Joeran Sander Bosma
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
Research Subjects
Joeran Sander Bosma has a total of 347 co-authors. Here are the top co-authors by number of shared publications:
- Henkjan Huisman
Radboud University Medical Center, Radboud Institute for Health Sciences, Nijmegen
10 shared publications - Alessa Hering
Radboudumc
6 shared publications - Maarten de Rooij
Department of Medical Imaging, Radboud University Medical Center, Nijmegen, The Netherlands
6 shared publications - Anindo Saha
Diagnostic Image Analysis Group, Radboud University Medical Center, Nijmegen, The Netherlands
5 shared publications - Luc Builtjes
Diagnostic Image Analysis Group, Medical Imaging Department, Radboud University Medical Center, Geert Grooteplein Zuid 10, Nijmegen, 6525 GA, The Netherlands, 31 614021323
5 shared publications - Bram van Ginneken
Radboud University Medical Center
4 shared publications
Latest Publications
- Open Access
- 6 October 2026
- Published by Elsevier in SSRN Electronic Journal
GHMSCHRFT: Transformer-based de-identification with realistic surrogate replacement for Dutch clinical reports
Background and Objective: Clinical free text can only be reused once protected health information (PHI) has been removed or replaced. No openly available transformer model for Dutch de-identification has been trained on real clinical text, and Hiding in Plain Sight (HIPS), which replaces detected PHI with realistic surrogates so that missed identifiers blend in, had no Dutch implementation. We developed GHMSCHRFT, an open pipeline that pairs transformer-based PHI detection with Dutch HIPS, and evaluated both its detection and the residual risk of its output.Methods: An XLM-RoBERTa-large model pretrained on Dutch medical text was fine-tuned on 3,941 reports from two hospitals, annotated for 24 PHI categories, and combined with a Dutch HIPS surrogate module. Detection was tested on 980 internal reports and on 340 reports from a third hospital not used for development, and compared with four existing systems in their default configuration. Detection recall counted a PHI span as found when all its tokens were tagged as PHI, whatever the predicted category, because such spans are replaced. Five readers then searched 100 processed reports, including every internal report with missed PHI, for identifiers that had survived.Results: Entity-level F1-score was 0.943 internally and 0.950 externally, and detection recall was 0.983 and 0.986. On the categories all systems share, GHMSCHRFT reached a detection recall of 0.985 and 0.986, against 0.790 and 0.883 for the strongest comparator, a neural Dutch tagger, and 0.682 or lower for the other three. The 100 reports contained 21 residual PHI spans among 1,039 surrogates. The readers together found 5 of the 21, and no reader's precision exceeded 8.3%, consistent with random selection among all PHI-like candidates.Conclusions: Near-identical performance at an unseen hospital suggests that GHMSCHRFT can be used at other Dutch hospitals without retraining, and it detected substantially more PHI than existing systems. After surrogate replacement, readers could not reliably tell missed identifiers apart from the surrogates around them, although this is a detectability test on 21 spans rather than a re-identification risk assessment. The model, surrogate module, annotation guidelines and evaluation protocol are openly released.
- 6 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessGHMSCHRFT: Transformer-based de-identification with realistic surrogate replacement for Dutch clinical reports
Background and Objective: Clinical free text can only be reused once protected health information (PHI) has been removed or replaced. No openly available transformer model for Dutch de-identification has been trained on real clinical text, and Hiding in Plain Sight (HIPS), which replaces detected PHI with realistic surrogates so that missed identifiers blend in, had no Dutch implementation. We developed GHMSCHRFT, an open pipeline that pairs transformer-based PHI detection with Dutch HIPS, and evaluated both its detection and the residual risk of its output.Methods: An XLM-RoBERTa-large model pretrained on Dutch medical text was fine-tuned on 3,941 reports from two hospitals, annotated for 24 PHI categories, and combined with a Dutch HIPS surrogate module. Detection was tested on 980 internal reports and on 340 reports from a third hospital not used for development, and compared with four existing systems in their default configuration. Detection recall counted a PHI span as found when all its tokens were tagged as PHI, whatever the predicted category, because such spans are replaced. Five readers then searched 100 processed reports, including every internal report with missed PHI, for identifiers that had survived.Results: Entity-level F1-score was 0.943 internally and 0.950 externally, and detection recall was 0.983 and 0.986. On the categories all systems share, GHMSCHRFT reached a detection recall of 0.985 and 0.986, against 0.790 and 0.883 for the strongest comparator, a neural Dutch tagger, and 0.682 or lower for the other three. The 100 reports contained 21 residual PHI spans among 1,039 surrogates. The readers together found 5 of the 21, and no reader's precision exceeded 8.3%, consistent with random selection among all PHI-like candidates.Conclusions: Near-identical performance at an unseen hospital suggests that GHMSCHRFT can be used at other Dutch hospitals without retraining, and it detected substantially more PHI than existing systems. After surrogate replacement, readers could not reliably tell missed identifiers apart from the surrogates around them, although this is a detectability test on 21 spans rather than a re-identification risk assessment. The model, surrogate module, annotation guidelines and evaluation protocol are openly released.
