Marcel Haas
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Marcel Haas has a total of 75 co-authors. Here are the top co-authors by number of shared publications:
- Marco Spruit
Leiden University
11 shared publications - Bram van Dijk
Leiden University Medical Center, The Netherlands
5 shared publications - Jim Achterberg
Leiden University Medical Center
5 shared publications - Rimke Cathelijne Vos
Public Health and Primary Care/Health Campus The Hague, Leiden University Medical Center, Leiden, Zuid Holland, Netherlands
5 shared publications - Saif Ul Islam
University of Warwick, UK
5 shared publications - Carsten Maple
University of Warwick
4 shared publications
Latest Publications
- Open Access
- 26 October 2026
- Published by Public Library of Science (PLoS) in PLOS Digital Health
- Vol. 5 (10) , e0001105
Refocusing algorithmic fairness on feature-level bias: A diagnostic approach using dutch EHR data
While algorithmic fairness research in healthcare has predominantly focused on disparities in model performance, less attention has been given to the underlying data structures that may drive such disparities. High-level fairness metrics often obscure the deeper feature-level dynamics necessary for a critical and context-aware assessment of fairness. To address this gap, we propose and apply a diagnostic framework termed Feature-Level Bias Identification And Sensemaking (FL-BIAS). As a case study, we conducted a secondary analysis of a retrospective cross-sectional cohort study using electronic health record data from Dutch general practitioners, linked with sociodemographic data from Statistics Netherlands. The dataset included 112,872 patients, of whom 16.2% had a non-Western migration background. Hospitalization in the following year was modeled using Johns Hopkins Aggregated Diagnosis Groups (ADGs). We trained logistic regression and XGBoost models on different subgroup datasets to evaluate performance disparities using fairness metrics. To analyze feature-level contributions to predictions and errors, we applied Kernel Shapley Values and Cohort Shapley methods. Exploratory analysis revealed significant differences in socioeconomic status (SES) and ADG distributions between Dutch and non-Western groups, though Multiple Correspondence Analysis showed minimal structural variation. Mediation analysis indicated that the effect of migration background on hospitalization was largely mediated by SES, with potential unobserved confounding. While standard fairness metrics indicated modest bias in favor of non-Western patients, deeper feature-level analyses revealed subgroup-specific patterns of variable importance that suggest potentially less favorable underlying conditions. For instance, malignancy (ADG 32) had a stronger predictive impact among non-Western patients but contributed less to false negatives compared to Dutch patients, possibly reflecting structural disparities in cancer diagnosis and care. These findings highlight the need for contextual, multi-level evaluations of algorithmic bias. Fairness in healthcare algorithms must be approached as a socio-technical challenge, requiring multidisciplinary collaboration to uncover root causes in the processes that generate the data.
- 26 October 2026
- Published by Public Library of Science (PLoS) in PLOS Digital Health
- Vol. 5 (10) , e0001105
Open AccessRefocusing algorithmic fairness on feature-level bias: A diagnostic approach using dutch EHR data
While algorithmic fairness research in healthcare has predominantly focused on disparities in model performance, less attention has been given to the underlying data structures that may drive such disparities. High-level fairness metrics often obscure the deeper feature-level dynamics necessary for a critical and context-aware assessment of fairness. To address this gap, we propose and apply a diagnostic framework termed Feature-Level Bias Identification And Sensemaking (FL-BIAS). As a case study, we conducted a secondary analysis of a retrospective cross-sectional cohort study using electronic health record data from Dutch general practitioners, linked with sociodemographic data from Statistics Netherlands. The dataset included 112,872 patients, of whom 16.2% had a non-Western migration background. Hospitalization in the following year was modeled using Johns Hopkins Aggregated Diagnosis Groups (ADGs). We trained logistic regression and XGBoost models on different subgroup datasets to evaluate performance disparities using fairness metrics. To analyze feature-level contributions to predictions and errors, we applied Kernel Shapley Values and Cohort Shapley methods. Exploratory analysis revealed significant differences in socioeconomic status (SES) and ADG distributions between Dutch and non-Western groups, though Multiple Correspondence Analysis showed minimal structural variation. Mediation analysis indicated that the effect of migration background on hospitalization was largely mediated by SES, with potential unobserved confounding. While standard fairness metrics indicated modest bias in favor of non-Western patients, deeper feature-level analyses revealed subgroup-specific patterns of variable importance that suggest potentially less favorable underlying conditions. For instance, malignancy (ADG 32) had a stronger predictive impact among non-Western patients but contributed less to false negatives compared to Dutch patients, possibly reflecting structural disparities in cancer diagnosis and care. These findings highlight the need for contextual, multi-level evaluations of algorithmic bias. Fairness in healthcare algorithms must be approached as a socio-technical challenge, requiring multidisciplinary collaboration to uncover root causes in the processes that generate the data.
