Matthew Watson
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Matthew Watson has a total of 32 co-authors. Here are the top co-authors by number of shared publications:
- Noura Al Moubayed
Department of Computer Science, Durham University, Durham, UK
12 shared publications - David Budgen
Department of Computer Science, Durham University, Durham, U.K.
4 shared publications Bashar Awwad Shiekh Hasan
Department of Computer Science, Durham University, Durham, UK
3 shared publications- Ali Alessa
Department of Information Technology Programs, Institute of Public Administration, Riyadh, Saudi Arabia
2 shared publications Benedict Jones
Evergreen Life, Manchester, UK
2 shared publications- Chris Kennelly
Cievert Ltd Newcastle upon Tyne United Kingdom
2 shared publications
Latest Publications
- Open Access
- 8 July 2026
- Published by Elsevier in Medical Image Analysis
- Vol. 113, 104191
Controllable synthesis of dermoscopic images using diffusion models for enhanced computer aided diagnosis and detection
- 8 July 2026
- Published by Elsevier in Medical Image Analysis
- Vol. 113, 104191
Open AccessControllable synthesis of dermoscopic images using diffusion models for enhanced computer aided diagnosis and detection
- Open Access
- 13 June 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
Generalizable multilingual medical text anonymization using generative instruction tuning
Medical research depends on access to high quality data that protects patient privacy. Free text in health records contains valuable clinical detail, yet it often includes sensitive personal information that must be removed before use. Current approaches rely on manually created training data and focus mainly on narrow domains. They are difficult to scale to new medical fields and languages. This study aims to address these limitations by developing a framework that supports privacy-preserving use of medical text across diverse settings. The study introduces an annotation-free framework for training and adapting LLM-based anonymization models across diverse medical domains. Our reproducible framework includes the development of a generative medical anonymization model, leveraging synthetic data and instruction tuning of generative LLMs. Performance is evaluated on both synthetic test sets and on patient requests from a digital triage service. Accuracy, recall, precision, and the ability to maintain the original meaning of non-sensitive text are assessed. Here we show that generative models trained with the synthetic framework reach performance that exceeds strong baseline systems across several medical domains. The models preserve non-sensitive text with high fidelity and anonymize sensitive information with high accuracy. They perform well even when trained on small datasets, generalize to unseen clinical fields, and support anonymization in multiple languages without requiring additional training data in those languages. The study presents a reproducible, annotation-free approach that enables the development of effective anonymization models for medical text. The framework reduces reliance on real patient data, lowers the cost of adaptation to new settings, and supports wider use of unstructured clinical information for research and service improvement.
- 13 June 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
Open AccessGeneralizable multilingual medical text anonymization using generative instruction tuning
Medical research depends on access to high quality data that protects patient privacy. Free text in health records contains valuable clinical detail, yet it often includes sensitive personal information that must be removed before use. Current approaches rely on manually created training data and focus mainly on narrow domains. They are difficult to scale to new medical fields and languages. This study aims to address these limitations by developing a framework that supports privacy-preserving use of medical text across diverse settings. The study introduces an annotation-free framework for training and adapting LLM-based anonymization models across diverse medical domains. Our reproducible framework includes the development of a generative medical anonymization model, leveraging synthetic data and instruction tuning of generative LLMs. Performance is evaluated on both synthetic test sets and on patient requests from a digital triage service. Accuracy, recall, precision, and the ability to maintain the original meaning of non-sensitive text are assessed. Here we show that generative models trained with the synthetic framework reach performance that exceeds strong baseline systems across several medical domains. The models preserve non-sensitive text with high fidelity and anonymize sensitive information with high accuracy. They perform well even when trained on small datasets, generalize to unseen clinical fields, and support anonymization in multiple languages without requiring additional training data in those languages. The study presents a reproducible, annotation-free approach that enables the development of effective anonymization models for medical text. The framework reduces reliance on real patient data, lowers the cost of adaptation to new settings, and supports wider use of unstructured clinical information for research and service improvement.
