Munib Mesinovic
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Munib Mesinovic has a total of 45 co-authors. Here are the top co-authors by number of shared publications:
- Tingting Zhu
University of Oxford, UK
6 shared publications Peter Watkinson
Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, UK
4 shared publicationsTingting Zhu
Department of Engineering Science, University of Oxford, Oxford, UK
4 shared publications- Bich Hai Ho
Oxford University Clinical Research Unit, Ho Chi Minh City, Vietnam
2 shared publications - Catherine Louise Thwaites
Oxford University Clinical Research Unit, Ho Chi Minh City, Viet Nam
2 shared publications - Hong Phuc Vo
Ho Chi Minh City Oxford University Clinical Research Unit
2 shared publications
Latest Publications
- Open Access
- 9 October 2026
- Published by MDPI AG in Preprints.org
Rethinking the Evaluation of Generative AI in Clinical Diagnosis and Prognosis
Generative AI increasingly supports clinical diagnosis and prognosis, but safe deployment into clinical practice requires standardized evaluation. Evaluation practices from 2022 to 2025 are surveyed, and a five-dimensional framework is introduced: (I) Factual and Contextual Correctness, (II) Safety, Alignment, and Robustness, (III) Reasoning Quality, (IV) Calibration and Uncertainty, and (V) Clinical Anthropomorphism. Current automated and human evaluation methods rely heavily on accuracy, lack multimodal assessment, and are not aligned with human judgment. Emerging benchmarks are assessed and linked to our framework. A hybrid evaluation approach combining automated scalability with expert clinical judgment is recommended for safe and effective deployment.
- 9 October 2026
- Published by MDPI AG in Preprints.org
Open AccessRethinking the Evaluation of Generative AI in Clinical Diagnosis and Prognosis
Generative AI increasingly supports clinical diagnosis and prognosis, but safe deployment into clinical practice requires standardized evaluation. Evaluation practices from 2022 to 2025 are surveyed, and a five-dimensional framework is introduced: (I) Factual and Contextual Correctness, (II) Safety, Alignment, and Robustness, (III) Reasoning Quality, (IV) Calibration and Uncertainty, and (V) Clinical Anthropomorphism. Current automated and human evaluation methods rely heavily on accuracy, lack multimodal assessment, and are not aligned with human judgment. Emerging benchmarks are assessed and linked to our framework. A hybrid evaluation approach combining automated scalability with expert clinical judgment is recommended for safe and effective deployment.
- 7 August 2026
- Published by Springer Nature in Nature Biomedical Engineering
- Vol. 10 (8) , 1536-1556
Causal graph neural networks for healthcare
- 7 August 2026
- Published by Springer Nature in Nature Biomedical Engineering
- Vol. 10 (8) , 1536-1556
Causal graph neural networks for healthcare
- Open Access
- 26 May 2026
- Published by Association for Computing Machinery (ACM) in ACM Computing Surveys
- Vol. 58 (15) , 1-36
Clinical XAI: A Tutorial on Concepts, Methods, and Modalities from Linear Models to LLMs
Explainable AI (XAI) in clinical risk prediction helps make applied machine learning models transparent and clinically accountable. This tutorial covers recent progress across EHR, text, imaging, genomics, and multimodal data, with a later focus on emerging large language model (LLM) systems. We aimed to standardise terminology and distinguish terms such as interpretability, explanation, fairness, bias, trust, and transparency. We also provide a decision map linking methods to evaluation requirements. We review 97 unique studies (2019–2024) using a scheme that distinguishes alignment with knowledge (AK), expert review (ER), decision impact (DI), workflow integration (WI), and patient-facing (PF) assessment. We find that SHAP with tree ensembles is the most popular method in tabular EHRs, attention and CAM methods are common in imaging, and saliency-plus enrichment testing in genomics. A minority of papers report clinician-involved validation, and reporting of rater agreement, calibration, and subgroup effects is inconsistent. We, thus, propose a tiered, modality-agnostic framework, comprising AK+ER as the minimum standard, followed by DI (accuracy, time, confidence), and then WI (usage, outcomes), and provide design checklists, metrics, and pitfalls. Our aim is a roadmap, grounded in clinical practice, for developing, stress-testing, and deploying XAI in risk prediction.
- 26 May 2026
- Published by Association for Computing Machinery (ACM) in ACM Computing Surveys
- Vol. 58 (15) , 1-36
Open AccessClinical XAI: A Tutorial on Concepts, Methods, and Modalities from Linear Models to LLMs
Explainable AI (XAI) in clinical risk prediction helps make applied machine learning models transparent and clinically accountable. This tutorial covers recent progress across EHR, text, imaging, genomics, and multimodal data, with a later focus on emerging large language model (LLM) systems. We aimed to standardise terminology and distinguish terms such as interpretability, explanation, fairness, bias, trust, and transparency. We also provide a decision map linking methods to evaluation requirements. We review 97 unique studies (2019–2024) using a scheme that distinguishes alignment with knowledge (AK), expert review (ER), decision impact (DI), workflow integration (WI), and patient-facing (PF) assessment. We find that SHAP with tree ensembles is the most popular method in tabular EHRs, attention and CAM methods are common in imaging, and saliency-plus enrichment testing in genomics. A minority of papers report clinician-involved validation, and reporting of rater agreement, calibration, and subgroup effects is inconsistent. We, thus, propose a tiered, modality-agnostic framework, comprising AK+ER as the minimum standard, followed by DI (accuracy, time, confidence), and then WI (usage, outcomes), and provide design checklists, metrics, and pitfalls. Our aim is a roadmap, grounded in clinical practice, for developing, stress-testing, and deploying XAI in risk prediction.
