Maarten van Smeden
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Maarten van Smeden has a total of 547 co-authors. Here are the top co-authors by number of shared publications:
- Ben Van Calster
Department of Development and Regeneration, KU Leuven, Leuven, Belgium
20 shared publications - Gary Stephen Collins
Centre for Statistics in Medicine, Nuffield Department of Orthopaedics, Rheumatology and Musculoskeletal Sciences, University of Oxford, Oxford, UK
14 shared publications - Laure Wynants
Department of Development and Regeneration, KU Leuven, Leuven, Belgium
14 shared publications - Rolf Hendrik Herman Groenwold
Leiden University Medical Center
13 shared publications Karel G.M. Moons
Julius Center for Health Sciences and Primary Care, UMC Utrecht, Utrecht, The Netherlands
10 shared publicationsJohannes B. Reitsma
Julius Center for Health Sciences and Primary Care, University Medical Centre Utrecht, Utrecht University, Utrecht, The Netherlands
8 shared publications
Latest Publications
- Open Access
- 3 October 2026
- Published by Springer Nature in npj Digital Medicine
Shapley value explanations for clinical prediction models: a scoping review and guide
Shapley value explanations are increasingly used to provide added transparency for “black-box” prediction models in healthcare. In this article, we provide an introductory overview of Shapley value explanations and describe their use through a detailed scoping review of 100 randomly selected, peer-reviewed publications that used Shapley explanations as part of their clinical research. We focus on the use of Shapley value explanations to explain the predictions of clinical prediction models that use tabular input features (e.g., predictors that measure patient characteristics like blood pressure and age) to predict health outcomes in individuals. In our review of the literature, we found that the methods used to compute Shapley value explanations were often underreported; e.g., 91% of publications did not disclose the algorithm that was used to compute Shapley value explanations, and 97% did not disclose the source or size of the reference (baseline) population used in the calculations. We identify four dominant motivations for using Shapley explanations in the literature (identification of key features, clinical decision support, trustworthiness, and exploratory analyses), elaborate on commonly found (mis) interpretations, and discuss challenges associated with Shapley value explanations. Finally, we provide practical recommendations for use of Shapley values in the context of clinical prediction models.
- 3 October 2026
- Published by Springer Nature in npj Digital Medicine
Open AccessShapley value explanations for clinical prediction models: a scoping review and guide
Shapley value explanations are increasingly used to provide added transparency for “black-box” prediction models in healthcare. In this article, we provide an introductory overview of Shapley value explanations and describe their use through a detailed scoping review of 100 randomly selected, peer-reviewed publications that used Shapley explanations as part of their clinical research. We focus on the use of Shapley value explanations to explain the predictions of clinical prediction models that use tabular input features (e.g., predictors that measure patient characteristics like blood pressure and age) to predict health outcomes in individuals. In our review of the literature, we found that the methods used to compute Shapley value explanations were often underreported; e.g., 91% of publications did not disclose the algorithm that was used to compute Shapley value explanations, and 97% did not disclose the source or size of the reference (baseline) population used in the calculations. We identify four dominant motivations for using Shapley explanations in the literature (identification of key features, clinical decision support, trustworthiness, and exploratory analyses), elaborate on commonly found (mis) interpretations, and discuss challenges associated with Shapley value explanations. Finally, we provide practical recommendations for use of Shapley values in the context of clinical prediction models.
- Open Access
- 25 March 2026
- Published by Elsevier in Journal of Clinical Epidemiology
- Vol. 194, 112256
Preliminary appraisal of machine learning based prediction models
- 25 March 2026
- Published by Elsevier in Journal of Clinical Epidemiology
- Vol. 194, 112256
Open AccessPreliminary appraisal of machine learning based prediction models
- Top Cited PapersOpen Access
- 4 March 2026
- Published by Oxford University Press (OUP) in European Heart Journal - Digital Health
- Vol. 7 (3) , ztag038
Signal or noise? Evaluating commonly used attribution methods for explaining deep neural networks in electrocardiogram classification
Aims: Attribution-based explainability methods are widely used in electrocardiogram (ECG) analysis to interpret predictions from ‘black-box’ deep neural networks (DNNs). To be useful in clinical applications, attribution methods must produce explanations that are both clear and reflective of the model’s inner workings. This study evaluates 12 attribution methods in DNN-based ECG classification. Methods and results: We analysed 12 attribution methods using a dataset of 873 710 median beat ECGs spanning nine diagnostic classes. Methods were applied to convolutional neural network-based models trained for ECG classification. Performance was evaluated across four experiments: inter-method similarity, self-consistency, dependence on model weights, and ability to identify features important for model inference. All task models achieved an area under the receiver operating curve above 0.95. Attribution methods demonstrated low correlation and high variability across inter-method comparisons. Self-consistency across random model initializations was moderate for most methods (mean correlation 0.41–0.65). Randomizing model weights led to rapid loss of correlation, although some methods did not converge to zero. Perturbation of input data revealed differences in how well attribution methods identified features relevant to model performance. Conclusion: Attribution methods demonstrated limited reliability, instability across model variants and incomplete dependence on learned parameters, constraining their utility in high-stakes settings such as healthcare. These findings suggest that attribution techniques should be used cautiously and supported by task-specific sanity checks. Approaches grounded in rigorous validation, inherently interpretable modelling or counterfactual explanations may better support clinically meaningful insight.