- Open Access
- 19 March 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
Multimodal models for skin cancer classification using clinical freetext and dermatoscopic images
Background Skin cancer is one of the most prevalent cancers globally, with early detection critical to ensure reduced mortality risk. To aid early detection, machine learning (ML) skin cancer detection models have been proposed, currently with a focus on dermatoscopic imaging only. However, freetext may provide extra diagnostic information that is not present in images alone. Methods We constructed a multimodal dataset comprising 5481 dermatoscopic images from 4538 patients, including patient metadata and clinical notes, with binary labels (benign vs. malignant, 7% malignant). To assess and mitigate bias from leading language, we developed a clinical text preprocessing pipeline combining regular expressions and large language models, enabling multiple levels of filtering. We train multimodal ML models on this dataset to explore the effect of freetext on model performance. Results Our results show that incorporating unfiltered text significantly improves classification performance (0.970 AUROC) compared to visual data alone (0.909 AUROC); even with leading language removed, performance gains persist (0.948 AUROC). Conclusions This work benchmarks clinical freetext inclusion in skin lesion classification, demonstrating that clinical text contributes predictive value beyond that available in images alone. The model’s high performance on unfiltered clinical text highlights the high levels of bias, and possible shortcutting, present in this text which may make it unsuitable for inclusion in some ML models. By systematically filtering clinical notes via our proposed technique, we show that multimodal models retain improved accuracy while reducing bias. These results provide practical guidance for integrating clinical text into real-world skin cancer detection systems and establish a foundation for future multimodal research in dermatology.
- 19 March 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
Open AccessMultimodal models for skin cancer classification using clinical freetext and dermatoscopic images
Background Skin cancer is one of the most prevalent cancers globally, with early detection critical to ensure reduced mortality risk. To aid early detection, machine learning (ML) skin cancer detection models have been proposed, currently with a focus on dermatoscopic imaging only. However, freetext may provide extra diagnostic information that is not present in images alone. Methods We constructed a multimodal dataset comprising 5481 dermatoscopic images from 4538 patients, including patient metadata and clinical notes, with binary labels (benign vs. malignant, 7% malignant). To assess and mitigate bias from leading language, we developed a clinical text preprocessing pipeline combining regular expressions and large language models, enabling multiple levels of filtering. We train multimodal ML models on this dataset to explore the effect of freetext on model performance. Results Our results show that incorporating unfiltered text significantly improves classification performance (0.970 AUROC) compared to visual data alone (0.909 AUROC); even with leading language removed, performance gains persist (0.948 AUROC). Conclusions This work benchmarks clinical freetext inclusion in skin lesion classification, demonstrating that clinical text contributes predictive value beyond that available in images alone. The model’s high performance on unfiltered clinical text highlights the high levels of bias, and possible shortcutting, present in this text which may make it unsuitable for inclusion in some ML models. By systematically filtering clinical notes via our proposed technique, we show that multimodal models retain improved accuracy while reducing bias. These results provide practical guidance for integrating clinical text into real-world skin cancer detection systems and establish a foundation for future multimodal research in dermatology.
- 27 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. PP (15209210) , 1-11
AttenCraft: Attention-based Disentanglement of Multiple Concepts for Text-to-Image Customization
Text-to-image (T2I) customization empowers users to adapt the T2I diffusion model to new concepts absent in the pre-training dataset. On this basis, capturing multiple new concepts from a single image has emerged as a new task, allowing the model to learn multiple concepts simultaneously or discard unwanted concepts. However, multiple-concept disentanglement remains a key challenge. Existing disentanglement models often exhibit two main issues: feature fusion and asynchronous learning across different concepts. To address these issues, we propose AttenCraft, an attention-based method for multiple-concept disentanglement. Our method uses attention maps to generate accurate masks for each concept in a single initialization step, aiding in concept disentanglement without requiring mask preparation from humans or specialized models. Moreover, we introduce an adaptive algorithm based on attention scores to estimate sampling ratios for different concepts, promoting balanced feature acquisition and synchronized learning. AttenCraft also introduces a feature-retaining training framework that employs various loss functions to enhance feature recognition and prevent fusion. Extensive experiments show that our model effectively mitigates these two issues, achieving state-of-the-art image fidelity and comparable prompt fidelity to baseline models.
