New Diagnostic Tool Probes Feature-Level Bias in Dutch Health Algorithms

A team of Dutch researchers has developed a diagnostic framework that looks beneath the surface of algorithmic fairness metrics in healthcare, examining how individual data features drive biased predictions. The framework, called Feature-Level Bias Identification And Sensemaking (FL-BIAS), was applied to electronic health record data from Dutch general practitioners, and its findings suggest that standard fairness scores can mask deeper, potentially less favorable patterns hidden within the data.

The study, led by Cathleen S. Parsons and colleagues, addressed a persistent gap in the field. While most algorithmic fairness research in healthcare has concentrated on disparities in model performance, far less attention has been paid to the underlying data structures that may generate those disparities. High-level fairness metrics, the authors argue, often obscure the feature-level dynamics needed for a critical, context-aware assessment of fairness.

As a case study, the researchers conducted a secondary analysis of a retrospective cross-sectional cohort using electronic health records from Dutch general practitioners, linked with sociodemographic data from Statistics Netherlands. The dataset covered 112,872 patients, of whom 16.2 percent had a non-Western migration background. Hospitalization in the following year was modeled using Johns Hopkins Aggregated Diagnosis Groups (ADGs).

To test for disparities, the team trained logistic regression and XGBoost models on different subgroup datasets and evaluated them with standard fairness metrics. They then applied Kernel Shapley Values and Cohort Shapley methods to trace how individual features contributed to predictions and to errors, providing a feature-level view that conventional metrics cannot offer.

Exploratory analysis revealed significant differences in socioeconomic status (SES) and ADG distributions between Dutch and non-Western patient groups, although Multiple Correspondence Analysis showed minimal structural variation in the data overall. Mediation analysis indicated that the effect of migration background on hospitalization was largely mediated by SES, with the authors noting the possibility of unobserved confounding.

At the level of standard fairness metrics, the models appeared to show modest bias in favor of non-Western patients. Yet the deeper feature-level analyses told a more nuanced story, uncovering subgroup-specific patterns of variable importance that suggest potentially less favorable underlying conditions for some groups.

One striking example involved malignancy, recorded as ADG 32. This diagnosis group had a stronger predictive impact among non-Western patients but contributed less to false negatives compared with Dutch patients, possibly reflecting structural disparities in cancer diagnosis and care.

The authors emphasize that fairness in healthcare algorithms must be approached as a socio-technical challenge. Their findings highlight the need for contextual, multi-level evaluations of algorithmic bias, and they call for multidisciplinary collaboration to uncover the root causes embedded in the processes that generate health data.

Subject of Research: Feature-level bias identification in healthcare prediction algorithms using Dutch electronic health record data

Article Title: Refocusing algorithmic fairness on feature-level bias: A diagnostic approach using dutch EHR data

Article References: Parsons, C. S., Girwar, S.-A. M., Azimi, S., Spruit, M. R., & Haas, M. R. (2026). Refocusing algorithmic fairness on feature-level bias: A diagnostic approach using dutch EHR data. PLOS Digital Health, 5(10), e0001105. https://doi.org/10.1371/journal.pdig.0001105

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001105

Keywords: algorithmic fairness, healthcare AI, electronic health records, FL-BIAS framework, Shapley values, XGBoost, logistic regression, socioeconomic status, migration background, hospitalization prediction, diagnosis groups, socio-technical bias