<p>Type 2 diabetes has become an urban epidemic influenced by neighbourhood environments. However, conventional risk models focusing solely on individual factors fail to account for these neighbourhood influences and often require detailed patient data that may not be available. To address this gap, we developed an integrated approach combining machine learning and causal inference to map type 2 diabetes risk at the neighbourhood level. Using demographic, health, and socioeconomic data from 1,149 Census Tracts (CTs; the neighbourhood unit in this study) in a large metropolitan region, we trained seven machine learning models to identify neighbourhoods with high diabetes prevalence. Although neighbourhood-level diabetes data were available for this study area, our model’s high predictive accuracy on external validation data (area under the curve (AUC) = 0.95), particularly from a distinct geographical region, suggests potential utility for predicting diabetes risk in other Canadian regions or elsewhere where such data are unavailable, provided comparable covariates are available and the model is locally retrained and validated using spatially aware procedures. The top models achieved high recall (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(>90\%\)</EquationSource> </InlineEquation>) and AUC up to 0.96 on test data, indicating accurate identification of high-risk neighbourhoods with few missed high-risk areas. Survey-derived neighbourhood health indicators, including obesity rate, physical inactivity, and median age were strong predictors of diabetes prevalence. We then applied a Causal Forest approach to estimate conditional average treatment effects (CATE, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\tau\)</EquationSource> </InlineEquation>) for selected potentially modifiable factors and summarized the results with the mean <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\bar{\tau }\)</EquationSource> </InlineEquation>. Higher work stress (<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\bar{\tau }= 0.312\)</EquationSource> </InlineEquation>) and daily smoking (<InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(\bar{\tau }= 0.155\)</EquationSource> </InlineEquation>) were moderately associated with increased risk, whereas better mental health (<InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(\bar{\tau }\approx -1.1\)</EquationSource> </InlineEquation>) was protective, highlighting mental health as a priority for further evaluation, especially in neighbourhoods predicted to have high diabetes prevalence. These findings could help identify modifiable neighbourhood-level factors for local prevention efforts and inform equity-oriented planning in diverse urban populations. Prospective or quasi-experimental studies are needed to evaluate intervention effects. Our integrated machine-learning and causal framework lays the groundwork for precision public health, suggesting that modifiable neighbourhood factors may indicate diabetes risk when patient-level data are scarce. Furthermore, the pipeline is conceptually adaptable to other chronic diseases influenced by social and environmental determinants and may inform targeted prevention beyond type 2 diabetes, contingent on disease-specific feature sets and external validation.</p>

Mapping neighbourhood-level drivers of type 2 diabetes for precision public health using predictive and causal machine learning

期刊论文
出版模式：

开放获取
发表日期：

2026年1月5日

查看全文

Mohammad Noaeen,
Amirhosein Rostami,
Ibrahim Ghanem,
Olli Saarela,
Karim Keshavjee,
Jeffrey R. Brook,
Zahra Shakeri

摘要

Type 2 diabetes has become an urban epidemic influenced by neighbourhood environments. However, conventional risk models focusing solely on individual factors fail to account for these neighbourhood influences and often require detailed patient data that may not be available. To address this gap, we developed an integrated approach combining machine learning and causal inference to map type 2 diabetes risk at the neighbourhood level. Using demographic, health, and socioeconomic data from 1,149 Census Tracts (CTs; the neighbourhood unit in this study) in a large metropolitan region, we trained seven machine learning models to identify neighbourhoods with high diabetes prevalence. Although neighbourhood-level diabetes data were available for this study area, our model’s high predictive accuracy on external validation data (area under the curve (AUC) = 0.95), particularly from a distinct geographical region, suggests potential utility for predicting diabetes risk in other Canadian regions or elsewhere where such data are unavailable, provided comparable covariates are available and the model is locally retrained and validated using spatially aware procedures. The top models achieved high recall ( \(>90\%\) ) and AUC up to 0.96 on test data, indicating accurate identification of high-risk neighbourhoods with few missed high-risk areas. Survey-derived neighbourhood health indicators, including obesity rate, physical inactivity, and median age were strong predictors of diabetes prevalence. We then applied a Causal Forest approach to estimate conditional average treatment effects (CATE, \(\tau\) ) for selected potentially modifiable factors and summarized the results with the mean \(\bar{\tau }\) . Higher work stress ( \(\bar{\tau }= 0.312\) ) and daily smoking ( \(\bar{\tau }= 0.155\) ) were moderately associated with increased risk, whereas better mental health ( \(\bar{\tau }\approx -1.1\) ) was protective, highlighting mental health as a priority for further evaluation, especially in neighbourhoods predicted to have high diabetes prevalence. These findings could help identify modifiable neighbourhood-level factors for local prevention efforts and inform equity-oriented planning in diverse urban populations. Prospective or quasi-experimental studies are needed to evaluate intervention effects. Our integrated machine-learning and causal framework lays the groundwork for precision public health, suggesting that modifiable neighbourhood factors may indicate diabetes risk when patient-level data are scarce. Furthermore, the pipeline is conceptually adaptable to other chronic diseases influenced by social and environmental determinants and may inform targeted prevention beyond type 2 diabetes, contingent on disease-specific feature sets and external validation.