This repo is based on another paper: https://github.com/adymaharana/ObesityDL
This is a project for the class: CSCI499 Artificial Intelligence for Social Good
Link to paper: paper
In this paper, we adopt a novel method for the task of predicting the prevalence of mental health issues in a city. By utilizing socioeconomic data, as well as built environmental features derived from satellite imagery, to predict the prevalence of mental health issues in a city, the method provides an alternative to telephone and in-person interviews. Through this method, we demonstrate that built environmental features do correlate with the prevalence of mental health issues. Our multi-modal model achieves a maximum of 0.86 R2 score for Intra-City predictions and a maximum of 0.78 R2 score for Multi-city and Cross-City predictions. Hence, our proposed method demonstrates a reliable, uniform method to predict mental health issues, and could be applied to other cities in the world that are not included in this study.
- Obtain census tract data per city selected
- Obtain shapefile per city
- Obtain socioeconomic features per census tract
- Obtain images from google static maps api
- Extract image features per image within census tract
- Mean all image features within census tract to extract 1 data point
- Combine socioeconomic features and extracted image features per census tract
- Regress to predict prevalence of mental health
Notebook | Function |
---|---|
Main.ipynb | EDA |
socioeconomic_features.ipynb | Extract socioeconomic features |
Download_imgs.ipynb | Download and Extract Satellite Images |
Model_eval.ipynb | Regression model results |
Trained on Memphis, Tested on Los Angeles:
Multi-city performance of Gradient Boosting Regression with Multi-modal features