- Open Access
- 28 September 2026
- Published by JMIR Publications Inc. in JMIR mHealth and uHealth
- Vol. 14, e79512
A Cardiovascular Care Pathway Supported by Remote Patient Management in Dutch Primary Care: Cost-Effectiveness, Budget Impact, and Workload Analysis
Background: Remote patient management (RPM) that supports patient self-monitoring of vital parameters and lifestyle factors may improve cardiovascular risk management (CVRM) in primary care. However, large-scale implementation remains limited, partly due to insufficient evidence on long-term value for money, budget impact, and implications for health care professionals' workload. Objective: This study aimed to estimate the long-term cost-effectiveness, 5-year health care budget impact, and expected changes in general practitioner (GP) and practice nurse (PN) workload associated with the CVRM-Box intervention in the Netherlands. Methods: We conducted a model-based economic evaluation comparing CVRM-Box with care as usual in Dutch primary care integrated CVRM programs. CVRM-Box is a multicomponent RPM intervention comprising a digital blood pressure (BP) monitor, digital weight scale, step counter/activity tracker, and a mobile app, with measurements transferred to the GP practice for periodic review. A time-inhomogeneous cohort Markov model simulated lifetime transitions among health states, including at-risk, post–myocardial infarction (MI), poststroke, recurrent events, cardiovascular death, and noncardiovascular death. Cardiovascular risks were modeled using prediction equations (SCORE2, SCORE2-OP, and SMART2) populated with subgroup-specific risk factor profiles. Intervention effects were modeled as changes in systolic BP derived from a matched cohort study of CVRM-Box. Other key parameters (costs and utilities) were similarly obtained from the matched cohort study, or routine primary care data, and published sources. Outcomes included incremental cost-effectiveness ratios (ICERs, indicating cost per quality-adjusted life year [QALY] gained), cost-effectiveness probabilities, 5-year health care budget impact, and an exploratory workload analysis estimating annual changes in visit time and remote consultation frequency for GPs and PNs. Probabilistic and scenario uncertainty analyses were performed. Results: In the overall population, CVRM-Box increased costs and QALYs versus care as usual, yielding an ICER of €17,340/QALY gained (EUR €1=US $1.11 as of 29 December 2023) and a 60% probability of cost-effectiveness at a willingness-to-pay threshold of €20,000/QALY. Cost-effectiveness was more favorable in higher-risk subgroups (uncontrolled BP and/or prior MI/stroke), with ≥70% probability of cost-effectiveness at €20,000/QALY, whereas in the lower-risk subgroup with controlled BP and no prior MI/stroke, the intervention is unlikely to be cost-effective (ICER: €42,384/QALY). The 5-year health care budget impact was €2.9 million and €662.2 million for regional and national rollout, respectively. In a typical Dutch primary care practice, the CVRM-Box reduced PN workload by 25.9 h and 44.7 remote consultations annually, while the change in GP workload was negligible (0.5 h; 3.3 remote consultations). Conclusions: The CVRM-Box multicomponent RPM intervention is likely to be cost-effective among high-risk subgroups but not in lower-risk groups. The intervention further reduces PN workload but not GP workload. Results primarily apply to the Dutch context and may not generalize to health care systems with different payment/incentive arrangements. Additionally, long-term outcomes were modeled using prediction models rather than observed cardiovascular events. Nonetheless, our findings provide an argument for implementation in primary care for higher-risk subgroups.
- 28 September 2026
- Published by JMIR Publications Inc. in JMIR mHealth and uHealth
- Vol. 14, e79512
Open AccessA Cardiovascular Care Pathway Supported by Remote Patient Management in Dutch Primary Care: Cost-Effectiveness, Budget Impact, and Workload Analysis
Background: Remote patient management (RPM) that supports patient self-monitoring of vital parameters and lifestyle factors may improve cardiovascular risk management (CVRM) in primary care. However, large-scale implementation remains limited, partly due to insufficient evidence on long-term value for money, budget impact, and implications for health care professionals' workload. Objective: This study aimed to estimate the long-term cost-effectiveness, 5-year health care budget impact, and expected changes in general practitioner (GP) and practice nurse (PN) workload associated with the CVRM-Box intervention in the Netherlands. Methods: We conducted a model-based economic evaluation comparing CVRM-Box with care as usual in Dutch primary care integrated CVRM programs. CVRM-Box is a multicomponent RPM intervention comprising a digital blood pressure (BP) monitor, digital weight scale, step counter/activity tracker, and a mobile app, with measurements transferred to the GP practice for periodic review. A time-inhomogeneous cohort Markov model simulated lifetime transitions among health states, including at-risk, post–myocardial infarction (MI), poststroke, recurrent events, cardiovascular death, and noncardiovascular death. Cardiovascular risks were modeled using prediction equations (SCORE2, SCORE2-OP, and SMART2) populated with subgroup-specific risk factor profiles. Intervention effects were modeled as changes in systolic BP derived from a