to evaluate each model’s performance. Using a variety of indicators, we assessed the algorithms’ ability to predict leptospirosis hotspots. These included cross-validation scores on the training data and the unseen hold-out dataset with the trained algorithm. Furthermore, sensitivity, accuracy, and F1-score were assessed to choose the best model for hotspot prediction.

Based on the analysis, the best model for predicting leptospirosis hotspot areas was the LGBM algorithm. Compared to RF (0.13) and SVM (0.14), LGBM achieved the highest precision (0.16), suggesting a more robust capacity to categorise actual hotspot locations correctly. When identifying all hotspots, LGBM did not show the highest value for sensitivity (0.43) (RF: 0.50, SVM: 0.41). However, LGBM obtained the greatest F1-score (0.23), balancing sensitivity and precision. This can be inferred that LGBM finds the best possible balance between minimising false positives, predicting locations as hotspots when they are not, and finding actual hotspots. Table 3 provides a thorough analysis of each model’s performance comparison.

The most significant factors influencing the model’s predictions can be identified through feature importance analysis. Each input variable is scored according to how well it predicts the target variable (hotspot areas) in the model. A higher score denotes a more substantial influence on hotspot prediction. Notably, only the LGBM and RF algorithms allow feature-importance computation. Rainfall was found to be the most important factor by LGBM, followed by LST (Fig. 3). On the other hand, RF showed the opposite significance hierarchy.