Daphné Chopard
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
Daphné Chopard has a total of 33 co-authors. Here are the top co-authors by number of shared publications:
- Irena Spasić
School of Computer Science and Informatics, Cardiff University, Cardiff, United Kingdom of Great Britain and Northern Ireland
3 shared publications - Claire Johnson
Cardiff Univ, Ctr Trials Res, Cardiff, Wales
2 shared publications - Matthias Sebastian Treder
School of Computer Science & Informatics, Cardiff University, Cardiff, United Kingdom
2 shared publications - Monica Busse
Centre for Trials Research, School of Medicine, Cardiff University
2 shared publications - Nagheen Ahmed
Cardiff Univ, Ctr Trials Res, Cardiff, Wales
2 shared publications - Padraig Corcoran
School of Computer Science and Informatics, Cardiff University, Cardiff, United Kingdom of Great Britain and Northern Ireland
2 shared publications
Latest Publications
- Open Access
A Foundation Model for Intensive Care: Unlocking Generalization across Tasks and Domains at Scale
BACKGROUND: Early recognition of physiological deterioration in critically ill patients enables timely intervention, yet much of the predictive information in routinely collected intensive care data remains underutilized. Existing prediction models are typically developed for a single task at a single institution and often degrade across hospitals with differing patient populations and treatment patterns. Whether a model pretrained on harmonized multi-institutional data can generalize across both hospitals and tasks (dual zero-shot) without task adaptation or local retraining has not been systematically evaluated. METHODS: We harmonized 16 critical care datasets from ICUs and emergency departments across the United States, Europe, and Asia, comprising over 1.1 million patient stays and over one billion data points. We pretrained ICareFM, a transformer-based foundation model, on 130 harmonized clinical concepts to estimate the probability that physiological values will cross clinician-specified thresholds within a given time horizon, enabling flexible event definitions without retraining. We evaluated ICareFM on seven prediction tasks (circulatory, respiratory, kidney, and liver organ failure, hyperglycemia, sepsis, and mortality) across nine external ICU cohorts using a local patient equivalence (LPE) framework that quantifies how many labeled patients a locally trained model requires to match ICareFM performance. RESULTS: Without task- or hospital-specific training (dual zero-shot), ICareFM achieved a median AuROC of 0.837 (95% CI, 0.797–0.858) across held-out datasets and seven prediction tasks, and outperformed task-matched clinical scores used in intensive care by AuROC +0.049 (95% CI, 0.031–0.071). It matched the performance of specialized local models trained on over 1,000 labeled patient stays. With task- and hospital-specific adaptation (staged adaptation), locally trained models required a median of 14,709 patient stays to match ICareFM; in 84% of settings, the available local training data was insufficient to train a model that would outperform staged-adapted ICareFM by a 0.5% AuROC margin. External validation on large cohorts at Robert Bosch Krankenhaus (Germany) and Charité – Universitätsmedizin Berlin (Germany) revealed staged-adapted ICareFM outperformed local models trained on over 60,000 and 100,000 patient stays, respectively. Beyond the ICU, adapted models outperformed locally trained alternatives in 9/10 emergency department and general ward settings. Pairing ICareFM with large language models enabled natural language risk queries that outperformed both native LLM prediction and clinical scoring baselines. CONCLUSIONS: A foundation model pretrained on harmonized, multi-continental critical care data generalized across hospitals and clinical tasks, reducing local data requirements and, with staged adaptation, matching or outperforming locally trained models in the majority of evaluated settings. The data harmonization code, processing pipelines, and model weights are released to support independent validation and further research.
