Analyzing Diabetics for Food Access Training on the Map with CBAM
Analyzing Diabetics for Food Access Training on the Map with CBAM
复制标题
DOI：
10.1109/wiiat50758.2020.00072
复制
发表时间：
2020-12
期刊：
影响因子：
--
通讯作者：
Huaze Xie;Yuanyuan Wang;Yukiko Kawai
中科院分区：
文献类型：
--
作者：
Huaze Xie;Yuanyuan Wang;Yukiko Kawai
通过医学文本数据挖掘得到的知识图可以用来存储疾病-症状关系。医学研究人员期望直接从电子病历中提取高质量的疾病属性关系，以满足临床决策。在本研究中，我们将影响2型糖尿病的因素与地理信息特征相关联，从人群的角度观察糖尿病患者的食物获取情况。本文以糖尿病患者的临床指标和食品购买习惯为依据，结合相关信息，探讨不合理饮食在某些地区的高发病率和诱因。我们选取了人类住区的地理数据，降低了时间和空间的复杂性，并通过Google Maps API中的食物获取构建了栅格数据的统计方法，以评估2012-2016年间美国2型糖尿病的公共记录。此外，我们使用基于注意力的卷积网络模型来适应和改进糖尿病特征的属性关系结构和空间分布。该模型的初始准确率对地理特征不敏感(平均准确率为49%，召回率为53%)，但通过加入CBAM模型来改进卷积核和栅格单元近似预测密度方法的性能，使得地标特征分类得到了显著的提高。最后，结合地理特征预测，对易感患者在食物获取方式的选择上提出建议。我们的地理信息栅格化方法也可以应用于地图上其他疾病特征的研究。
The knowledge graph obtained through medical text data mining can be used to store the disease-symptom relationship. Medical researchers expect to extract high-quality disease attribute relationships directly from electronic medical records to satisfy clinical decision-making. In our research, we correlated the influencing attributes in T2DM with geographic information features to observe the food acquisition of diabetic patients from a population perspective. In this article, we used clinical indicators and food purchase habits of diabetic patients with coordinated information to discuss the high incidence and incentives caused by an unreasonable diet in certain areas. We selected the geographic data of human settlements to reduce the complexity of time and space and constructed a statistical method of raster data through the food acquisition in the Google Maps API to evaluate the public records of type 2 diabetes in the United States from 2012 to 2016. Also, we use an attention-based convolutional network model to adapt and improve the attribute relationship structure of diabetes features and spatial distribution. The initial accuracy of the model is not sensitive to geographic features (the average accuracy is 49%, the recall rate is 53%), but by adding the CBAM model to improve the performance of the convolution kernel and raster unit approximate predictive density method, the placemark feature classification has been significantly improved. At last, combining the geographical feature prediction we can suggest for susceptible patients in food access method selection. Our geographic information rasterization method also might be applied to the study of other disease characteristics on the map.