matched cohort study of CVRM-Box. Other key parameters (costs and utilities) were similarly obtained from the matched cohort study, or routine primary care data, and published sources. Outcomes included incremental cost-effectiveness ratios (ICERs, indicating cost per quality-adjusted life year [QALY] gained), cost-effectiveness probabilities, 5-year health care budget impact, and an exploratory workload analysis estimating annual changes in visit time and remote consultation frequency for GPs and PNs. Probabilistic and scenario uncertainty analyses were performed. Results: In the overall population, CVRM-Box increased costs and QALYs versus care as usual, yielding an ICER of €17,340/QALY gained (EUR €1=US $1.11 as of 29 December 2023) and a 60% probability of cost-effectiveness at a willingness-to-pay threshold of €20,000/QALY. Cost-effectiveness was more favorable in higher-risk subgroups (uncontrolled BP and/or prior MI/stroke), with ≥70% probability of cost-effectiveness at €20,000/QALY, whereas in the lower-risk subgroup with controlled BP and no prior MI/stroke, the intervention is unlikely to be cost-effective (ICER: €42,384/QALY). The 5-year health care budget impact was €2.9 million and €662.2 million for regional and national rollout, respectively. In a typical Dutch primary care practice, the CVRM-Box reduced PN workload by 25.9 h and 44.7 remote consultations annually, while the change in GP workload was negligible (0.5 h; 3.3 remote consultations). Conclusions: The CVRM-Box multicomponent RPM intervention is likely to be cost-effective among high-risk subgroups but not in lower-risk groups. The intervention further reduces PN workload but not GP workload. Results primarily apply to the Dutch context and may not generalize to health care systems with different payment/incentive arrangements. Additionally, long-term outcomes were modeled using prediction models rather than observed cardiovascular events. Nonetheless, our findings provide an argument for implementation in primary care for higher-risk subgroups.
- Open Access
- 23 July 2026
- Published by SAGE Publications in European Journal of Criminology
Leveraging synthetic data to improve open science practices in criminology – releasing a synthetic homicide dataset and dashboard
In criminology, the call for transparency and openness in research practices is resounding louder than ever. FAIR ( f indable, a ccessible, i nteroperable and r eusable) and open data are recognized as a necessity to enable replication of criminological research, enhance future empirical studies, foster collaboration and inform evidence-based policymaking. Yet, criminologists are faced with an intricate web of privacy regulations governing the dissemination of sensitive data about crime victims, perpetrators, witnesses or other stakeholders. In this article, we examine criminology-specific challenges posed by regulations like General Data Protection Regulation, which mandate safeguarding of personally identifiable information, often rendering original data unsuitable for open sharing. To reconcile this dilemma, we discuss the use of synthetic data generation as a tool to overcome (parts of) this challenge. Synthetic data entails creating artificial datasets that mimic the statistical characteristics of real data while obfuscating identifying information. By utilizing synthetic data, criminologists can both adhere to stringent regulations while disseminating data and facilitating open science practices. This article discusses benefits, as well as ethical considerations and limitations of synthetic data. As a proof-of-concept, we showcase the synthesis and publicly release a synthetic version of the Dutch Homicide Monitor, using an existing open-source tool for data synthetization.
- 23 July 2026
- Published by SAGE Publications in European Journal of Criminology
Open AccessLeveraging synthetic data to improve open science practices in criminology – releasing a synthetic homicide dataset and dashboard
In criminology, the call for transparency and openness in research practices is resounding louder than ever. FAIR ( f indable, a ccessible, i nteroperable and r eusable) and open data are recognized as a necessity to enable replication of criminological research, enhance future empirical studies, foster collaboration and inform evidence-based policymaking. Yet, criminologists are faced with an intricate web of privacy regulations governing the dissemination of sensitive data about crime victims, perpetrators, witnesses or other stakeholders. In this article, we examine criminology-specific challenges posed by regulations like General Data Protection Regulation, which mandate safeguarding of personally identifiable information, often rendering original data unsuitable for open sharing. To reconcile this dilemma, we discuss the use of synthetic data generation as a tool to overcome (parts of) this challenge. Synthetic data entails creating artificial datasets that mimic the statistical characteristics of real data while obfuscating identifying information. By utilizing synthetic data, criminologists can both adhere to stringent regulations while disseminating data and facilitating open science practices. This article discusses benefits, as well as ethical considerations and limitations of synthetic data. As a proof-of-concept, we showcase the synthesis and publicly release a synthetic version of the Dutch Homicide Monitor, using an existing open-source tool for data synthetization.