- 4 March 2026
- Published by Oxford University Press (OUP) in European Heart Journal - Digital Health
- Vol. 7 (3) , ztag038
Top Cited PapersOpen AccessSignal or noise? Evaluating commonly used attribution methods for explaining deep neural networks in electrocardiogram classification
Aims: Attribution-based explainability methods are widely used in electrocardiogram (ECG) analysis to interpret predictions from ‘black-box’ deep neural networks (DNNs). To be useful in clinical applications, attribution methods must produce explanations that are both clear and reflective of the model’s inner workings. This study evaluates 12 attribution methods in DNN-based ECG classification. Methods and results: We analysed 12 attribution methods using a dataset of 873 710 median beat ECGs spanning nine diagnostic classes. Methods were applied to convolutional neural network-based models trained for ECG classification. Performance was evaluated across four experiments: inter-method similarity, self-consistency, dependence on model weights, and ability to identify features important for model inference. All task models achieved an area under the receiver operating curve above 0.95. Attribution methods demonstrated low correlation and high variability across inter-method comparisons. Self-consistency across random model initializations was moderate for most methods (mean correlation 0.41–0.65). Randomizing model weights led to rapid loss of correlation, although some methods did not converge to zero. Perturbation of input data revealed differences in how well attribution methods identified features relevant to model performance. Conclusion: Attribution methods demonstrated limited reliability, instability across model variants and incomplete dependence on learned parameters, constraining their utility in high-stakes settings such as healthcare. These findings suggest that attribution techniques should be used cautiously and supported by task-specific sanity checks. Approaches grounded in rigorous validation, inherently interpretable modelling or counterfactual explanations may better support clinically meaningful insight.
- Open Access
- 27 August 2025
- Published by Springer Nature in Spine Deformity
- Vol. 14 (1) , 311
Correction: What can we learn from scoliosis in children with the 22q11.2 deletion syndrome? Prognostic factors at pre-adolescent age for fast progressive, mild and self-resolving forms during adolescence
- 27 August 2025
- Published by Springer Nature in Spine Deformity
- Vol. 14 (1) , 311
Open AccessCorrection: What can we learn from scoliosis in children with the 22q11.2 deletion syndrome? Prognostic factors at pre-adolescent age for fast progressive, mild and self-resolving forms during adolescence
- Top Cited PapersOpen Access
- 31 July 2025
- Published by The Ewha Medical Journal, Ewha Womans University College of Medicine in Ewha Medical Journal
- Vol. 48 (3) , e48
TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods: a Korean translation
TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods: a Korean translation
- 31 July 2025
- Published by The Ewha Medical Journal, Ewha Womans University College of Medicine in Ewha Medical Journal
- Vol. 48 (3) , e48
Top Cited PapersOpen AccessTRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods: a Korean translation
TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods: a Korean translation
- Top Cited PapersOpen Access
PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods
The Prediction model Risk Of Bias ASsessment Tool (PROBAST) is used to assess the quality, risk of bias, and applicability of prediction models or algorithms and of prediction model/algorithm studies. Since PROBAST’s introduction in 2019, much progress has been made in the methodology for prediction modelling and in the use of artificial intelligence, including machine learning, techniques. An update to PROBAST-2019 is thus needed. This article describes the development of PROBAST+AI. PROBAST+AI consists of two distinctive parts: model development and model evaluation. For model development, PROBAST+AI users assess quality and applicability using 16 targeted signalling questions. For model evaluation, PROBAST+AI users assess the risk of bias and applicability using 18 targeted signalling questions. Both parts contain four domains: participants and data sources, predictors, outcome, and analysis. Applicability of the prediction model is rated for the participants and data sources, predictors, and outcome domains. PROBAST+AI may replace the original PROBAST tool and allows all key stakeholders (eg, model developers, AI companies, researchers, editors, reviewers, healthcare professionals, guideline developers, and policy organisations) to examine the quality, risk of bias, and applicability of any type of prediction model in the healthcare sector, irrespective of whether regression modelling or AI techniques are used.
