Machine Learning Based Modeling to Estimate Potential Distribution of Current Remerging hMPV in Asia Regions

1. Introduction

The first occurrences of hMPV were firstly reported by Hoogen et al. (2001) following a suspected case in the Netherland, a country in the temperate regions. A few years after that, Rao et al. (2004) reported a case in Pune, India. It indicates a virus's potential to spread over a distance. After that, the occurrences of hMPV have been reported in several locations around the world, ranging from Canada, the United Kingdom, and Australia (Nissen et al. 2002). This situation indicates the ability of the virus to spread and be regulated by environmental variables, including climate variables. Since the virus has the ability to adapt to the new locations, then it requires an approach to estimate its potential distribution. A method used widely to study the potential distribution is using machine learning.

Recently, numerous machine learning-based modeling has been used to estimate the disease potential distribution, including viruses and COVID-19 (He et al. 2021, Pramanik et al. 2021, Pan et al. 2024). As an example, Yu et al. (2019) have used the model to predict the potential geographical distribution of the H7N9 avian influenza virus in several locations in China where known as high-risk zones for H7N9 occurrences. Following the progress and advantages of the distribution model, it is important to elaborate the bioclimatic variables responsible for present distribution as well as the potential distribution of hMPV across Asian regions. Hence, the present study used the machine learning-based potential distribution modeling approach to determine the risk areas of hMPV and climatic sensitivity influencing the hMPV occurrences based on reports and existing data. The machine learning then is considered the only efficient approach for controlling and preventing the hMPV potential spreads and also to identify the determinant bioclimatic risk of hMPV.

2. Materials and Methods

The research was conducted on several Asian countries where hMPV has been reported recently. The modeling methodology followed recent study (Ren et al. 2020), including hMPV occurrence, bioclimatic variables, and model evaluation.

2.1. Study Site

The study was implemented in several Asian countries where hMPV has been reported recently. Deep analysis was implemented in the regions that have the most hMPV reports, and the data can be used for modeling.

2.2. hMPV Occurrences

HMPV just reemerged recently, and this has limited the data collection. Most of the data and hMPV occurrences were based on the news where the geographical locations of hMPV occurrences were mentioned. The geographical locations then were tracked to record the geocoordinates. If the locations were a city, then a geocoordinate of the center of the city were retrieved. The geocoordinates were tabulated as longitude and latitude and saved as CSV formats for further analyses. The threshold for hMPV occurrences data recording was 9^th January 2025.

2.3. Bioclimatic Variables

This study, in accordance with Dong et al. (2023) and Arshad et al. (2022), incorporated a variety of 19 bioclimatic variables (Table 1) range from Bio 1 to Bio 19 and retrieved from the WordClim global climate database (www.worldclim.org), specifically version 2.0 [33]. This database is widely used including in Asia (Rana et al. 2017) and has been used extensively to develop potential distribution model (Khanum et al. 2013). Grids were used to visualize the bioclimatic variables. These grids have a spatial resolution of roughly 25 m² at the equator and less elsewhere (Fick et al. 2013).

In order to produce an accurate and unbiased potential distribution model, those bioclimatic variables were selected based on the selection and exploration of bioclimatic variables having a major contribution. Each bioclimatic variable's contribution to the final model was assessed using Jackknife analysis. If those bioclimatic variables having a low percentage contribution or with values less than 6%, then those bioclimatic variables were considered having less impacts on distribution model (Wei et al. 2018).

2.4. Multicollinearity Test

The distribution model was developed upon the association of hMPV occurrences with 19 bioclimatic variables. To gain the unbiased distribution model, the bioclimatic variables needed to be checked for their possibility of collinearity among variables. Here in this study, by using Pearson's correlation tests (Préau et al. 2018) on 19 bioclimatic variables with ranges of Bio 1 – Bio 19, a multicollinearity test was carried out to create a model with improved performance with fewer variables and to prevent collinearity between the variable. Highly cross-correlated variables with r² values larger than 0.8 were eliminated, while variables with r² values less than 0.8 were retained for additional examination in the context of geographical distribution modeling. When multicollinearity arises, a variable's predictive capacity is unstable and unreliable because it has a strong correlation with other variables in the model (As’ary et al. 2022). The chosen bioclimatic variables to be employed were Bio 1, 9, and 15 based on the results of the multicollinearity test (Table 1).

2.5. hMPV Potential Distribution Modeling

In order to produce estimated suitable distribution areas of hMPV across Asian regions, this study uses machine learning (Dolci and Peruzzi 2022) within R applications version 3.6.3 (Mao et al. 2021). The R applications used to support the machine learning analyses and create the hMPV suitability models include the"dismo" and "sp" packages (Khan et al. 2022), the "rgdal" package (Bivand 2022), and the "raster" library (Lemenkova 2020). For developing the suitable distribution model, Bio 1, 9, 4, and 15 were included in the model's inputs.