- 27 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. PP (15209210) , 1-11
AttenCraft: Attention-based Disentanglement of Multiple Concepts for Text-to-Image Customization
Text-to-image (T2I) customization empowers users to adapt the T2I diffusion model to new concepts absent in the pre-training dataset. On this basis, capturing multiple new concepts from a single image has emerged as a new task, allowing the model to learn multiple concepts simultaneously or discard unwanted concepts. However, multiple-concept disentanglement remains a key challenge. Existing disentanglement models often exhibit two main issues: feature fusion and asynchronous learning across different concepts. To address these issues, we propose AttenCraft, an attention-based method for multiple-concept disentanglement. Our method uses attention maps to generate accurate masks for each concept in a single initialization step, aiding in concept disentanglement without requiring mask preparation from humans or specialized models. Moreover, we introduce an adaptive algorithm based on attention scores to estimate sampling ratios for different concepts, promoting balanced feature acquisition and synchronized learning. AttenCraft also introduces a feature-retaining training framework that employs various loss functions to enhance feature recognition and prevent fusion. Extensive experiments show that our model effectively mitigates these two issues, achieving state-of-the-art image fidelity and comparable prompt fidelity to baseline models.
- 15 January 2025
- Published by Springer Nature in Internal and Emergency Medicine
- Vol. 20 (3) , 681-690
The variable relationship between the National Early Warning Score on admission to hospital, the primary discharge diagnosis, and in-hospital mortality
Background Patients with an elevated admission National Early Warning Score (NEWS) are more likely to die while in hospital. However, it is not known if this increased mortality risk is the same for all diagnoses. The aim of this study was to determine and compare the increased risk of in-hospital mortality associated with an elevated NEWS and different primary discharge diagnoses in unselected emergency admissions to a UK university teaching hospital. Methods A non-interventional observational study of 122,321 consecutive, unselected, adult patients with complete data admitted as an emergency between 2014 and 2022. Results The overall in-hospital mortality was 4.3%. Eighty diagnostic groupings accounted for 85.8% of all admissions and 89.4% of all in-hospital deaths. Depending on diagnostic grouping, the risk of mortality associated with an admission NEWS ≥ 3 ranged from 2.3- to 100-fold. For example, the in-hospital mortality of COPD patients increased from 1.9% for those with admission NEWS < 3 to 35.6% for those with NEWS ≥ 3, for chest pain mortality increased from 0.1 to 3.9%, and for patients with an opiate overdose from 0.2 to 7.7%. Conversely, for admission NEWS < 3, aspiration pneumonia and intracranial hemorrhage had in-hospital mortalities of 13.7% and 12.1%, respectively. Discussion There is enormous variation in the mortality risk associated with an increased admission NEWS in different commonly encountered diagnoses. Therefore, the mortality risk of some ‘low risk’ conditions can be dramatically increased if their admission NEWS is elevated, whereas some ‘high risk’ conditions are still likely to die even if their admission NEWS is low.
