In a recent study published in European Heart Journal, Researchers used state-of-the-art artificial intelligence (AI) techniques and analysis to correlate built environment characteristics identified in the AI ​​model with observed variance in coronary heart disease (CHD). We evaluated the gender. Specifically, the team used custom convolutional neural networks (CNNs), linear mixed effects models (LMEMs), and activation maps to identify associations of CHD-related characteristics and improve health outcomes at the census tract level. predicted.

The first-of-its-kind study used more than 530,000 Google Street Views (GSVs) to train and evaluate the model, and the results show that AI algorithms can create cities of the future that significantly reduce CHD burden. This suggests the possibility of designing.

Research: Artificial intelligence-based assessment of built environment and coronary artery disease prevalence from Google Street View. Image credit: yanto kw / Shutterstock

Assessing the potential of machine vision in CHD, GSV, and architectural environments

Coronary heart disease (CHD), also known as coronary artery disease (CAD), is a chronic, potentially life-threatening, non-communicable disease characterized by the deposition of plaque along the walls of the coronary arteries. The movement of oxygenated arteries is impeded or completely blocked. Blood to the heart. This buildup usually occurs gradually. It begins in childhood, progresses slowly, and eventually manifests as CHD later in life.

Despite decades of research and significant scientific advances in CHD risk detection and prevention, CHD remains the leading cause of heart disease-related mortality, especially in the United States, where it is estimated to be well over 50%. I am. Percentage of all cardiac mortality (approximately 400,000 deaths in 2020 alone). Recent evidence suggests that nontraditional risk factors such as race, income, culture, and education may play a significant role in the pathology of his CHD.

Environmental factors such as temperature and environmental pollution (noise and air) are also thought to be involved in the disease, but evidence to support these hypotheses remains lacking. A large repository of “built” urban features (buildings, green spaces, roads) enables place-specific he CHD risk detection and is the first step for policy-based medical interventions.

“Large-scale integrated assessments of the environment at the neighborhood level would facilitate rapid and complete assessments of CHD impacts. However, such data are scarce, in part due to the cost of neighborhood audits. Machine vision approaches such as Google Street View (GSV) have been the go-to approach for virtual neighborhood audits since their launch in 2007. It’s becoming more and more popular.”

Google Street View (GSV) is an imaging technology that powers many Google applications, including Google Maps and Google Earth. First published in 2007, this image dataset is primarily crowdsourced and displays interactive panoramas stitched together from VR photos, covering nearly 100% of the United States. Unrelated research exploiting the previously untapped potential of GSV has shown that ground truthing by humans can be improved, especially when using machine learning algorithms to classify and evaluate features of the built environment from his GSV images. A technology comparable in accuracy has been established.

About research

This study aims to assess the built environment across seven U.S. cities using GSV imagery and use these results to estimate CHD prevalence at the census tract level. Census tract-level data (2015-2016) from the 2018 Centers for Disease Control and Prevention (CDC) Population-Level Analysis and Community Estimates (PLACES) and the Robert Wood Johnson Foundation. This dataset consisted of American adults (18 years and older) with clinically confirmed angina or CHD status (positive or negative) from 789 census tracts across Bellevue, Washington. . Brownsville, Texas. Cleveland, Ohio; Denver, Colorado. Detroit, Michigan. Fremont, California. and Kansas City, Kansas.

Data collected as part of this study included de-identified demographic and socio-economic (DSE, age, race, gender, education level, income, occupation) factors and medical history. The image dataset consists of over 530,000 of his images from his GSV servers, with Google’s image classification intact. Imagine that data extraction was performed using a deep CNN (DCNN) called Places365CNN, which is the default extractor for the Places database. Considering the similarity between GSV and Places image feature classification, Places365CNN was found to be robust to the current research data extraction after training with over 10 million training images.

To investigate the association between raw DCNN extracted features (N = 4096) and region-level CHD prevalence, researchers applied three independent machine learning (ML) models: extra-tree regression (ET); and Random Forest Regression (RF) was trained and tested. ), and light gradient boost machine regressor (LGBM). All three models were cross-validated 10 times to improve the predictive accuracy and achieve robustness of the models. Following model training, multilevel regression analyzes using both linear fixed-effects and random-effects models were performed, with variables adjusted for age, gender, income, race, and education level.

“…we employed Grad-CAM technology to create a saliency map that highlights these salient features in the original GSV image. This process allows the CNN to determine what environmental features are We provide some explanation for what we believe is related to the prevalence of CHD in the neighborhood.”

Research results and key points

Geographic CHD prevalence was found to vary significantly, with Bellevue having a median prevalence of 4.70%, while Cleveland had a much higher prevalence of 8.70%. The features extracted by DCNN were found to contain more than 4,096 ML-classified features. A highlight of this study is that these extracted features alone explained 63% of the observed cross-regional variation in CHD prevalence.

“We found a small number of extreme values ​​in certain census tracts in Detroit and Cleveland that were underestimated by the model. CHD prevalence in these underrepresented census tracts was often 12% or more. When we examined the CNN-extracted features using t-SNE, we noticed a clustering of census tracts with similar CHD prevalence values.”

Multilevel modeling revealed that DSE factors (particularly age, gender, and education) were more accurate predictors of CHD than GSV characteristics. These results demonstrate that while GSV features may certainly be useful in highlighting specific built environment information related to CHD prevalence at the neighborhood level, this technique can be used to provide potential ways to It suggests that additional calculations (e.g. Grad-CAM method) are required before . Identifying built environment information.

“The results of our research provide a proof of concept that uses machine vision to identify features of urban networks that are associated with risk. It has the potential to quickly identify exposed areas and enable targeted interventions.”