- Open Access
- 15 July 2026
- Published by Elsevier in SSRN Electronic Journal
Developing the Neighborhood-level Dutch Environmental Stress Index (D-ESI) Integrating Multiple Environmental Domains
Background: Multiple environmental stressors influence population health, and a composite measure can help identify environmentally disadvantaged neighborhoods. Environmental conditions often differ between urban and rural areas. In this study, we developed a Dutch Environmental Stress Index (D-ESI) and characterized its spatial distribution across urbanicity strata. Methods: We identified neighborhood-level variables across three environmental domains: physico-chemical, built environment, and sociodemographic. We obtained data from Statistics Netherlands (CBS), Climate Impact Atlas, and National Georegister (NGR) for the year 2023. Principal component analysis (PCA) was applied within each domain to construct domain-specific ESIs. A composite ESI was calculated as the mean of three domain-specific indices. In addition, urbanicity-stratified indices were created. Construct and discriminant validity were assessed. Findings: The composite ESI had an overall median value of -0·04 (range: -3·73–12·29) and was higher in urban areas. Spatial patterns showed elevated composite ESI values in the urban outskirts of major Dutch metropolises and across the central and southern regions of the country. Physico-chemical and sociodemographic stress were concentrated in urban areas, whereas built environment stress showed opposing patterns. Urbanicity-stratified analyses demonstrated moderate within-stratum variability. Interpretation: Urban areas tended to have better service availability and infrastructure but poorer air quality, higher noise pollution, greater population pressure, and less favorable socioeconomic position (SEP). However, neighborhoods with similar levels of urbanization can experience different combinations of physico-chemical, built environment, and sociodemographic stress. The ESIs may support evidence-based policymaking for environmental and public health interventions and serve as exposure proxies for epidemiological studies on disease risk.
- 15 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDeveloping the Neighborhood-level Dutch Environmental Stress Index (D-ESI) Integrating Multiple Environmental Domains
Background: Multiple environmental stressors influence population health, and a composite measure can help identify environmentally disadvantaged neighborhoods. Environmental conditions often differ between urban and rural areas. In this study, we developed a Dutch Environmental Stress Index (D-ESI) and characterized its spatial distribution across urbanicity strata. Methods: We identified neighborhood-level variables across three environmental domains: physico-chemical, built environment, and sociodemographic. We obtained data from Statistics Netherlands (CBS), Climate Impact Atlas, and National Georegister (NGR) for the year 2023. Principal component analysis (PCA) was applied within each domain to construct domain-specific ESIs. A composite ESI was calculated as the mean of three domain-specific indices. In addition, urbanicity-stratified indices were created. Construct and discriminant validity were assessed. Findings: The composite ESI had an overall median value of -0·04 (range: -3·73–12·29) and was higher in urban areas. Spatial patterns showed elevated composite ESI values in the urban outskirts of major Dutch metropolises and across the central and southern regions of the country. Physico-chemical and sociodemographic stress were concentrated in urban areas, whereas built environment stress showed opposing patterns. Urbanicity-stratified analyses demonstrated moderate within-stratum variability. Interpretation: Urban areas tended to have better service availability and infrastructure but poorer air quality, higher noise pollution, greater population pressure, and less favorable socioeconomic position (SEP). However, neighborhoods with similar levels of urbanization can experience different combinations of physico-chemical, built environment, and sociodemographic stress. The ESIs may support evidence-based policymaking for environmental and public health interventions and serve as exposure proxies for epidemiological studies on disease risk.
- Open Access
- 9 July 2026
- Published by JMIR Publications Inc. in JMIR Medical Informatics
- Vol. 14, e80377
Quantifying the Predictive Power of Social Determinants of Health in Cardiovascular Disease and Type 2 Diabetes Progression Using XGBoost: Retrospective Cohort Study
Background: Cardiovascular diseases (CVDs) and type 2 diabetes (DM2) are influenced not only by biomedical risk factors but also by social determinants of health (SDOH). While the inclusion of SDOH in predictive models is increasingly advocated, few studies have quantified their specific contribution in a high-risk clinical cohort using robust statistical and machine-learning approaches. Objective: This study aims to quantify the added predictive value of SDOH in predicting CVD or DM2 disease onset within 5 years, within 10 years, and at any time during follow-up among individuals already at elevated risk and to compare this added value across multiple modeling setups and frameworks. Methods: We used a large, linked dataset of over 58,000 inclusion events from the Extramural Leiden University Medical Center Academic Network data warehouse in the Netherlands, combining structured coded diagnosis and medication records from general practitioners with individual-level socioeconomic data from Statistics Netherlands. Individuals aged 30 years and older without prior DM2 or CVD were followed to assess disease progression. We trained Cox proportional hazards (CPH) and Extreme Gradient Boosting (XGBoost) models to predict progression to DM2 or CVD within 5 and 10 years and overall. All analyses were performed using the R programming language. Experiments included comparisons of Systematic Coronary Risk Evaluation 2, CPH, and XGBoost models; evaluation of time-bound and survival-based formulations; and quantification of SDOH impact using feature subset XGBoost models and Shapley additive explanations (SHAP)–based importance. Results: For the 5-year prediction of CVD or DM2, the combined XGBoost model using biomedical and SDOH predictors achieved an area under the receiver operating characteristic curve (AUC) of 0.738, significantly outperforming the biomedical-only model (AUC=0.728; P =.01) and the SDOH-only model (AUC=0.691; P <.001). For 10-year CVD prediction, XGBoost achieved an AUC of 0.729, outperforming CPH (AUC=0.718; P =.02) and Systematic Coronary Risk Evaluation 2 (AUC=0.697; P <.001). For overall event prediction, XGBoost again performed best (AUC=0.719), significantly higher than CPH (AUC=0.704; P <.001). SHAP analyses showed that biomedical predictors contributed most strongly on a per-feature basis, while a subset of SDOH variables, particularly income- and benefit-related indicators, provided complementary predictive signal and ranked among the most influential predictors. Conclusions: Incorporating SDOH improved the prediction of CVD and DM2 onset in a clinically defined high-risk cohort. Across hundreds of linked predictors, SDOH provided measurable incremental discrimination beyond biomedical risk factors, and income- and benefit-related variables ranked among the most influential features. SHAP analyses indicated that this added value was largely driven by a limited subset of highly informative social predictors. These findings support integrating structured SDOH into clinically actionable risk stratification models.