Top Cited PapersOpen AccessPROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods
The Prediction model Risk Of Bias ASsessment Tool (PROBAST) is used to assess the quality, risk of bias, and applicability of prediction models or algorithms and of prediction model/algorithm studies. Since PROBAST’s introduction in 2019, much progress has been made in the methodology for prediction modelling and in the use of artificial intelligence, including machine learning, techniques. An update to PROBAST-2019 is thus needed. This article describes the development of PROBAST+AI. PROBAST+AI consists of two distinctive parts: model development and model evaluation. For model development, PROBAST+AI users assess quality and applicability using 16 targeted signalling questions. For model evaluation, PROBAST+AI users assess the risk of bias and applicability using 18 targeted signalling questions. Both parts contain four domains: participants and data sources, predictors, outcome, and analysis. Applicability of the prediction model is rated for the participants and data sources, predictors, and outcome domains. PROBAST+AI may replace the original PROBAST tool and allows all key stakeholders (eg, model developers, AI companies, researchers, editors, reviewers, healthcare professionals, guideline developers, and policy organisations) to examine the quality, risk of bias, and applicability of any type of prediction model in the healthcare sector, irrespective of whether regression modelling or AI techniques are used.
- Open Access
- 20 March 2025
- Published by Springer Nature in Spine Deformity
- Vol. 13 (4) , 1179-1187
What can we learn from scoliosis in children with the 22q11.2 deletion syndrome? Prognostic factors at pre-adolescent age for fast progressive, mild and self-resolving forms during adolescence
Introduction: Longitudinal data starting before adolescence and before curve onset, may elucidate prognostic factors for later scoliotic curve development. The aim is to predict the maximum curve acceleration (MCA; °/month) and the final curve progression in a cohort of 22q11.2DS subjects screened for scoliosis. Methods: Scoliosis screening starts immediately after 22q11.2DS diagnosis. A minimum of 2 years follow-up, two assessments, Risser 0, open triradiate cartilage at start, were the inclusion criteria. Risser ≥ 3 corresponded to skeletally matured. Linear and logistic binary mixed effect models accounting for patients nested into multiple measurement occasions were created to predict MCA during adolescence and progressors (progression to ≥ 30) versus non-progressors (no scoliosis or < 30 at last follow-up). Results: 161 subjects (59% females) with a mean baseline age 8.7 ± 2.4 years, follow-up of 4.2 ± 2.4 years and having reached skeletal maturity. Ultimately, 19 subjects became progressors and 142 became non-progressors. Curve magnitude at baseline was 8.8 ± 5.9° (range 0–50), at final follow-up 11.6 ± 12.4 (0–77). The mean curve acceleration was + 0.1 ± 0.5°, respectively + 0.2 ± 0.5°for non-progressors vs progressors during the acceleration phase. A linear mixed model showed that the triradiate cartilage closure accelerates MCA by 2.6 when adjusted for age and female gender. In a logistic mixed model, when the triradiate cartilage closes, the OR of reaching the MCA before the next follow-up is increased by 4.60 (CI95% 2.34–8.90 p < 0.001). No evidence for prognostic value of Risser in all derivated models. Conclusion: We found no evidence for the parameters in the coronal, sagittal nor transverse plane before curve onset acting as prognostic factors for curve behavior. In the prediction model on a longitudinal database that starts in many patients before scoliosis, no clear radiographic discriminant for later progressive scoliosis could be identified. The closure of the triradiate cartilage resulted as the best sign of pubertal spurt onset and scoliosis acceleration.