- 1 October 2026
- Published by Radiological Society of North America (RSNA) in Radiology
- Vol. 321 (1)
Artificial Intelligence for Prostate Cancer Detection and Screening on MRI in Global Populations: The SCARLET-1 Study
In this study of 22,445 prostate MRI examinations from 22 countries, an AI system’s performance was comparable to radiology readings for prostate cancer detection in external diagnostic and screening datasets.
- 1 October 2026
- Published by Radiological Society of North America (RSNA) in Radiology
- Vol. 321 (1)
Artificial Intelligence for Prostate Cancer Detection and Screening on MRI in Global Populations: The SCARLET-1 Study
In this study of 22,445 prostate MRI examinations from 22 countries, an AI system’s performance was comparable to radiology readings for prostate cancer detection in external diagnostic and screening datasets.
- 1 August 2026
- Published by Radiological Society of North America (RSNA) in Radiology
- Vol. 320 (2)
Diagnostically Calibrated AI Image Quality Framework Applied to Prostate Cancer Detection at MRI
Deep learning framework-derived image quality thresholds progressively excluding low-quality MRI scans correlated with improved diagnostic performance of an artificial intelligence model and radiologist accuracy in detecting clinically significant prostate cancer.
- 1 August 2026
- Published by Radiological Society of North America (RSNA) in Radiology
- Vol. 320 (2)
Diagnostically Calibrated AI Image Quality Framework Applied to Prostate Cancer Detection at MRI
Deep learning framework-derived image quality thresholds progressively excluding low-quality MRI scans correlated with improved diagnostic performance of an artificial intelligence model and radiologist accuracy in detecting clinically significant prostate cancer.
- Top Cited PapersOpen Access
- 30 April 2026
- Published by JMIR Publications Inc. in JMIR Medical Informatics
- Vol. 14 (1) , e81500
Large Language Model Automated Extraction of Clinical Signs and Symptoms From Emergency Department Reports for Machine Learning Prediction Models: Development and Validation Study
Background: Most clinically relevant information in emergency department (ED) visits is documented in free text, limiting reuse for research and clinical decision support. Despite growing interest in large language model (LLM)–based feature extraction, very few studies have examined it directly on ED reports. Existing work has mainly addressed binary tasks and rarely evaluated their impact on downstream prediction models. Furthermore, evidence for small multilingual LLMs remains limited, especially for underrepresented languages such as Dutch. Locally deployable LLMs could enable automated feature extraction for decision support systems without increasing physician workload. Objective: We aim to evaluate whether a small open-source LLM (Qwen 2.5:14B) can automatically extract 16 clinical signs and symptoms from ED reports and use these as input for an appendicitis prediction model. LLM performance under minimal and optimized 0-shot prompts was assessed against researcher annotations (reference standard) and physician annotations. Methods: This retrospective study used 336 ED reports from patients presenting with acute abdominal pain to a Dutch teaching hospital (2016-2023). One hundred reports were randomly selected to develop a minimal and an optimized 0-shot prompt strategy. The remaining 236 reports, reserved for evaluation, were annotated by 2 ED physicians and processed by the LLM to extract 16 signs and symptoms, covering binary, multiclass, and multilabel classification tasks. These features were used as input to the HIVE (History, Intake, Vitals, Examination) appendicitis prediction model. LLM extraction accuracy, sensitivity, and specificity were measured against the researcher’s (reference standard) and physician annotations. The HIVE model’s area under the receiver operating characteristic curve was evaluated using LLM-extracted vs physician-annotated features. Results: Among 336 ED reports from patients with acute abdominal pain (median age 41, IQR 22‐62 years, 205/336, 61% female), 50% (167/336) had appendicitis. The LLM achieved weighted average accuracies of 0.910 (95% CI (0.018) with minimal prompts and 0.929 (95% CI ±0.016) with optimized prompts, vs 0.961 (95% CI ±0.012) and 0.951 (95% CI ±0.015) for physicians. Corresponding HIVE model area under the receiver operating characteristic curves were 0.871 (95% CI ±0.019) and 0.911 (95% CI ±0.014) with LLM inputs under the minimal and optimized prompts, compared to 0.917 (95% CI ±0.015) and 0.924 (95% CI ±0.018) for physician inputs. Conclusions: A small locally deployable multilingual LLM can approach physician-level accuracy in extracting structured binary, multiclass, and multilabel clinical data from free-text Dutch ED reports, while preserving patient privacy, interpretability, and statistical transparency for downstream diagnostic modeling.
