# Spatial prediction of the probability of liver fluke infection using a geographic weighted regression (GWR) model in waterways connecting the Mekong River, Sakon Nakhon of Thailand

## TL;DR

This study uses a spatial model to predict liver fluke infections in water sources in Thailand, showing improved accuracy with a hexagonal grid approach.

## Contribution

The study introduces a novel spatial prediction model using a geographic-weighted regression with hexagonal grids for liver fluke infection prediction.

## Key findings

- Hexagonal grid models outperformed point-based models with R2 values up to 58.7% and RMSE reductions of 77.1%.
- Soil drainage and road proximity were identified as key factors influencing infection probability.
- The most accurate model predicted over 95% of grids with high accuracy.

## Abstract

Liver flukes (Opisthorchis viverrine, OV) infections in water sources continue to persist in Sakon Nakhon Province, which is linked to the Mekong River. The agency's traditional infection data comprises the locations of infected water sources. However, this data is insufficient for developing a predictive model for infections within the sub-basin. When analyzed alongside independent variables, represented as identical points, it lacks the necessary information to generate a trend line that produces a reliable coefficient. This study implemented a spatial model that integrates a geographic-weighted regression (GWR) framework with appropriate weighting as a prototype. This approach improves the selection of independent variables by shifting from a point-based methodology to a weighted hexagonal grid.

A set of eight independent variables land use, soil drainage, road network, water sources, streamlines, surface temperature, NDMI (Normalized Difference Moisture Index), and NDVI (Normalized Difference Vegetation Index) was initially weighted. This study developed three linear models based on the Geographically Weighted Regression (GWR) model. It demonstrates the advantages of utilizing a hexagonal grid instead of a point grid. The three alternative models were tested with various independent variables and employed a factor-by-factor averaging approach, which necessitates the hexagonal grid size as a counterweight to ensure fairness across the entire grid, rather than relying solely on point data. A mathematical model was developed to calculate the average of each factor in order to achieve equality across a hexagonal grid area. Subsequently, the correlation was tested, and the alternative models were grouped. The resulting dendrogram includes three models.

The results of the GWR comparison test were derived from both infected and hexagonal water source data. Models constructed from hexagonal grids consistently outperformed all alternative models, with R2 values improving to 58.7 %, 41.1 %, and 53.2 % for Model-1, Model-2, and Model-3, respectively. The RMSE also showed significant improvement, decreasing to 77.1 %, 60.2 %, and 67.2 %. Additionally, the model's accuracy was evaluated using AUC values of 0.725, 0.652, and 0.707, indicating that the developed model can effectively predict water source infections. Model-1 emerged as the most representative across all tests, incorporating soil drainage factors and road proximity as key influences on water source infection. Finally, the results are presented as infection prediction maps for each grid, highlighting areas of both overestimation and underestimation. The most accurate prediction model identified that over 95 % of grids had a high degree of accuracy. This study is anticipated to be applicable to infections caused by other water-mediated parasites.