- 20 March 2025
- Published by Springer Nature in Spine Deformity
- Vol. 13 (4) , 1179-1187
Open AccessWhat can we learn from scoliosis in children with the 22q11.2 deletion syndrome? Prognostic factors at pre-adolescent age for fast progressive, mild and self-resolving forms during adolescence
Introduction: Longitudinal data starting before adolescence and before curve onset, may elucidate prognostic factors for later scoliotic curve development. The aim is to predict the maximum curve acceleration (MCA; °/month) and the final curve progression in a cohort of 22q11.2DS subjects screened for scoliosis. Methods: Scoliosis screening starts immediately after 22q11.2DS diagnosis. A minimum of 2 years follow-up, two assessments, Risser 0, open triradiate cartilage at start, were the inclusion criteria. Risser ≥ 3 corresponded to skeletally matured. Linear and logistic binary mixed effect models accounting for patients nested into multiple measurement occasions were created to predict MCA during adolescence and progressors (progression to ≥ 30) versus non-progressors (no scoliosis or < 30 at last follow-up). Results: 161 subjects (59% females) with a mean baseline age 8.7 ± 2.4 years, follow-up of 4.2 ± 2.4 years and having reached skeletal maturity. Ultimately, 19 subjects became progressors and 142 became non-progressors. Curve magnitude at baseline was 8.8 ± 5.9° (range 0–50), at final follow-up 11.6 ± 12.4 (0–77). The mean curve acceleration was + 0.1 ± 0.5°, respectively + 0.2 ± 0.5°for non-progressors vs progressors during the acceleration phase. A linear mixed model showed that the triradiate cartilage closure accelerates MCA by 2.6 when adjusted for age and female gender. In a logistic mixed model, when the triradiate cartilage closes, the OR of reaching the MCA before the next follow-up is increased by 4.60 (CI95% 2.34–8.90 p < 0.001). No evidence for prognostic value of Risser in all derivated models. Conclusion: We found no evidence for the parameters in the coronal, sagittal nor transverse plane before curve onset acting as prognostic factors for curve behavior. In the prediction model on a longitudinal database that starts in many patients before scoliosis, no clear radiographic discriminant for later progressive scoliosis could be identified. The closure of the triradiate cartilage resulted as the best sign of pubertal spurt onset and scoliosis acceleration.
- Open Access
- 25 February 2025
- Published by Elsevier in Journal of Clinical Epidemiology
- Vol. 181, 111732
Assessing the quality of prediction models in health care using the Prediction model Risk Of Bias ASsessment Tool (PROBAST): an evaluation of its use and practical application
- 25 February 2025
- Published by Elsevier in Journal of Clinical Epidemiology
- Vol. 181, 111732
Open AccessAssessing the quality of prediction models in health care using the Prediction model Risk Of Bias ASsessment Tool (PROBAST): an evaluation of its use and practical application
- Top Cited PapersOpen Access
- 26 January 2025
- Published by Wiley in Statistics in Medicine
- Vol. 44 (3-4) , e10320
The Harms of Class Imbalance Corrections for Machine Learning Based Prediction Models: A Simulation Study
Introduction: Risk prediction models are increasingly used in healthcare to aid in clinical decision‐making. In most clinical contexts, model calibration (i.e., assessing the reliability of risk estimates) is critical. Data available for model development are often not perfectly balanced with the modeled outcome (i.e., individuals with vs. without the event of interest are not equally prevalent in the data). It is common for researchers to correct for class imbalance, yet, the effect of such imbalance corrections on the calibration of machine learning models is largely unknown.Methods: We studied the effect of imbalance corrections on model calibration for a variety of machine learning algorithms. Using extensive Monte Carlo simulations we compared the out‐of‐sample predictive performance of models developed with an imbalance correction to those developed without a correction for class imbalance across different data‐generating scenarios (varying sample size, the number of predictors, and event fraction). Our findings were illustrated in a case study using MIMIC‐III data.Results: In all simulation scenarios, prediction models developed without a correction for class imbalance consistently had equal or better calibration performance than prediction models developed with a correction for class imbalance. The miscalibration introduced by correcting for class imbalance was characterized by an over‐estimation of risk and was not always able to be corrected with re‐calibration.Conclusion: Correcting for class imbalance is not always necessary and may even be harmful to clinical prediction models which aim to produce reliable risk estimates on an individual basis.
- 26 January 2025
- Published by Wiley in Statistics in Medicine
- Vol. 44 (3-4) , e10320
Top Cited PapersOpen AccessThe Harms of Class Imbalance Corrections for Machine Learning Based Prediction Models: A Simulation Study
Introduction: Risk prediction models are increasingly used in healthcare to aid in clinical decision‐making. In most clinical contexts, model calibration (i.e., assessing the reliability of risk estimates) is critical. Data available for model development are often not perfectly balanced with the modeled outcome (i.e., individuals with vs. without the event of interest are not equally prevalent in the data). It is common for researchers to correct for class imbalance, yet, the effect of such imbalance corrections on the calibration of machine learning models is largely unknown.Methods: We studied the effect of imbalance corrections on model calibration for a variety of machine learning algorithms. Using extensive Monte Carlo simulations we compared the out‐of‐sample predictive performance of models developed with an imbalance correction to those developed without a correction for class imbalance across different data‐generating scenarios (varying sample size, the number of predictors, and event fraction). Our findings were illustrated in a case study using MIMIC‐III data.Results: In all simulation scenarios, prediction models developed without a correction for class imbalance consistently had equal or better calibration performance than prediction models developed with a correction for class imbalance. The miscalibration introduced by correcting for class imbalance was characterized by an over‐estimation of risk and was not always able to be corrected with re‐calibration.Conclusion: Correcting for class imbalance is not always necessary and may even be harmful to clinical prediction models which aim to produce reliable risk estimates on an individual basis.