Open AccessA Foundation Model for Intensive Care: Unlocking Generalization across Tasks and Domains at Scale
BACKGROUND: Early recognition of physiological deterioration in critically ill patients enables timely intervention, yet much of the predictive information in routinely collected intensive care data remains underutilized. Existing prediction models are typically developed for a single task at a single institution and often degrade across hospitals with differing patient populations and treatment patterns. Whether a model pretrained on harmonized multi-institutional data can generalize across both hospitals and tasks (dual zero-shot) without task adaptation or local retraining has not been systematically evaluated. METHODS: We harmonized 16 critical care datasets from ICUs and emergency departments across the United States, Europe, and Asia, comprising over 1.1 million patient stays and over one billion data points. We pretrained ICareFM, a transformer-based foundation model, on 130 harmonized clinical concepts to estimate the probability that physiological values will cross clinician-specified thresholds within a given time horizon, enabling flexible event definitions without retraining. We evaluated ICareFM on seven prediction tasks (circulatory, respiratory, kidney, and liver organ failure, hyperglycemia, sepsis, and mortality) across nine external ICU cohorts using a local patient equivalence (LPE) framework that quantifies how many labeled patients a locally trained model requires to match ICareFM performance. RESULTS: Without task- or hospital-specific training (dual zero-shot), ICareFM achieved a median AuROC of 0.837 (95% CI, 0.797–0.858) across held-out datasets and seven prediction tasks, and outperformed task-matched clinical scores used in intensive care by AuROC +0.049 (95% CI, 0.031–0.071). It matched the performance of specialized local models trained on over 1,000 labeled patient stays. With task- and hospital-specific adaptation (staged adaptation), locally trained models required a median of 14,709 patient stays to match ICareFM; in 84% of settings, the available local training data was insufficient to train a model that would outperform staged-adapted ICareFM by a 0.5% AuROC margin. External validation on large cohorts at Robert Bosch Krankenhaus (Germany) and Charité – Universitätsmedizin Berlin (Germany) revealed staged-adapted ICareFM outperformed local models trained on over 60,000 and 100,000 patient stays, respectively. Beyond the ICU, adapted models outperformed locally trained alternatives in 9/10 emergency department and general ward settings. Pairing ICareFM with large language models enabled natural language risk queries that outperformed both native LLM prediction and clinical scoring baselines. CONCLUSIONS: A foundation model pretrained on harmonized, multi-continental critical care data generalized across hospitals and clinical tasks, reducing local data requirements and, with staged adaptation, matching or outperforming locally trained models in the majority of evaluated settings. The data harmonization code, processing pipelines, and model weights are released to support independent validation and further research.
- Open Access
Comparison of carbon dioxide control during pressure controlled versus pressure regulated volume controlled ventilation in children (CoCO2): protocol for a pilot digital randomized controlled trial
Introduction: Digital trials are a promising strategy to increase the evidence base for common interventions and may convey considerable efficiency benefits in trial conduct. Although paediatric intensive care units (PICUs) are rich in routine electronic data, highly pragmatic digital trials in this field remain scarce. There are unmet evidence needs for optimal mechanical ventilation modes in paediatric intensive care. We aim to test the feasibility of a digital PICU trial comparing two modes of invasive mechanical ventilation using carbon dioxide (CO 2 ) control as the outcome measure. Methods and analysis: Single-centre, open-labelled, randomized controlled pilot trial with two parallel treatment arms comparing pressure control (PC) vs pressure-regulated volume control (PRVC). Patients are eligible if aged 2 kg, have an arterial line, and require >60 minutes of mechanical ventilation during PICU hospitalization at the University Children’s Hospital Zurich. Exclusion criteria include cardiac shunt lesions, pulmonary hypertension under treatment, and intracranial hypertension. CO 2 is measured using three methods: end-tidal (continuous), transcutaneous (continuous), and blood gas analyses (intermittent). Baseline, intervention, and outcome data are collected electronically from the patients’ routine electronic health records. The primary feasibility outcome is adherence to the assigned ventilation mode, while the primary physiological outcome is the proportion of time spent within the target range of CO 2 (end-tidal, normocarbia defined as CO 2 ≥ 4.5, ≤ 6 kPa). Both outcomes are captured digitally every minute from randomization until censoring (48 hours after randomization, extubation, discharge, or death, whichever comes first). Analysis will occur on an intention-to-treat basis. We aim to enrol 60 patients in total. Recruitment started in January 2024 and is planned to continue for 6 months. Ethics and dissemination: This study received ethical approval (BASEC 2022-00829). Study results will be disseminated through publication in a peer-reviewed journal and other media like podcasts. Trial registration number: NCT058431 ARTICLE SUMMARY: This study compares two commonly used modes of invasive mechanical ventilation in a randomized design. The trial will provide feasibility data to inform the conduct of digital trials by using electronic patient data directly extracted from the source systems, minimizing manual data collection and associated bias and thereby increasing local readiness for more efficient clinical trial conduct. Outcomes of this pilot trial relate to feasibility and physiological measures; future larger trials should also explore patient-centred outcomes. Blinding is not possible due to the nature of the intervention. Technical issues that may affect the availability or accuracy of data may arise and will be documented. Some aspects of digital trials, such as electronic informed consent, are not implemented in this trial.