- 9 July 2026
- Published by JMIR Publications Inc. in JMIR Medical Informatics
- Vol. 14, e80377
Open AccessQuantifying the Predictive Power of Social Determinants of Health in Cardiovascular Disease and Type 2 Diabetes Progression Using XGBoost: Retrospective Cohort Study
Background: Cardiovascular diseases (CVDs) and type 2 diabetes (DM2) are influenced not only by biomedical risk factors but also by social determinants of health (SDOH). While the inclusion of SDOH in predictive models is increasingly advocated, few studies have quantified their specific contribution in a high-risk clinical cohort using robust statistical and machine-learning approaches. Objective: This study aims to quantify the added predictive value of SDOH in predicting CVD or DM2 disease onset within 5 years, within 10 years, and at any time during follow-up among individuals already at elevated risk and to compare this added value across multiple modeling setups and frameworks. Methods: We used a large, linked dataset of over 58,000 inclusion events from the Extramural Leiden University Medical Center Academic Network data warehouse in the Netherlands, combining structured coded diagnosis and medication records from general practitioners with individual-level socioeconomic data from Statistics Netherlands. Individuals aged 30 years and older without prior DM2 or CVD were followed to assess disease progression. We trained Cox proportional hazards (CPH) and Extreme Gradient Boosting (XGBoost) models to predict progression to DM2 or CVD within 5 and 10 years and overall. All analyses were performed using the R programming language. Experiments included comparisons of Systematic Coronary Risk Evaluation 2, CPH, and XGBoost models; evaluation of time-bound and survival-based formulations; and quantification of SDOH impact using feature subset XGBoost models and Shapley additive explanations (SHAP)–based importance. Results: For the 5-year prediction of CVD or DM2, the combined XGBoost model using biomedical and SDOH predictors achieved an area under the receiver operating characteristic curve (AUC) of 0.738, significantly outperforming the biomedical-only model (AUC=0.728; P =.01) and the SDOH-only model (AUC=0.691; P <.001). For 10-year CVD prediction, XGBoost achieved an AUC of 0.729, outperforming CPH (AUC=0.718; P =.02) and Systematic Coronary Risk Evaluation 2 (AUC=0.697; P <.001). For overall event prediction, XGBoost again performed best (AUC=0.719), significantly higher than CPH (AUC=0.704; P <.001). SHAP analyses showed that biomedical predictors contributed most strongly on a per-feature basis, while a subset of SDOH variables, particularly income- and benefit-related indicators, provided complementary predictive signal and ranked among the most influential predictors. Conclusions: Incorporating SDOH improved the prediction of CVD and DM2 onset in a clinically defined high-risk cohort. Across hundreds of linked predictors, SDOH provided measurable incremental discrimination beyond biomedical risk factors, and income- and benefit-related variables ranked among the most influential features. SHAP analyses indicated that this added value was largely driven by a limited subset of highly informative social predictors. These findings support integrating structured SDOH into clinically actionable risk stratification models.
- Open Access
- 7 July 2026
- Published by Springer Nature in BMC Public Health
- Vol. 26 (1)
Identifying co-occurring neighbourhood environmental patterns and their association with health behaviours in a Dutch urban population at high cardiometabolic risk
Background: Pattern-based approaches to characterising neighbourhood environments have been applied across cities and regions, but it is less clear whether they can identify independent environmental effects within a single dense urban area, the scale at which most local public-health interventions are designed. Furthermore, research has primarily focused on general populations, leaving a gap in understanding how environmental factors relate to health behaviours in high-risk groups. This study aimed to (1) describe how a broad set of food-environment, green-space and walkability indicators co-occur in a highly urbanised Dutch city and how the resulting patterns relate to population density and neighbourhood socioeconomic position; and (2) explore whether these patterns are associated with weight status, diet quality and physical activity in primary-care patients at high cardiovascular risk. Methods: This cross-sectional study used baseline data from the Healthy Heart study in The Hague, Netherlands ( N = 475 participants from 73 postal codes). Twenty-three indicators across food environment, green space and walkability domains were assessed at two spatial scales. Correlation analyses examined intercorrelations among indicators and contextual characteristics. Neighbourhood patterns were identified using principal component analysis (PCA). Associations of component scores with weight status, diet quality and physical activity were examined using (multinomial) logistic and linear regression, adjusted for individual and neighbourhood-level covariates. Results: Four components explained 87% of the variance in the environmental indicators. The dominant neighbourhood pattern, characterised by high food-retail density, high walkability and low vegetation cover, correlated very strongly with population density and this coupling persisted within the most highly urbanised stratum. The “Relative food environment advantage” pattern showed the most consistent association across models with lower odds of being overweight, but also counterintuitively with lower fruit and vegetable diet quality. There was some indication of effect modification by neighbourhood socioeconomic position for this pattern. Conclusions: Within a single dense urban area, neighbourhood characteristics cluster into structured patterns in urban settings. Food environment measures remained coupled to population density even within high-density areas, fundamentally challenging independent effect estimation. Modest, paradoxical associations suggest limited environmental influence on lifestyle behaviours and weight in medically supervised high-risk populations.