- 15 January 2025
- Published by Springer Nature in Internal and Emergency Medicine
- Vol. 20 (3) , 681-690
The variable relationship between the National Early Warning Score on admission to hospital, the primary discharge diagnosis, and in-hospital mortality
Background Patients with an elevated admission National Early Warning Score (NEWS) are more likely to die while in hospital. However, it is not known if this increased mortality risk is the same for all diagnoses. The aim of this study was to determine and compare the increased risk of in-hospital mortality associated with an elevated NEWS and different primary discharge diagnoses in unselected emergency admissions to a UK university teaching hospital. Methods A non-interventional observational study of 122,321 consecutive, unselected, adult patients with complete data admitted as an emergency between 2014 and 2022. Results The overall in-hospital mortality was 4.3%. Eighty diagnostic groupings accounted for 85.8% of all admissions and 89.4% of all in-hospital deaths. Depending on diagnostic grouping, the risk of mortality associated with an admission NEWS ≥ 3 ranged from 2.3- to 100-fold. For example, the in-hospital mortality of COPD patients increased from 1.9% for those with admission NEWS < 3 to 35.6% for those with NEWS ≥ 3, for chest pain mortality increased from 0.1 to 3.9%, and for patients with an opiate overdose from 0.2 to 7.7%. Conversely, for admission NEWS < 3, aspiration pneumonia and intracranial hemorrhage had in-hospital mortalities of 13.7% and 12.1%, respectively. Discussion There is enormous variation in the mortality risk associated with an increased admission NEWS in different commonly encountered diagnoses. Therefore, the mortality risk of some ‘low risk’ conditions can be dramatically increased if their admission NEWS is elevated, whereas some ‘high risk’ conditions are still likely to die even if their admission NEWS is low.
- Open Access
- 4 December 2024
- Published by BMJ in BMJ Health & Care Informatics
- Vol. 31 (1) , e101088
Performance of machine learning versus the national early warning score for predicting patient deterioration risk: a single-site study of emergency admissions
Objectives: Increasing operational pressures on emergency departments (ED) make it imperative to quickly and accurately identify patients requiring urgent clinical intervention. The widespread adoption of electronic health records (EHR) makes rich feature patient data sets more readily available. These large data stores lend themselves to use in modern machine learning (ML) models. This paper investigates the use of transformer-based models to identify critical deterioration in unplanned ED admissions, using free-text fields, such as triage notes, and tabular data, including early warning scores (EWS). Design: A retrospective ML study. Setting: A large ED in a UK university teaching hospital. Methods: We extracted rich feature sets of routine clinical data from the EHR and systematically measured the performance of tree- and transformer-based models for predicting patient mortality or admission to critical care within 24 hours of presentation to ED. We compared our proposed models to the National EWS (NEWS). Results: Models were trained on 174 393 admission records. We found that models including free-text triage notes outperform structured tabular data models, achieving an average precision of 0.92, compared with 0.75 for tree-based models and 0.12 for NEWS. Conclusions: Our findings suggests that machine learning models using free-text data have the potential to improve clinical decision-making in the ED; our techniques significantly reduce alert rate while detecting most high-risk patients missed by NEWS.
- 4 December 2024
- Published by BMJ in BMJ Health & Care Informatics
- Vol. 31 (1) , e101088
Open AccessPerformance of machine learning versus the national early warning score for predicting patient deterioration risk: a single-site study of emergency admissions
Objectives: Increasing operational pressures on emergency departments (ED) make it imperative to quickly and accurately identify patients requiring urgent clinical intervention. The widespread adoption of electronic health records (EHR) makes rich feature patient data sets more readily available. These large data stores lend themselves to use in modern machine learning (ML) models. This paper investigates the use of transformer-based models to identify critical deterioration in unplanned ED admissions, using free-text fields, such as triage notes, and tabular data, including early warning scores (EWS). Design: A retrospective ML study. Setting: A large ED in a UK university teaching hospital. Methods: We extracted rich feature sets of routine clinical data from the EHR and systematically measured the performance of tree- and transformer-based models for predicting patient mortality or admission to critical care within 24 hours of presentation to ED. We compared our proposed models to the National EWS (NEWS). Results: Models were trained on 174 393 admission records. We found that models including free-text triage notes outperform structured tabular data models, achieving an average precision of 0.92, compared with 0.75 for tree-based models and 0.12 for NEWS. Conclusions: Our findings suggests that machine learning models using free-text data have the potential to improve clinical decision-making in the ED; our techniques significantly reduce alert rate while detecting most high-risk patients missed by NEWS.