Open AccessComparison of carbon dioxide control during pressure controlled versus pressure regulated volume controlled ventilation in children (CoCO2): protocol for a pilot digital randomized controlled trial
Introduction: Digital trials are a promising strategy to increase the evidence base for common interventions and may convey considerable efficiency benefits in trial conduct. Although paediatric intensive care units (PICUs) are rich in routine electronic data, highly pragmatic digital trials in this field remain scarce. There are unmet evidence needs for optimal mechanical ventilation modes in paediatric intensive care. We aim to test the feasibility of a digital PICU trial comparing two modes of invasive mechanical ventilation using carbon dioxide (CO 2 ) control as the outcome measure. Methods and analysis: Single-centre, open-labelled, randomized controlled pilot trial with two parallel treatment arms comparing pressure control (PC) vs pressure-regulated volume control (PRVC). Patients are eligible if aged 2 kg, have an arterial line, and require >60 minutes of mechanical ventilation during PICU hospitalization at the University Children’s Hospital Zurich. Exclusion criteria include cardiac shunt lesions, pulmonary hypertension under treatment, and intracranial hypertension. CO 2 is measured using three methods: end-tidal (continuous), transcutaneous (continuous), and blood gas analyses (intermittent). Baseline, intervention, and outcome data are collected electronically from the patients’ routine electronic health records. The primary feasibility outcome is adherence to the assigned ventilation mode, while the primary physiological outcome is the proportion of time spent within the target range of CO 2 (end-tidal, normocarbia defined as CO 2 ≥ 4.5, ≤ 6 kPa). Both outcomes are captured digitally every minute from randomization until censoring (48 hours after randomization, extubation, discharge, or death, whichever comes first). Analysis will occur on an intention-to-treat basis. We aim to enrol 60 patients in total. Recruitment started in January 2024 and is planned to continue for 6 months. Ethics and dissemination: This study received ethical approval (BASEC 2022-00829). Study results will be disseminated through publication in a peer-reviewed journal and other media like podcasts. Trial registration number: NCT058431 ARTICLE SUMMARY: This study compares two commonly used modes of invasive mechanical ventilation in a randomized design. The trial will provide feasibility data to inform the conduct of digital trials by using electronic patient data directly extracted from the source systems, minimizing manual data collection and associated bias and thereby increasing local readiness for more efficient clinical trial conduct. Outcomes of this pilot trial relate to feasibility and physiological measures; future larger trials should also explore patient-centred outcomes. Blinding is not possible due to the nature of the intervention. Technical issues that may affect the availability or accuracy of data may arise and will be documented. Some aspects of digital trials, such as electronic informed consent, are not implemented in this trial.
- Open Access
- 24 December 2021
- Published by JMIR Publications Inc. in JMIR Medical Informatics
- Vol. 9 (12) , e28632
Text Mining of Adverse Events in Clinical Trials: Deep Learning Approach
Background: Pharmacovigilance and safety reporting, which involve processes for monitoring the use of medicines in clinical trials, play a critical role in the identification of previously unrecognized adverse events or changes in the patterns of adverse events. Objective: This study aims to demonstrate the feasibility of automating the coding of adverse events described in the narrative section of the serious adverse event report forms to enable statistical analysis of the aforementioned patterns. Methods: We used the Unified Medical Language System (UMLS) as the coding scheme, which integrates 217 source vocabularies, thus enabling coding against other relevant terminologies such as the International Classification of Diseases–10th Revision, Medical Dictionary for Regulatory Activities, and Systematized Nomenclature of Medicine). We used MetaMap, a highly configurable dictionary lookup software, to identify the mentions of the UMLS concepts. We trained a binary classifier using Bidirectional Encoder Representations from Transformers (BERT), a transformer-based language model that captures contextual relationships, to differentiate between mentions of the UMLS concepts that represented adverse events and those that did not. Results: The model achieved a high F1 score of 0.8080, despite the class imbalance. This is 10.15 percent points lower than human-like performance but also 17.45 percent points higher than that of the baseline approach. Conclusions: These results confirmed that automated coding of adverse events described in the narrative section of serious adverse event reports is feasible. Once coded, adverse events can be statistically analyzed so that any correlations with the trialed medicines can be estimated in a timely fashion.