- 30 April 2026
- Published by JMIR Publications Inc. in JMIR Medical Informatics
- Vol. 14 (1) , e81500
Top Cited PapersOpen AccessLarge Language Model Automated Extraction of Clinical Signs and Symptoms From Emergency Department Reports for Machine Learning Prediction Models: Development and Validation Study
Background: Most clinically relevant information in emergency department (ED) visits is documented in free text, limiting reuse for research and clinical decision support. Despite growing interest in large language model (LLM)–based feature extraction, very few studies have examined it directly on ED reports. Existing work has mainly addressed binary tasks and rarely evaluated their impact on downstream prediction models. Furthermore, evidence for small multilingual LLMs remains limited, especially for underrepresented languages such as Dutch. Locally deployable LLMs could enable automated feature extraction for decision support systems without increasing physician workload. Objective: We aim to evaluate whether a small open-source LLM (Qwen 2.5:14B) can automatically extract 16 clinical signs and symptoms from ED reports and use these as input for an appendicitis prediction model. LLM performance under minimal and optimized 0-shot prompts was assessed against researcher annotations (reference standard) and physician annotations. Methods: This retrospective study used 336 ED reports from patients presenting with acute abdominal pain to a Dutch teaching hospital (2016-2023). One hundred reports were randomly selected to develop a minimal and an optimized 0-shot prompt strategy. The remaining 236 reports, reserved for evaluation, were annotated by 2 ED physicians and processed by the LLM to extract 16 signs and symptoms, covering binary, multiclass, and multilabel classification tasks. These features were used as input to the HIVE (History, Intake, Vitals, Examination) appendicitis prediction model. LLM extraction accuracy, sensitivity, and specificity were measured against the researcher’s (reference standard) and physician annotations. The HIVE model’s area under the receiver operating characteristic curve was evaluated using LLM-extracted vs physician-annotated features. Results: Among 336 ED reports from patients with acute abdominal pain (median age 41, IQR 22‐62 years, 205/336, 61% female), 50% (167/336) had appendicitis. The LLM achieved weighted average accuracies of 0.910 (95% CI (0.018) with minimal prompts and 0.929 (95% CI ±0.016) with optimized prompts, vs 0.961 (95% CI ±0.012) and 0.951 (95% CI ±0.015) for physicians. Corresponding HIVE model area under the receiver operating characteristic curves were 0.871 (95% CI ±0.019) and 0.911 (95% CI ±0.014) with LLM inputs under the minimal and optimized prompts, compared to 0.917 (95% CI ±0.015) and 0.924 (95% CI ±0.018) for physician inputs. Conclusions: A small locally deployable multilingual LLM can approach physician-level accuracy in extracting structured binary, multiclass, and multilabel clinical data from free-text Dutch ED reports, while preserving patient privacy, interpretability, and statistical transparency for downstream diagnostic modeling.
