SatHealth
收藏资源简介:
SatHealth是一个多模态公共卫生数据集,结合了基于卫星的环境数据、卫星图像、从医疗索赔估计的全面疾病流行率以及社会健康决定因素(SDoH)指标。数据集包括来自Google Maps的超过40万张航拍卫星图像,每张图像覆盖约500米宽的方形区域。此外,我们使用来自MarketScan数据库的医疗索赔数据来估计所有疾病的区域流行率。至于SDoH,我们使用美国社区调查(ACS)的美国人口普查数据计算得出的社会剥夺指数(SDI)。我们还设计了一个多模态融合框架,以无缝地整合来自SatHealth的各种异构多模态环境数据源,并提供用户友好的区域环境嵌入,以便进行下游分析和后续研究。我们首先通过统计分析验证和量化了环境-疾病关系,反映了城乡健康状况的差异。之后,我们在两个临床任务上使用了数据集:区域公共卫生建模(例如,基于环境数据预测区域SDI分数和疾病流行率)和个性化疾病风险预测(例如,使用环境数据增强个人疾病风险预测)。实验结果表明,生活环境信息可以显著提高AI模型的性能和时空泛化能力。最后,我们部署了一个基于Web的应用程序,用户可以探索和访问SatHealth数据以及区域嵌入向量。我们的区域嵌入可以插入到任何具有地理空间信息的临床AI中,为将环境因素纳入临床AI开发铺平了道路。我们从俄亥俄州的Ohio O-SUDDEn项目开始开发SatHealth。然而,我们使用的所有卫星数据都具有全球覆盖范围,MarketScan的患者级医疗索赔具有美国覆盖范围。因此,我们框架的环境因素处理流程可以轻松适应其他地区。我们还提供了GitHub上的代码,以便用户可以为不同兴趣领域创建数据和嵌入。我们也将逐步更新SatHealth以覆盖美国。
SatHealth is a multimodal public health dataset that integrates satellite-based environmental data, satellite imagery, comprehensive disease prevalence estimates derived from medical claims, and Social Determinants of Health (SDoH) metrics. The dataset includes over 400,000 aerial satellite images sourced from Google Maps, each covering a square region approximately 500 meters in width. Additionally, we utilized medical claims data from the MarketScan database to estimate regional disease prevalence across all conditions. For SDoH metrics, we computed the Social Deprivation Index (SDI) using U.S. Census data sourced from the American Community Survey (ACS). We further developed a multimodal fusion framework to seamlessly integrate diverse heterogeneous multimodal environmental data sources from SatHealth, and deliver user-friendly regional environmental embeddings to support downstream analysis and subsequent research. We first validated and quantified the environment-disease relationship through statistical analysis, which reveals disparities in health outcomes between urban and rural areas. Subsequently, we employed the dataset for two clinical tasks: regional public health modeling (e.g., predicting regional SDI scores and disease prevalence using environmental data) and personalized disease risk prediction (e.g., enhancing individual-level disease risk prediction with environmental data). Experimental results demonstrate that residential environment information can significantly improve the performance and spatiotemporal generalization capability of AI models. Finally, we deployed a web-based application that enables users to explore and access SatHealth data and regional embedding vectors. Our regional embeddings can be integrated into any clinical AI system equipped with geospatial information, paving the way for incorporating environmental factors into clinical AI development. We began developing SatHealth starting from the Ohio O-SUDDEn project in Ohio. However, all satellite data we utilized has global coverage, while the patient-level medical claims from MarketScan are restricted to the United States. Accordingly, the environmental factor processing pipeline of our framework can be readily adapted for other regions. We additionally provide open-source code on GitHub, allowing users to generate custom data and embeddings for various research domains of interest. We will also gradually expand the coverage of SatHealth to include the entire United States over time.



