Fairness and bias correction in machine learning for depression prediction: results from four study populations

Aug 5, 2026

Stigma and inequality in mental healthcare are amplified when scientific data are uneven across populations and machine learning models are trained without properly accounting for subgroup differences. A systematic study analyzes bias in depression-prediction ML models across four datasets spanning distinct countries and populations. Standard ML approaches show regularly biased behavior, while mitigation methods, including a post-hoc approach, can reduce unfair bias levels. No single model achieves equality of outcomes, highlighting fairness analysis and transparent debiasing reporting.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

AccessFree to download

Introduction

Depression risk and exposome-based prediction

Algorithmic bias and protected attributes

Goal of the study and datasets

Bias assessment and mitigation

arXiv :2211 .05321v3 [ cs .LG] 26 Oct 2023 Fairness and bias correction in machine learning for depression prediction: results from four study populations Vien Ngoc Dang 1,* , Anna Cascarano1 , Rosa H. Mulder2 , Charlotte Cecil3 , Maria A. Zuluaga4 , Jernimo Hernndez-Gonzlez5 , and Karim Lekadir 1 1 Departament de Matemtiques i Informtica, Facultat de Matemtiques i Informtica, Universitat de Barcelona, Spain 2 Department of Pediatrics, Erasmus University Rotterdam, the Netherlands 4 Department of Child and Adolescent Psychiatry/Psychology, ErasmusMC-Sophia, Rotterdam, the Netherlands 4 Data Science Department, EURECOM, Sophia Antipolis, France 5 Departament d’Informtica, Matemtica Aplicada i Estad´ıstica, Universitat de Girona, Spain * [dangn@ub.edu](dangn@ub.edu) ABSTRACT A significant level of stigma and inequality exists in mental healthcare, especially in under-served populations. Inequalities are reflected in the data collected for scientific purposes. When not properly accounted for, machine learning (ML) models leart from data can reinforce these structural inequalities or biases. Here, we present a systematic study of bias in ML models designed to predict depression in four different case studies covering different countries and populations. We find that standard ML approaches show regularly biased behaviors. We also show that mitigation techniques, both standard and our own post-hoc method, can be effective in reducing the level of unfair bias. No single best ML model for depression prediction provides equality of outcomes. This emphasizes the importance of analyzing fairness during model selection and transparent reporting about the impact of debiasing interventions. Finally, we provide practical recommendations to develop bias-aware ML models for depression risk prediction. Introduction Depression is a leading cause of disability worldwide, a major risk factor for the global burden of disease, and can even lead to suicide [1, 2] . Taking into account that the global prevalence of depression increased by 25% during the COVID-19 outbreak [3], being able to identify those individuals at risk would be of great value to enable the application of personalized preventive measures. To this end, it is necessary to characterize the factors leading up to the development of depression. Research to date points to the importance of both genetic and environmental factors (as well as their interaction) in the etiology of depression [4, 5] . Furthermore, environmental factors have been shown to co-occur, exerting cumulative effects on depression risk. The totality of these environmental influences is often referred to as the exposome and includes environmental and lifestyle factors, as well as traumatic life events [6] . Exposome data does not only provide an alternative picture, it is also relatively inexpensive and easy to acquire, typically through questionnaires [7] . Motivated by the successful application of machine learning (ML) in different contexts of the medical domain, there is a spike in the use of ML for the detection, diagnosis, and treatment of depression [8–10] . Specifically, supervised ML methods are commonly used to learn predictive models from historical data, which are then applied to predict possible illness development in new cases and patients. Recently, concerns have been raised about algorithmic bias [8] and the undesirable ability of ML models of amplifying unfair behaviors masked in past practice, that is, in the data used for model learning. The term “algorithmic bias” refers to differences in the predictive power of models when applied to different subgroups of the population. These differences are particularly worrying if they are found when the subgroups are determined according to some protected attribute such as ethnicity, sex, or age. The subgroup that is adversely impacted by the bias of the ML model is known as the unprivileged group, and the subgroup that is unfairly be

FAQs

Why is fairness important in machine learning for depression prediction?

Bias can reinforce structural inequalities when models learn from data that differ across populations, leading to unequal predictive power across protected subgroups.

What data sources were used to study bias?

The study evaluates depression prediction using four public datasets: LONGSCAN, FUUS, NHANES, and the UK Biobank (UKB).

Do mitigation techniques eliminate unfair bias completely?

Mitigation methods can reduce unfair bias, but the results show that no single best ML model guarantees equal outcomes across groups.

Aug 5, 2026

Upload by Ava Thompson