- Open Access
- 2 November 2024
- Published by BMJ in BMJ Oncology
- Vol. 3 (1) , e000430
From prediction to practice: mitigating bias and data shift in machine-learning models for chemotherapy-induced organ dysfunction across unseen cancers
Objectives: Routine monitoring of renal and hepatic function during chemotherapy ensures that treatment-related organ damage has not occurred and clearance of subsequent treatment is not hindered; however, frequency and timing are not optimal. Model bias and data heterogeneity concerns have hampered the ability of machine learning (ML) to be deployed into clinical practice. This study aims to develop models that could support individualised decisions on the timing of renal and hepatic monitoring while exploring the effect of data shift on model performance. Methods and analysis: We used retrospective data from three UK hospitals to develop and validate ML models predicting unacceptable rises in creatinine/bilirubin post cycle 3 for patients undergoing treatment for the following cancers: breast, colorectal, lung, ovarian and diffuse large B-cell lymphoma. Results: We extracted 3614 patients with no missing blood test data across cycles 1–6 of chemotherapy treatment. We improved on previous work by including predictions post cycle 3. Optimised for sensitivity, we achieve F2 scores of 0.7773 (bilirubin) and 0.6893 (creatinine) on unseen data. Performance is consistent on tumour types unseen during training (F2 bilirubin: 0.7423, F2 creatinine: 0.6820). Conclusion: Our technique highlights the effectiveness of ML in clinical settings, demonstrating the potential to improve the delivery of care. Notably, our ML models can generalise to unseen tumour types. We propose gold-standard bias mitigation steps for ML models: evaluation on multisite data, thorough patient population analysis, and both formalised bias measures and model performance comparisons on patient subgroups. We demonstrate that data aggregation techniques have unintended consequences on model bias.
- 2 November 2024
- Published by BMJ in BMJ Oncology
- Vol. 3 (1) , e000430
Open AccessFrom prediction to practice: mitigating bias and data shift in machine-learning models for chemotherapy-induced organ dysfunction across unseen cancers
Objectives: Routine monitoring of renal and hepatic function during chemotherapy ensures that treatment-related organ damage has not occurred and clearance of subsequent treatment is not hindered; however, frequency and timing are not optimal. Model bias and data heterogeneity concerns have hampered the ability of machine learning (ML) to be deployed into clinical practice. This study aims to develop models that could support individualised decisions on the timing of renal and hepatic monitoring while exploring the effect of data shift on model performance. Methods and analysis: We used retrospective data from three UK hospitals to develop and validate ML models predicting unacceptable rises in creatinine/bilirubin post cycle 3 for patients undergoing treatment for the following cancers: breast, colorectal, lung, ovarian and diffuse large B-cell lymphoma. Results: We extracted 3614 patients with no missing blood test data across cycles 1–6 of chemotherapy treatment. We improved on previous work by including predictions post cycle 3. Optimised for sensitivity, we achieve F2 scores of 0.7773 (bilirubin) and 0.6893 (creatinine) on unseen data. Performance is consistent on tumour types unseen during training (F2 bilirubin: 0.7423, F2 creatinine: 0.6820). Conclusion: Our technique highlights the effectiveness of ML in clinical settings, demonstrating the potential to improve the delivery of care. Notably, our ML models can generalise to unseen tumour types. We propose gold-standard bias mitigation steps for ML models: evaluation on multisite data, thorough patient population analysis, and both formalised bias measures and model performance comparisons on patient subgroups. We demonstrate that data aggregation techniques have unintended consequences on model bias.
- Open Access
- 24 January 2023
- Published by Springer Nature in Scientific Reports
- Vol. 13 (1) , 1
Author Correction: Using model explanations to guide deep learning models towards consistent explanations for EHR data
- 24 January 2023
- Published by Springer Nature in Scientific Reports
- Vol. 13 (1) , 1
Open AccessAuthor Correction: Using model explanations to guide deep learning models towards consistent explanations for EHR data
- Open Access
- 18 November 2022
- Published by Springer Nature in Scientific Reports
- Vol. 12 (1) , 1-14
Using model explanations to guide deep learning models towards consistent explanations for EHR data
It has been shown that identical deep learning (DL) architectures will produce distinct explanations when trained with different hyperparameters that are orthogonal to the task (e.g. random seed, training set order). In domains such as healthcare and finance, where transparency and explainability is paramount, this can be a significant barrier to DL adoption. In this study we present a further analysis of explanation (in)consistency on 6 tabular datasets/tasks, with a focus on Electronic Health Records data. We propose a novel deep learning ensemble architecture that trains its sub-models to produce consistent explanations, improving explanation consistency by as much as 315% (e.g. from 0.02433 to 0.1011 on MIMIC-IV), and on average by 124% (e.g. from 0.12282 to 0.4450 on the BCW dataset). We evaluate the effectiveness of our proposed technique and discuss the implications our results have for both industrial applications of DL and explainability as well as future methodological work.