- 7 July 2026
- Published by Springer Nature in BMC Public Health
- Vol. 26 (1)
Open AccessIdentifying co-occurring neighbourhood environmental patterns and their association with health behaviours in a Dutch urban population at high cardiometabolic risk
Background: Pattern-based approaches to characterising neighbourhood environments have been applied across cities and regions, but it is less clear whether they can identify independent environmental effects within a single dense urban area, the scale at which most local public-health interventions are designed. Furthermore, research has primarily focused on general populations, leaving a gap in understanding how environmental factors relate to health behaviours in high-risk groups. This study aimed to (1) describe how a broad set of food-environment, green-space and walkability indicators co-occur in a highly urbanised Dutch city and how the resulting patterns relate to population density and neighbourhood socioeconomic position; and (2) explore whether these patterns are associated with weight status, diet quality and physical activity in primary-care patients at high cardiovascular risk. Methods: This cross-sectional study used baseline data from the Healthy Heart study in The Hague, Netherlands ( N = 475 participants from 73 postal codes). Twenty-three indicators across food environment, green space and walkability domains were assessed at two spatial scales. Correlation analyses examined intercorrelations among indicators and contextual characteristics. Neighbourhood patterns were identified using principal component analysis (PCA). Associations of component scores with weight status, diet quality and physical activity were examined using (multinomial) logistic and linear regression, adjusted for individual and neighbourhood-level covariates. Results: Four components explained 87% of the variance in the environmental indicators. The dominant neighbourhood pattern, characterised by high food-retail density, high walkability and low vegetation cover, correlated very strongly with population density and this coupling persisted within the most highly urbanised stratum. The “Relative food environment advantage” pattern showed the most consistent association across models with lower odds of being overweight, but also counterintuitively with lower fruit and vegetable diet quality. There was some indication of effect modification by neighbourhood socioeconomic position for this pattern. Conclusions: Within a single dense urban area, neighbourhood characteristics cluster into structured patterns in urban settings. Food environment measures remained coupled to population density even within high-density areas, fundamentally challenging independent effect estimation. Modest, paradoxical associations suggest limited environmental influence on lifestyle behaviours and weight in medically supervised high-risk populations.
- Open Access
- 7 July 2026
- Published by Open Engineering Inc in engrXiv
Metrics That Matter: A Practical Survey on Tabular Synthetic Data Evaluation
Assessing the quality of synthetic data (SD) is vital to determine whether it can provide a viable alternative to real data. A wide variety of metrics exist to examine the three archetypal dimensions of SD evaluation: realism (fidelity), task-specific usefulness (utility), and remaining disclosure risk (privacy). Current work in SD generation often relies on the ad-hoc selection of evaluation metrics without a clear justification, while the suitability of metrics strongly depend on the dataset and other contextual factors. This paper surveys the field of SD evaluation, provides guidance regarding metric selection based on four key questions pertaining to the task, goal, data type, and domain of SD, and provides general practical recommendations on SD evaluation. Finally, experiments on an illustrative dataset of electronic health records show how researchers can bring our insights and recommendations for SD evaluation into practice. By doing so, we aim to support researchers and practitioners seeking to generate and evaluate SD.
- 7 July 2026
- Published by Open Engineering Inc in engrXiv
Open AccessMetrics That Matter: A Practical Survey on Tabular Synthetic Data Evaluation
Assessing the quality of synthetic data (SD) is vital to determine whether it can provide a viable alternative to real data. A wide variety of metrics exist to examine the three archetypal dimensions of SD evaluation: realism (fidelity), task-specific usefulness (utility), and remaining disclosure risk (privacy). Current work in SD generation often relies on the ad-hoc selection of evaluation metrics without a clear justification, while the suitability of metrics strongly depend on the dataset and other contextual factors. This paper surveys the field of SD evaluation, provides guidance regarding metric selection based on four key questions pertaining to the task, goal, data type, and domain of SD, and provides general practical recommendations on SD evaluation. Finally, experiments on an illustrative dataset of electronic health records show how researchers can bring our insights and recommendations for SD evaluation into practice. By doing so, we aim to support researchers and practitioners seeking to generate and evaluate SD.