- 12 March 2026
- Published by Springer Nature
- p. 48-54
Modular Pipeline for Rapidly Evaluating Foundation Models in Medical Imaging
- 12 March 2026
- Published by Springer Nature
- p. 48-54
Modular Pipeline for Rapidly Evaluating Foundation Models in Medical Imaging
- Open Access
- 12 March 2026
- Published by Springer Nature
- p. 92
Abstract: Leveraging Open-source Language Models for Clinical Information Extraction
- 12 March 2026
- Published by Springer Nature
- p. 92
Open AccessAbstract: Leveraging Open-source Language Models for Clinical Information Extraction
- Open Access
- 21 November 2025
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 199, 111321
Scalable Clinical Annotation with Location Evidence (SCALE)
- 21 November 2025
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 199, 111321
Open AccessScalable Clinical Annotation with Location Evidence (SCALE)
- Open Access
- 2 September 2025
- Published by Oxford University Press (OUP) in JAMIA Open
- Vol. 8 (5)
Leveraging open-source large language models for clinical information extraction in resource-constrained settings
Objective: We aimed to evaluate the zero-shot performance of open-source generative large language models (LLMs) on clinical information extraction from Dutch medical reports using the Diagnostic Report Analysis: General Optimization of NLP (DRAGON) benchmark. Methods: We developed and released the llm_extractinator framework, a scalable, open-source tool for automating information extraction from clinical texts using LLMs. We evaluated 9 multilingual open-source LLMs across 28 tasks in the DRAGON benchmark, covering classification, regression, and named entity recognition (NER). All tasks were performed in a zero-shot setting. Model performance was quantified using task-specific metrics and aggregated into a DRAGON utility score. Additionally, we investigated the effect of in-context translation to English. Results: Llama-3.3-70B achieved the highest utility score (0.760), followed by Phi-4-14B (0.751), Qwen-2.5-14B (0.748), and DeepSeek-R1-14B (0.744). These models outperformed or matched a fine-tuned RoBERTa baseline on 17 of 28 tasks, particularly in regression and structured classification. NER performance was consistently low across all models. Translation to English consistently reduced performance. Discussion: Generative LLMs demonstrated strong zero-shot capabilities on clinical natural language processing tasks involving structured inference. Models around 14B parameters performed well overall, with Llama-3.3-70B leading but at high computational cost. Generative models excelled in regression tasks, but were hindered by token-level output formats for NER. Translation to English reduced performance, emphasizing the need for native language support. Conclusion: Open-source generative LLMs provide a viable zero-shot alternative for clinical information extraction from Dutch medical texts, particularly in low-resource and multilingual settings.
- 2 September 2025
- Published by Oxford University Press (OUP) in JAMIA Open
- Vol. 8 (5)
Open AccessLeveraging open-source large language models for clinical information extraction in resource-constrained settings
Objective: We aimed to evaluate the zero-shot performance of open-source generative large language models (LLMs) on clinical information extraction from Dutch medical reports using the Diagnostic Report Analysis: General Optimization of NLP (DRAGON) benchmark. Methods: We developed and released the llm_extractinator framework, a scalable, open-source tool for automating information extraction from clinical texts using LLMs. We evaluated 9 multilingual open-source LLMs across 28 tasks in the DRAGON benchmark, covering classification, regression, and named entity recognition (NER). All tasks were performed in a zero-shot setting. Model performance was quantified using task-specific metrics and aggregated into a DRAGON utility score. Additionally, we investigated the effect of in-context translation to English. Results: Llama-3.3-70B achieved the highest utility score (0.760), followed by Phi-4-14B (0.751), Qwen-2.5-14B (0.748), and DeepSeek-R1-14B (0.744). These models outperformed or matched a fine-tuned RoBERTa baseline on 17 of 28 tasks, particularly in regression and structured classification. NER performance was consistently low across all models. Translation to English consistently reduced performance. Discussion: Generative LLMs demonstrated strong zero-shot capabilities on clinical natural language processing tasks involving structured inference. Models around 14B parameters performed well overall, with Llama-3.3-70B leading but at high computational cost. Generative models excelled in regression tasks, but were hindered by token-level output formats for NER. Translation to English reduced performance, emphasizing the need for native language support. Conclusion: Open-source generative LLMs provide a viable zero-shot alternative for clinical information extraction from Dutch medical texts, particularly in low-resource and multilingual settings.
- 1 September 2025
- Published by Radiological Society of North America (RSNA) in Radiology: Artificial Intelligence
- Vol. 7 (5) , e240485
Optimizing Federated Learning Configurations for MRI Prostate Segmentation and Cancer Detection: A Simulation Study
Federated learning enables collaborative model training, enhancing the performance and generalizability of trained models for MRI-based prostate gland segmentation and clinically significant prostate cancer detection compared with locally trained models.
- 1 September 2025
- Published by Radiological Society of North America (RSNA) in Radiology: Artificial Intelligence
- Vol. 7 (5) , e240485
Optimizing Federated Learning Configurations for MRI Prostate Segmentation and Cancer Detection: A Simulation Study
Federated learning enables collaborative model training, enhancing the performance and generalizability of trained models for MRI-based prostate gland segmentation and clinically significant prostate cancer detection compared with locally trained models.