- 18 November 2022
- Published by Springer Nature in Scientific Reports
- Vol. 12 (1) , 1-14
Open AccessUsing model explanations to guide deep learning models towards consistent explanations for EHR data
It has been shown that identical deep learning (DL) architectures will produce distinct explanations when trained with different hyperparameters that are orthogonal to the task (e.g. random seed, training set order). In domains such as healthcare and finance, where transparency and explainability is paramount, this can be a significant barrier to DL adoption. In this study we present a further analysis of explanation (in)consistency on 6 tabular datasets/tasks, with a focus on Electronic Health Records data. We propose a novel deep learning ensemble architecture that trains its sub-models to produce consistent explanations, improving explanation consistency by as much as 315% (e.g. from 0.02433 to 0.1011 on MIMIC-IV), and on average by 124% (e.g. from 0.12282 to 0.4450 on the BCW dataset). We evaluate the effectiveness of our proposed technique and discuss the implications our results have for both industrial applications of DL and explainability as well as future methodological work.
- Open Access
Agree to Disagree: When Deep Learning Models With Identical Architectures Produce Distinct Explanations
Deep Learning of neural networks has progressively become more prominent in healthcare with models reaching, or even surpassing, expert accuracy levels. However, these success stories are tainted by concerning reports on the lack of model transparency and bias against some medical conditions or patients' sub-groups. Explainable methods are considered the gateway to alleviate many of these concerns. In this study we demonstrate that the generated explanations are volatile to changes in model training that are perpendicular to the classification task and model structure. This raises further questions about trust in deep learning models for healthcare. Mainly, whether the models capture underlying causal links in the data or just rely on spurious correlations that are made visible via explanation methods. We demonstrate that the output of explainability methods on deep neural networks can vary significantly by changes of hyper-parameters, such as the random seed or how the training set is shuffled. We introduce a measure of explanation consistency which we use to highlight the identified problems on the MIMIC-CXR dataset. We find explanations of identical models but with different training setups have a low consistency: $\approx$ 33% on average. On the contrary, kernel methods are robust against any orthogonal changes, with explanation consistency at 94%. We conclude that current trends in model explanation are not sufficient to mitigate the risks of deploying models in real life healthcare applications.Open AccessAgree to Disagree: When Deep Learning Models With Identical Architectures Produce Distinct Explanations
Deep Learning of neural networks has progressively become more prominent in healthcare with models reaching, or even surpassing, expert accuracy levels. However, these success stories are tainted by concerning reports on the lack of model transparency and bias against some medical conditions or patients' sub-groups. Explainable methods are considered the gateway to alleviate many of these concerns. In this study we demonstrate that the generated explanations are volatile to changes in model training that are perpendicular to the classification task and model structure. This raises further questions about trust in deep learning models for healthcare. Mainly, whether the models capture underlying causal links in the data or just rely on spurious correlations that are made visible via explanation methods. We demonstrate that the output of explainability methods on deep neural networks can vary significantly by changes of hyper-parameters, such as the random seed or how the training set is shuffled. We introduce a measure of explanation consistency which we use to highlight the identified problems on the MIMIC-CXR dataset. We find explanations of identical models but with different training setups have a low consistency: $\approx$ 33% on average. On the contrary, kernel methods are robust against any orthogonal changes, with explanation consistency at 94%. We conclude that current trends in model explanation are not sufficient to mitigate the risks of deploying models in real life healthcare applications.