- Open Access
The Relationship Between Mobility and Population Health in Urban Areas
Urban mobility may provide insight into population health by capturing how residents connect to services, resources, and urban systems. This is relevant for communities facing higher disease burden and limited resources, where reduced connectivity may signal barriers to care, healthy environments, and participation. Mobility patterns are furthermore shaped by socioeconomic position, housing, environmental quality, facilities access, lifestyle patterns, and population composition. Mobility-health associations may therefore reflect underlying social and environmental disadvantage rather than mobility itself, risking misdirected public-health policy responses. This ecological cross-sectional study examined associations between aggregated mobile phone–based mobility and health outcomes in The Hague, Netherlands, from January-July 2019. Mobile phone mobility was measured as mean outgoing mobility distance across eight regions. Contextual and health indicators were available at neighbourhood-level and were aggregated or linked to regions where required. Health outcomes were operationalised as indicators of disease burden, including cardiometabolic medication prescriptions, polypharmacy, and a syndemic-based measure of interacting health conditions. Contextual domains were selected using spatial clustering and ordinary least squares models, after which residual mobility–health associations were assessed. Outgoing mobility varied across regions and was strongly patterned by contextual factors. Lifestyle, housing, physical environment, and income accounted for 73.2% of variance in outgoing mobility. After adjustment, residual mobility showed weak, non-significant associations with cardiometabolic medication prescription, polypharmacy, and the syndemic-based measure. Sensitivity analyses supported these findings. Aggregated mobility should not be interpreted as a straightforward independent determinant of health. Instead, it appears to function as an integrative marker of urban context, spatial structure, and population composition.
Open AccessThe Relationship Between Mobility and Population Health in Urban Areas
Urban mobility may provide insight into population health by capturing how residents connect to services, resources, and urban systems. This is relevant for communities facing higher disease burden and limited resources, where reduced connectivity may signal barriers to care, healthy environments, and participation. Mobility patterns are furthermore shaped by socioeconomic position, housing, environmental quality, facilities access, lifestyle patterns, and population composition. Mobility-health associations may therefore reflect underlying social and environmental disadvantage rather than mobility itself, risking misdirected public-health policy responses. This ecological cross-sectional study examined associations between aggregated mobile phone–based mobility and health outcomes in The Hague, Netherlands, from January-July 2019. Mobile phone mobility was measured as mean outgoing mobility distance across eight regions. Contextual and health indicators were available at neighbourhood-level and were aggregated or linked to regions where required. Health outcomes were operationalised as indicators of disease burden, including cardiometabolic medication prescriptions, polypharmacy, and a syndemic-based measure of interacting health conditions. Contextual domains were selected using spatial clustering and ordinary least squares models, after which residual mobility–health associations were assessed. Outgoing mobility varied across regions and was strongly patterned by contextual factors. Lifestyle, housing, physical environment, and income accounted for 73.2% of variance in outgoing mobility. After adjustment, residual mobility showed weak, non-significant associations with cardiometabolic medication prescription, polypharmacy, and the syndemic-based measure. Sensitivity analyses supported these findings. Aggregated mobility should not be interpreted as a straightforward independent determinant of health. Instead, it appears to function as an integrative marker of urban context, spatial structure, and population composition.
- Open Access
- 21 May 2026
- Published by IOS Press
- Vol. 336, 934-938
OpenExtract: Automated Data Extraction for Systematic Reviews in Health
This study presents OpenExtract, an open-source pipeline for automated data extraction in large-scale systematic literature reviews. The pipeline queries large language models (LLMs) to predict data entries based on relevant sections of scientific articles. To test the efficacy of OpenExtract, we apply it to a systematic literature review in digital health and compare its outputs with those of human researchers. OpenExtract achieves precision and recall scores of > 0.8 in this task, indicating that it can be effective at extracting data automatically and efficiently. OpenExtract: https://github.com/JimAchterbergLUMC/OpenExtract.
- 21 May 2026
- Published by IOS Press
- Vol. 336, 934-938
Open AccessOpenExtract: Automated Data Extraction for Systematic Reviews in Health
This study presents OpenExtract, an open-source pipeline for automated data extraction in large-scale systematic literature reviews. The pipeline queries large language models (LLMs) to predict data entries based on relevant sections of scientific articles. To test the efficacy of OpenExtract, we apply it to a systematic literature review in digital health and compare its outputs with those of human researchers. OpenExtract achieves precision and recall scores of > 0.8 in this task, indicating that it can be effective at extracting data automatically and efficiently. OpenExtract: https://github.com/JimAchterbergLUMC/OpenExtract.