A Jackknife test was employed within the model to identify the contribution and influence of each bioclimatic variable on the hMPV potential distribution areas model (Promnikorn et al. 2019). Following that, the analysis findings from machine learning models predicting hMPV potential distribution area ranges were imported into GIS for visualization and further spatial analysis (Hijmans et al. 2012). The QGIS platform version 2.16 was used in this study. Wei et al. (2018) state that there are five suitability levels for prospective potential distribution areas on the used machine learning model map: for very low, low, moderate, high, and very high hMPV potential.

3. Results

3.1. Regions with Reported hMPV Cases Overlaid with Temperatures and Precipitation

As of 9^th January 2025, several regions or countries have reported hMPV cases, as can be seen in Figure 1. The countries with reported hMPV were mostly still experiencing winter seasons. In northern and high-latitude regions, hMPV is occurring in temperate regions with temperatures lower than 25°C. While hMPV were also occurring in the southern regions near the equator in tropical countries with temperature ranges of 25–30 °C. In comparing with the precipitation layers, the regions impacted by hMPV are overlapped with precipitation, as can be seen in the southern regions near the equator. In this region, the presence of hMPV is coincided with the precipitation ranges of 8–20 mm. At the same time, the northern and temperate regions has no precipitation. The Indian subcontinent has reported the presence of hMPV at least in the six locations. The first group, consisting of three locations, was in the higher latitude closed to the temperate regions, and another three hMPV occurrences were reported in the lower latitude latitude closed to the tropicsl regions. While no further information or data, especially in regions at lower latitudes and near the equator in Southeast Asia, whether the hMPV was due to the local transmission or transmitted from a person who has travelled from the impacted regions.

3.2. Contribution of Bioclimatic Variables to the Suitable Distribution Areas of hMPV

The multicollinearity tests have resulted in selecting three determinant bioclimatic variables, including Bio 1, 9, and 15, as selected variables. Those selected bioclimatic variables have varied contributions to the area suitable for hMPV distributions in studied areas, as can be seen in Table 2. From the results, the mean temperature of the driest quarter, followed by the precipitation seasonality and the annual mean temperature, were the climate variables that have the most contribution to the suitable areas of hMPV currently in the studied areas with the values of 50.45%, 35.01%, and 14.53%.

3.3. Response Curves of Bioclimatic Variables

The response curves show the relationships between the probability of occurrence and the suitability level of hMPV with each selected bioclimatic variable (Figure 2). The response curves of the suitable distribution areas of hMPV were drawn for the selected three bioclimatic variables. The existence probability of each bioclimatic variable to hMPV showed that with the increase in the bioclimatic variable value, the occurrence probability of suitable distribution areas for hMPV showed a trend of first increasing and then decreasing rapidly. In this study, the range of bioclimatic variables when the occurrence probability is greater than 0.6 was used to represent the bioclimatic characteristics of the hMPV suitable distribution areas. The bioclimatic characteristics of the distribution area of hMPV are as follows: the annual mean temperature (Bio 1) is limited at 27.0 °C, the mean temperature of the driest quarter (Bio 9) is limited at 24.0 °C, and the precipitation seasonality (Bio 15) is limited at 10.

3.4. Potential hMPV Distribution Areas Under Current Climate

Potential hMPV distribution areas under current climate in the India subcontinent region have been modeled using data from 9^th January 2025, as can be seen in Figure 3. Since there are several occurrences, then potential distribution analyses can be implemented. In India there are at least three potential distribution areas in the northern and southern parts of this subcontinent region. The potential areas were divided into five classes of suitability ranging from very low to very high. The first large high-risk areas are predicted in central and northern parts with north latitude of 16⁰-22⁰ and east longitude of 75⁰-79⁰. In these areas, very high areas were observed with north latitude of 17⁰-19⁰ and east longitude of 76⁰-78⁰, indicating preference areas toward the east direction. The second areas were observed in very high north latitude of 21⁰-23⁰ and east longitude of 72⁰-75⁰, indicating preferences in temperate regions. The third area was observed in the southern parts with north latitude of 13⁰-17⁰ and east longitude of 73⁰-75⁰, indicating potential preferences in tropical regions.

5. Limitations of the Study

The limitation of this study is represented in the low AUC value. This value is related to the numbers of samples or the presences of the hMPV cases recorded. Most of the cases used in this study are recorded from news, and it is very limited. As a preliminary study, this study only recorded limited hMPV cases and do not distinguish between hMPV local transmission or not. Despite the limited records, this study at least has succeeded in providing a baseline for hMPV potential distribution modeling. For future study and to deal with the limitations of the study, there is plenty of opportunity for improvement. It ranges from a comparison study among modeling algorithms and further study using climate change scenarios following the latest CMIP 6 scenarios.

6. Conclusions

As modeled in this study, this study has succeeded in estimating and even providing scientific evidence of potential spreads of hMPV. It has provided an insight that hMPV has restricted its spread to the center of the subcontinent or has a preference for the higher latitude or temperate regions. While there is still a patch in the southern parts of the subcontinent, or lower latitude, that indicates potential of hMPV preferences in the tropical regions. These potential areas indicate the risks of hMPV adapting and spreading in the tropical regions, and this can deliver significant health risks, considering several countries presented in the Asia and Southeast Asia regions are tropical countries.