Nationwide AI Models for Depression Prediction: Regional Generalizability and Culturally Sensitive Variable Selection

Abstract

The prevalence of psychiatric disorders is increasing globally. Major depressive disorder (MDD) is the most burdensome, contributing to growing strains on healthcare systems. Artificial intelligence (AI) for healthcare is deemed a potential solution but requires large-scale, high-quality data; therefore, we use Denmark as a case study with comprehensive register data and long psychiatric wait times. Although Denmark is often considered a homogeneous society, its sociodemographic and socioeconomic factors differ across the country, including immigration rates. There are also regional differences in the prevalence of psychiatric diagnoses. However, we have yet to understand how AI models account for such differences. We examine how regional and sociodemographic inequities affect MDD diagnosis and AI-based prediction. First, we conduct an observational analysis showing disparities in MDD diagnosis odds across regions of Denmark, as well as patient sociodemographic and socioeconomic factors. Second, we develop two MDD prediction models using tabular transformer architecture that differ in how they represent heritage and immigration-related information. We assess the performance of the two predictive models across the five administrative regions in Denmark, and their generalizability when trained on data from only a single region. We observe that model performance varies across Denmark’s five administrative regions, along with immigration status and heritage. Incorporating more nuanced sociodemographic and socioeconomic variables improves predictive performance across all regions and most sociodemographic groups. These findings illustrate the importance of fairness assessments in psychiatric AI and the need to account for regional, sociodemographic, and socioeconomic inequities when developing and deploying AI in high-risk healthcare settings.