- 20 August 2025
- Published by Radiological Society of North America (RSNA) in Radiology: Artificial Intelligence
- p. e240636
Characterizing the Impact of Training Data on Generalizability: Application in Deep Learning to Estimate Lung Nodule Malignancy Risk
“Just Accepted” papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To investigate the relationship between training data volume and performance of a deep learning AI algorithm developed to assess the malignancy risk of pulmonary nodules detected on low-dose CT scans in lung cancer screening. Materials and Methods This retrospective study used a dataset of 16077 annotated nodules (1249 malignant, 14828 benign) from the National Lung Screening Trial (NLST) to systematically train an AI algorithm for pulmonary nodule malignancy risk prediction across various stratified subsets ranging from 1.25% to the full dataset. External testing was conducted using data from the Danish Lung Cancer Screening Trial (DLCST) to determine the amount of training data at which the performance of the AI was statistically non-inferior to the AI trained on the full NLST cohort. A size-matched cancer-enriched subset of DLCST, where each malignant nodule had been paired in diameter with the closest two benign nodules, was used to investigate the amount of training data at which the performance of the AI algorithm was statistically non-inferior to the average performance of 11 clinicians. Results The external testing set included 599 participants (mean age 57.65 (SD 4.84) for females and mean age 59.03 (SD 4.94) for males) with 883 nodules (65 malignant, 818 benign). The AI achieved a mean AUC of 0.92 [95% CI: 0.88, 0.96] on the DLCST cohort when trained on the full NLST dataset. Training with 80% of NLST data resulted in non-inferior performance (mean AUC 0.92 [95%CI: 0.89, 0.96], P = .005). On the size-matched DLCST subset (59 malignant, 118 benign), the AI reached non-inferior clinician-level performance (mean AUC 0.82 [95% CI: 0.77, 0.86]) with 20% of the training data ( P = .02). Conclusion The deep learning AI algorithm demonstrated excellent performance in assessing pulmonary nodule malignancy risk, achieving clinical level performance with a fraction of the training data and reaching peak performance before utilizing the full dataset. ©RSNA, 2025
- 20 August 2025
- Published by Radiological Society of North America (RSNA) in Radiology: Artificial Intelligence
- p. e240636
Characterizing the Impact of Training Data on Generalizability: Application in Deep Learning to Estimate Lung Nodule Malignancy Risk
“Just Accepted” papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To investigate the relationship between training data volume and performance of a deep learning AI algorithm developed to assess the malignancy risk of pulmonary nodules detected on low-dose CT scans in lung cancer screening. Materials and Methods This retrospective study used a dataset of 16077 annotated nodules (1249 malignant, 14828 benign) from the National Lung Screening Trial (NLST) to systematically train an AI algorithm for pulmonary nodule malignancy risk prediction across various stratified subsets ranging from 1.25% to the full dataset. External testing was conducted using data from the Danish Lung Cancer Screening Trial (DLCST) to determine the amount of training data at which the performance of the AI was statistically non-inferior to the AI trained on the full NLST cohort. A size-matched cancer-enriched subset of DLCST, where each malignant nodule had been paired in diameter with the closest two benign nodules, was used to investigate the amount of training data at which the performance of the AI algorithm was statistically non-inferior to the average performance of 11 clinicians. Results The external testing set included 599 participants (mean age 57.65 (SD 4.84) for females and mean age 59.03 (SD 4.94) for males) with 883 nodules (65 malignant, 818 benign). The AI achieved a mean AUC of 0.92 [95% CI: 0.88, 0.96] on the DLCST cohort when trained on the full NLST dataset. Training with 80% of NLST data resulted in non-inferior performance (mean AUC 0.92 [95%CI: 0.89, 0.96], P = .005). On the size-matched DLCST subset (59 malignant, 118 benign), the AI reached non-inferior clinician-level performance (mean AUC 0.82 [95% CI: 0.77, 0.86]) with 20% of the training data ( P = .02). Conclusion The deep learning AI algorithm demonstrated excellent performance in assessing pulmonary nodule malignancy risk, achieving clinical level performance with a fraction of the training data and reaching peak performance before utilizing the full dataset. ©RSNA, 2025