- 24 December 2021
- Published by JMIR Publications Inc. in JMIR Medical Informatics
- Vol. 9 (12) , e28632
Open AccessText Mining of Adverse Events in Clinical Trials: Deep Learning Approach
Background: Pharmacovigilance and safety reporting, which involve processes for monitoring the use of medicines in clinical trials, play a critical role in the identification of previously unrecognized adverse events or changes in the patterns of adverse events. Objective: This study aims to demonstrate the feasibility of automating the coding of adverse events described in the narrative section of the serious adverse event report forms to enable statistical analysis of the aforementioned patterns. Methods: We used the Unified Medical Language System (UMLS) as the coding scheme, which integrates 217 source vocabularies, thus enabling coding against other relevant terminologies such as the International Classification of Diseases–10th Revision, Medical Dictionary for Regulatory Activities, and Systematized Nomenclature of Medicine). We used MetaMap, a highly configurable dictionary lookup software, to identify the mentions of the UMLS concepts. We trained a binary classifier using Bidirectional Encoder Representations from Transformers (BERT), a transformer-based language model that captures contextual relationships, to differentiate between mentions of the UMLS concepts that represented adverse events and those that did not. Results: The model achieved a high F1 score of 0.8080, despite the class imbalance. This is 10.15 percent points lower than human-like performance but also 17.45 percent points higher than that of the baseline approach. Conclusions: These results confirmed that automated coding of adverse events described in the narrative section of serious adverse event reports is feasible. Once coded, adverse events can be statistically analyzed so that any correlations with the trialed medicines can be estimated in a timely fashion.
- Open Access
- 9 March 2021
- Published by JMIR Publications Inc. in JMIR Preprints
Text Mining of Adverse Events in Clinical Trials: Deep Learning Approach (Preprint)
BACKGROUND Pharmacovigilance and safety reporting, which involve processes for monitoring the use of medicines in clinical trials, play a critical role in the identification of previously unrecognized adverse events or changes in the patterns of adverse events. OBJECTIVE This study aims to demonstrate the feasibility of automating the coding of adverse events described in the narrative section of the serious adverse event report forms to enable statistical analysis of the aforementioned patterns. METHODS We used the Unified Medical Language System (UMLS) as the coding scheme, which integrates 217 source vocabularies, thus enabling coding against other relevant terminologies such as the International Classification of Diseases–10th Revision, Medical Dictionary for Regulatory Activities, and Systematized Nomenclature of Medicine). We used MetaMap, a highly configurable dictionary lookup software, to identify the mentions of the UMLS concepts. We trained a binary classifier using Bidirectional Encoder Representations from Transformers (BERT), a transformer-based language model that captures contextual relationships, to differentiate between mentions of the UMLS concepts that represented adverse events and those that did not. RESULTS The model achieved a high F1 score of 0.8080, despite the class imbalance. This is 10.15 percent points lower than human-like performance but also 17.45 percent points higher than that of the baseline approach. CONCLUSIONS These results confirmed that automated coding of adverse events described in the narrative section of serious adverse event reports is feasible. Once coded, adverse events can be statistically analyzed so that any correlations with the trialed medicines can be estimated in a timely fashion.
- 9 March 2021
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessText Mining of Adverse Events in Clinical Trials: Deep Learning Approach (Preprint)
BACKGROUND Pharmacovigilance and safety reporting, which involve processes for monitoring the use of medicines in clinical trials, play a critical role in the identification of previously unrecognized adverse events or changes in the patterns of adverse events. OBJECTIVE This study aims to demonstrate the feasibility of automating the coding of adverse events described in the narrative section of the serious adverse event report forms to enable statistical analysis of the aforementioned patterns. METHODS We used the Unified Medical Language System (UMLS) as the coding scheme, which integrates 217 source vocabularies, thus enabling coding against other relevant terminologies such as the International Classification of Diseases–10th Revision, Medical Dictionary for Regulatory Activities, and Systematized Nomenclature of Medicine). We used MetaMap, a highly configurable dictionary lookup software, to identify the mentions of the UMLS concepts. We trained a binary classifier using Bidirectional Encoder Representations from Transformers (BERT), a transformer-based language model that captures contextual relationships, to differentiate between mentions of the UMLS concepts that represented adverse events and those that did not. RESULTS The model achieved a high F1 score of 0.8080, despite the class imbalance. This is 10.15 percent points lower than human-like performance but also 17.45 percent points higher than that of the baseline approach. CONCLUSIONS These results confirmed that automated coding of adverse events described in the narrative section of serious adverse event reports is feasible. Once coded, adverse events can be statistically analyzed so that any correlations with the trialed medicines can be estimated in a timely fashion.
- 27 September 2019
- Published by Springer Nature
- p. 71-82
A Deep Learning Approach to Self-expansion of Abbreviations Based on Morphology and Context Distance
- 27 September 2019
- Published by Springer Nature
- p. 71-82
A Deep Learning Approach to Self-expansion of Abbreviations Based on Morphology and Context Distance