- Open Access
Refocusing Algorithmic Fairness on Feature-Level Bias: A Diagnostic Approach Using Dutch EHR Data
While algorithmic fairness research in healthcare has predominantly focused on disparities in model performance, less attention has been given to the underlying data structures that may drive such disparities. High-level fairness metrics often obscure the deeper feature-level dynamics necessary for a critical and context-aware assessment of fairness. To address this gap, we propose and apply a diagnostic framework termed Feature-Level Bias Identification And Sensemaking (FL-BIAS). As a case study, we conducted a secondary analysis of a retrospective cross-sectional cohort study using electronic health record (EHR) data from Dutch general practitioners, linked with sociodemographic data from Statistics Netherlands. The dataset included 112,872 patients, of whom 16.2% had a non-Western migration background. Hospitalization in the following year was modeled using Johns Hopkins Aggregated Diagnosis Groups (ADGs). We trained logistic regression and XGBoost models on different subgroup datasets to evaluate performance disparities using fairness metrics. To analyze feature-level contributions to predictions and errors, we applied Shapley Value methods, including Kernel SHAP and Cohort Shapley. Exploratory analysis revealed significant differences in SES and ADG distributions between Dutch and non-Western groups, though Multiple Correspondence Analysis showed minimal structural variation. Mediation analysis indicated that the effect of migration background on hospitalization was largely mediated by SES, with potential unobserved confounding. While standard fairness metrics indicated modest bias in favor of non-Western patients, deeper feature-level analyses revealed subgroup-specific patterns of variable importance that suggest potentially less favorable underlying conditions. For instance, malignancy (ADG 32) had a stronger predictive impact among non-Western patients but contributed less to false negatives compared to Dutch patients, possibly reflecting structural disparities in cancer diagnosis and care. These findings highlight the need for contextual, multi-level evaluations of algorithmic bias. Fairness in healthcare AI must be approached as a socio-technical challenge, requiring multidisciplinary collaboration to uncover root causes and guide effective mitigation strategies. Author Summary: Unfair algorithms often originate in the data itself. To help detect such hidden biases, we combined known data science methods into a diagnostic approach named Feature-Level Bias Identification And Sensemaking (FL-BIAS). Using Dutch general practitioner records linked with national demographic information, we explored how patient migration background corrected for socioeconomic status affect predictions of hospital admissions. We discovered that certain medical conditions, such as cancer and chronic illness, contributed differently to predictions for Dutch compared to non-Western patients, revealing subtle but important patterns in how data reflects social inequalities. Our goal was to move beyond simple fairness scores and better understand how inequalities can be hidden in health data, supporting the design of fairer and more transparent healthcare AI tools that truly serve all patients.
Open AccessRefocusing Algorithmic Fairness on Feature-Level Bias: A Diagnostic Approach Using Dutch EHR Data
While algorithmic fairness research in healthcare has predominantly focused on disparities in model performance, less attention has been given to the underlying data structures that may drive such disparities. High-level fairness metrics often obscure the deeper feature-level dynamics necessary for a critical and context-aware assessment of fairness. To address this gap, we propose and apply a diagnostic framework termed Feature-Level Bias Identification And Sensemaking (FL-BIAS). As a case study, we conducted a secondary analysis of a retrospective cross-sectional cohort study using electronic health record (EHR) data from Dutch general practitioners, linked with sociodemographic data from Statistics Netherlands. The dataset included 112,872 patients, of whom 16.2% had a non-Western migration background. Hospitalization in the following year was modeled using Johns Hopkins Aggregated Diagnosis Groups (ADGs). We trained logistic regression and XGBoost models on different subgroup datasets to evaluate performance disparities using fairness metrics. To analyze feature-level contributions to predictions and errors, we applied Shapley Value methods, including Kernel SHAP and Cohort Shapley. Exploratory analysis revealed significant differences in SES and ADG distributions between Dutch and non-Western groups, though Multiple Correspondence Analysis showed minimal structural variation. Mediation analysis indicated that the effect of migration background on hospitalization was largely mediated by SES, with potential unobserved confounding. While standard fairness metrics indicated modest bias in favor of non-Western patients, deeper feature-level analyses revealed subgroup-specific patterns of variable importance that suggest potentially less favorable underlying conditions. For instance, malignancy (ADG 32) had a stronger predictive impact among non-Western patients but contributed less to false negatives compared to Dutch patients, possibly reflecting structural disparities in cancer diagnosis and care. These findings highlight the need for contextual, multi-level evaluations of algorithmic bias. Fairness in healthcare AI must be approached as a socio-technical challenge, requiring multidisciplinary collaboration to uncover root causes and guide effective mitigation strategies. Author Summary: Unfair algorithms often originate in the data itself. To help detect such hidden biases, we combined known data science methods into a diagnostic approach named Feature-Level Bias Identification And Sensemaking (FL-BIAS). Using Dutch general practitioner records linked with national demographic information, we explored how patient migration background corrected for socioeconomic status affect predictions of hospital admissions. We discovered that certain medical conditions, such as cancer and chronic illness, contributed differently to predictions for Dutch compared to non-Western patients, revealing subtle but important patterns in how data reflects social inequalities. Our goal was to move beyond simple fairness scores and better understand how inequalities can be hidden in health data, supporting the design of fairer and more transparent healthcare AI tools that truly serve all patients.