- Open Access
- 21 January 2025
- Published by Royal College of General Practitioners in British Journal of General Practice
- Vol. 75 (756) , e500-e508
Development and internal validation of a diagnostic prediction model for life-threatening events in callers with shortness of breath: a cross-sectional study in out-of-hours primary care
Background: Shortness of breath often prompts calls to out-of-hours primary care and is the prime reason for home visits by GPs. However, a diagnostic prediction model in this setting is lacking. Aim: To develop and internally validate a model predicting life-threatening events for out-of-hours primary care callers with shortness of breath. Design and setting: This cross-sectional study includes data from 1952 patients with shortness of breath who called Dutch out-of-hours primary care between September 2020 and August 2021. The trial was registered with the Netherlands Trial Register, number: NL9682. Method: Four logistic regression models were developed with life-threatening events as the outcome. The first model included age and gender (model 1) and then, successively, call characteristics (calling at night and someone else calling on behalf of the patient; model 2), symptoms (cough, fever, inability to speak full sentences and wheezing; model 3), and medical history and medication use (cardiovascular and/or pulmonary; model 4) were added. The models were internally validated using optimism correction via bootstrap with 1000 repetitions. Performance measures of discrimination (c-statistic) and calibration (calibration intercept and slope) were determined. Results: Approximately 17% (329/1952) of callers with shortness of breath had a life-threatening event. Model 3 performed best. This model exhibited good discriminative ability (internal validation c-statistic of 0.76, 95% confidence interval = 0.74 to 0.79) and was well calibrated. All models had a high net benefit compared with using no model. Models 3 and 4 had a higher net benefit compared with models 1 and 2. As models 3 and 4 were similar in terms of net benefit, the model with fewer parameters (model 3) is preferred. Conclusion: A prediction model consisting of age, gender, call characteristics, and symptoms holds promise for improving telephone triage of callers to out-of-hours primary care with shortness of breath. However, as the study period coincided with the COVID-19 pandemic, temporal validation is necessary before widespread implementation.
- 21 January 2025
- Published by Royal College of General Practitioners in British Journal of General Practice
- Vol. 75 (756) , e500-e508
Open AccessDevelopment and internal validation of a diagnostic prediction model for life-threatening events in callers with shortness of breath: a cross-sectional study in out-of-hours primary care
Background: Shortness of breath often prompts calls to out-of-hours primary care and is the prime reason for home visits by GPs. However, a diagnostic prediction model in this setting is lacking. Aim: To develop and internally validate a model predicting life-threatening events for out-of-hours primary care callers with shortness of breath. Design and setting: This cross-sectional study includes data from 1952 patients with shortness of breath who called Dutch out-of-hours primary care between September 2020 and August 2021. The trial was registered with the Netherlands Trial Register, number: NL9682. Method: Four logistic regression models were developed with life-threatening events as the outcome. The first model included age and gender (model 1) and then, successively, call characteristics (calling at night and someone else calling on behalf of the patient; model 2), symptoms (cough, fever, inability to speak full sentences and wheezing; model 3), and medical history and medication use (cardiovascular and/or pulmonary; model 4) were added. The models were internally validated using optimism correction via bootstrap with 1000 repetitions. Performance measures of discrimination (c-statistic) and calibration (calibration intercept and slope) were determined. Results: Approximately 17% (329/1952) of callers with shortness of breath had a life-threatening event. Model 3 performed best. This model exhibited good discriminative ability (internal validation c-statistic of 0.76, 95% confidence interval = 0.74 to 0.79) and was well calibrated. All models had a high net benefit compared with using no model. Models 3 and 4 had a higher net benefit compared with models 1 and 2. As models 3 and 4 were similar in terms of net benefit, the model with fewer parameters (model 3) is preferred. Conclusion: A prediction model consisting of age, gender, call characteristics, and symptoms holds promise for improving telephone triage of callers to out-of-hours primary care with shortness of breath. However, as the study period coincided with the COVID-19 pandemic, temporal validation is necessary before widespread implementation.