geographically weighted regression

Cancer remains a critical public health issue in India, with rising cases of breast cancer and cervical cancer. Accurate predictions and spatial analysis of cancer incidence are essential for shaping prevention strategies and targeting interventions in high-risk regions.

This study utilized a big data framework employing machine learning techniques from the SparkML library to predict cancer cases and analyze spatial distributions across Indian states from 2016 to 2021. Three machine learning models used Random Forest Regressor, Gradient Boosting Regressor, and Geographically Weighted Regression (GWR) were applied to the dataset. Spatial autocorrelation analysis used Moran’s I statistic to identify clustering patterns.

The spatial analysis revealed significant clustering of cancer cases, particularly in 2020, with a z-score of 2.23, a p-value of 0.02, and a Moran’s index of 0.15. Among the machine learning models, GWR achieved a predictive accuracy of 98% for both breast cancer and cervical cancer, while the Random Forest Regressor and Gradient Boosting Regressor achieved 95% and 97% accuracy, respectively, over the six-year period. Gradient Boosting outperformed other models in identifying key predictors and ensuring high predictive accuracy.

The findings highlight the efficacy of Gradient Boosting and GWR in predicting cancer incidence and analyzing spatial patterns. These models provide critical insights into cancer clustering and risk factors, supporting the development of targeted prevention strategies and policy interventions for high-risk regions in India. The results emphasize the utility of machine learning techniques in public health research and cancer control.

In 2022, gastric and breast cancer had a mortality rate of 6.8 cases, ranking fourth despite intervention efforts. A 2020 study by the International Agency for Research on Cancer found global variations in incidence rates. This study examines key risk factors and preventive measures [1-2].

This applied ecological study uses the MAGWGPRS (Multivariate Adaptive Geographically Weighted Generalized Poisson Regression Spline) model, integrating MARS and GWGPR, to analyze cancer registry data. The model identifies geographic variations and hotspots in cancer risk. Data sources include pathology reports, death records, biopsy data, and a non-communicable disease risk factor survey. The dataset comprises patient age, location, cancer case counts, and relevant risk factors. Analysis is conducted using R, with ArcGIS for map visualization.

According to the International Agency for Research on Cancer, key risk factors for stomach cancer include obesity, smoking, physical inactivity, poor nutrition, age, and population density. The MAGWGPRS model, a Geographically Weighted Model, identifies regional variations in these factors by weighting observations based on distance using a kernel function and optimizing the model with the GCV criterion. Our analysis highlights vegetable consumption, smoking, low physical activity, and age as the primary determinants of gastric cancer risk.

Our model identifies vegetable consumption, smoking, low physical activity, and aging as significant risk factors for gastric cancer. Further research is needed to refine obesity risk based on BMI criteria. The MAGWGPRS model is a valuable tool for identifying high-risk regions, enabling targeted interventions and prioritizing key risk factors across diverse geographic areas.

Ordinary linear regression (OLR) is one of the most common statistical techniques used in determining the association between the outcome variable and its related factors. This method determines the association that is assumed to be true for the whole study area – a global association. In the field of public health and social sciences, this assumption is not always true, especially when it is known that the relationship between variables varies across the study area. Therefore, in such a scenario, an OLR should be calibrated in a way to account for this spatial variability. In this paper, we demonstrate use of the geographically weighted regression (GWR) method to account for spatial heterogeneity. In GWR, local models are reported in which association varies according to the location accounting for the local variation in variables. This technique utilizes geographical weights in determining association between the outcome variable and its related factors. These geographical weights are relatively large (i.e. close to 1) for observations located near regression point than for the observations located farther from the regression point. In this paper, we demonstrated the application of GWR and its comparison with OLR using demographic and health survey (DHS) data from Tanzania. Here we have focused on determining the association between percentages of acute respiratory infection (ARI) in children with its related factors. From OLR, we found that the percentage of female with higher education had the largest significant association with ARI (P = 0.027). On the other hand, result from the GWR returned coefficients varying from -0.15 to -0.01 (P < 0.001) over the study area in contrast to the global coefficient from OLR model. We advocate that identifying significant spatially-varying association will help policymaker to recognize the local areas of interest and design targeted interventions.

Although intellectual disability (ID) is a common disability in Iran, there is no investigation on the spatial distribution pattern of these patients in national level and the spatial maps for recognition the areas with higher prevalence of IDs and local neighborhoods of these regions or effect of socio-demographic factor on this scattering is not still available. This proposition motivated us to assess the population with ID in our country.In a cross-sectional study, we applied Moran’s Index (Moran’s I) which includes information about the strength of the neighboring association between counties, as global univariate distribution assessment. A geographically weighted regression was used to explore relation between ID patient’s prevalence and some socio-demographic factors (migration and illiteracy rate, physician number (PN)/10,000 people and health-care centers (HCCs)/10,000 people).We found that spatial clusters of ID patients exist among Iran counties (Moran’s I = 0.36,P < 0.01) and in a rural area population groups (Moran’s I = 0.20,P < 0.01). Further, we detected spatial associations between ID patients and all of our investigated socio-demographic factors in national scale. In rural areas, illiteracy has high association with ID especially in the south region of Iran. Urban area has random pattern of ID patients both within and between the Iran counties (Moran’s I = 0.01,P > 0.3).According to the results, our Initial hypothesis about the existence ofspatial clusters in distribution ofpeople with ID in Iran was proven. Spatial autocorrelation between migration and illiteracy rate and prevalence of patients with ID was shown and was in agreement with our hypothesis. However, our supposition that the prevalence should have inverse relationship with PN and HCC was rejected.