2026 ECNP Congress

PS02-2036

Regional inequities in AI models for psychiatric diagnosis

Alissa Valentine — H. Coupland, P. Eike, E. Beaman, M. Osler, A. Feragen, M. Ganz · University of Copenhagen, Computer Science, Copenhagen, Denmark

2026 ECNP Congress

Background

The prevalence of psychiatric disorders is increasing globally. Major depressive disorder (MDD) is the costliest [1], contributing to growing strains on healthcare systems. In Denmark, this is reflected in long waiting times for psychiatric services, wherein rural regions experience the longest wait [2]. Artificial intelligence (AI) is often considered a tool to identify high-risk patient groups and reallocate resources. However, delays in obtaining care may lead to lower apparent prevalences of MDD in rural areas, such that models trained on nationwide data risk perpetuating regional disparities in mental healthcare. As such, this study is the first to use Denmark’s large-scale, high-quality registry data to explore regional heterogeneity in the prevalence of MDD and evaluate regional disparities in the performance of AI-driven MDD prediction models.

Methods

We leveraged Denmark’s nationwide registry data to create a cohort composed of individuals diagnosed with MDD between 2000-2018 (N=195,419) and a healthy reference group (N=168,224). Patient-level information derived from Danish register data included eight sociodemographic and five socioeconomic variables.

First, we conducted an observational association analysis using logistic regression to estimate the association between the likelihood of receiving an MDD diagnosis and living in each of Denmark’s five regions (North Jutland, Middle Jutland, South Denmark, Zealand, and the Capital Region) when controlling for sociodemographic and socioeconomic variables.

Second, we developed an MDD prediction model using tabular transformer architecture. The model was fine-tuned on nationwide data, then fine-tuned on region-specific data to explore the generalizability of ‘regionalized’ models. Model performance and fairness were evaluated using AUROC, accuracy, Brier score (BS), and calibration, including subgroup analyses by region and immigration status.

Results

The logistic regression revealed significant regional differences in the odds of receiving an MDD diagnosis when controlling for sociodemographic and socioeconomic variables, confirming regional heterogeneity.

When the MDD prediction model was fine-tuned on nationwide data, the greatest performance disparities were seen in the most rural region, North Jutland. North Jutland demonstrated the poorest reliability (BS=0.23) and the lowest AUROC (0.70). The model more frequently predicted a lower risk of MDD diagnosis in North Jutland than other regions, despite good calibration (0.006). The model showed similar performance for immigrants and people of Danish origin (AUROC= 0.71, accuracy=0.66, BS=0.22, calibration<0.005), but poorer performance for the descendants of immigrants (AUROC= 0.68, accuracy=0.63, BS=0.23, calibration<0.001).

Lastly, when fine-tuned on region-specific data, the model showed un-reciprocal generalizability across other regions. Whilst fine-tuning on nationwide data yielded the best performance for each region.

Conclusion

This work sheds light on the risks of deploying nationwide AI tools for MDD prediction, particularly for those living in rural regions, which were predicted to have lower risk. This may be the result of longer waiting times for psychiatric care in this region, demonstrating how AI models reproduce the biases present in their training data, leading to inequities in model performance. As such, fairness evaluations are critical to high-risk domains like psychiatry. Although these findings are focused on Denmark, they are broadly applicable to other healthcare systems with regional differences in insurance or healthcare access.

References

  1. Institute for Health Metrics and Evaluation (IHME), 2025. Global Burden of Disease Study Results. Available from: https://vizhub.healthdata.org/gbd-results. [Accessed: 13 January 2026]
  2. Jyllands-Posten, 2024. In some places, patients wait up to three years to see a psychiatrist – and nine months to get an appointment with an ophthalmologist. Available from: https://jyllands-posten.dk/indland/ECE17081920/nogle-steder-venter-patienter-op-til-tre-aar-paa-at-komme-til-psykiater-og-ni-maaneder-paa-at-faa-tid-hos-en-oejenlaege. [Accessed: 13 January 2026]