- RetractedOpen Access
- 16 January 2026
- Published by Springer Nature in npj Digital Medicine
- Vol. 9 (1)
RETRACTED ARTICLE: DynaGraph: interpretable dynamic graph learning for temporal electronic health records
Electronic health records (EHRs) capture evolving physiological processes, yet most machine learning models impose static or sequential assumptions that flatten their temporal and relational complexity. We introduce DynaGraph, a dynamic and interpretable graph learning framework that constructs evolving spatio-temporal graphs from multivariate clinical time-series. Unlike previous methods, DynaGraph learns the structure of relationships between different clinical variables over time without predefined graphs, integrates sequential embeddings with contrastive graph augmentation, and incorporates a pseudo-attention mechanism to reveal temporally resolved risk factors. Trained end-to-end with a novel multi-loss objective that combines focal, structural, and contrastive components, DynaGraph addresses two pervasive challenges in real-world clinical modelling: class imbalance and temporal instability. We evaluated DynaGraph on four large-scale EHR datasets totalling 40,856 patients: MIMIC-III (17,279 ICU admissions), eICU (1433 cardiac ICU patients), HiRID-ICU (33,000 patients), and EHRSHOT (2378 primary care patients). DynaGraph consistently outperforms 14 state-of-the-art baselines, achieving 6-8% relative improvements in area under the precision-recall curve (AUPRC) and significant gains in sensitivity (12-22% over leading methods). Beyond predictive performance, DynaGraph offers time-specific interpretability aligned with clinical reasoning, providing gradient-based feature importance scores at 3-hour intervals that identify which physiological relationships drive predictions. This framework explicitly models temporal attribution of risk factors across patient trajectories in a millisecond inference time.
- 16 January 2026
- Published by Springer Nature in npj Digital Medicine
- Vol. 9 (1)
RetractedOpen AccessRETRACTED ARTICLE: DynaGraph: interpretable dynamic graph learning for temporal electronic health records
Electronic health records (EHRs) capture evolving physiological processes, yet most machine learning models impose static or sequential assumptions that flatten their temporal and relational complexity. We introduce DynaGraph, a dynamic and interpretable graph learning framework that constructs evolving spatio-temporal graphs from multivariate clinical time-series. Unlike previous methods, DynaGraph learns the structure of relationships between different clinical variables over time without predefined graphs, integrates sequential embeddings with contrastive graph augmentation, and incorporates a pseudo-attention mechanism to reveal temporally resolved risk factors. Trained end-to-end with a novel multi-loss objective that combines focal, structural, and contrastive components, DynaGraph addresses two pervasive challenges in real-world clinical modelling: class imbalance and temporal instability. We evaluated DynaGraph on four large-scale EHR datasets totalling 40,856 patients: MIMIC-III (17,279 ICU admissions), eICU (1433 cardiac ICU patients), HiRID-ICU (33,000 patients), and EHRSHOT (2378 primary care patients). DynaGraph consistently outperforms 14 state-of-the-art baselines, achieving 6-8% relative improvements in area under the precision-recall curve (AUPRC) and significant gains in sensitivity (12-22% over leading methods). Beyond predictive performance, DynaGraph offers time-specific interpretability aligned with clinical reasoning, providing gradient-based feature importance scores at 3-hour intervals that identify which physiological relationships drive predictions. This framework explicitly models temporal attribution of risk factors across patient trajectories in a millisecond inference time.
- Open Access
- 16 January 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
Foundation models enable wearable signal screening for cardiovascular disease among people living with HIV
Background: Cardiovascular disease screening faces significant challenges in resource-limited settings, where infrastructure and computational constraints preclude advanced assessment. These constraints are particularly acute for people living with human immunodeficiency virus (HIV), who experience elevated cardiovascular risk yet often receive care in clinics without specialist diagnostic capacity. Pretrained physiological foundation models offer potential for low-cost screening using wearable sensors, though their applicability in resource-constrained settings remains unclear. Methods: We evaluate pretrained physiological embeddings from foundation models for cardiovascular disease detection using photoplethysmography signals from 80 people living with HIV in Ho Chi Minh City, Vietnam. Of 80 participants, 13 (16%) had cardiologist-confirmed cardiovascular disease. We compare strictly zero-shot deployment (NormWear without local training) with frozen PaPaGei embeddings plus locally trained classifier, alongside traditional approaches. Results: Here we show that the PaPaGei-embedding approach achieves area under the receiver operating characteristic curve 0.769 (95% confidence interval: 0.70, 0.84) and average precision 0.489 (0.37, 0.61) in this pilot cohort, numerically higher than zero-shot NormWear (0.610; 0.226), principal component analysis features (0.651; 0.208), and supervised clinical models (0.744; 0.433). This approach requires local labels for classifier training but avoids computationally intensive foundation model fine-tuning. However, given the small positive class size (13 cases), these findings require validation in larger cohorts. PaPaGei embeddings capture clinically coherent structure: patients on dolutegravir-based regimens cluster in low-risk regions, while those with high cholesterol variability occupy high-risk areas. Conclusions: These preliminary findings provide a potential methodological framework for deploying foundation models in resource-constrained settings, though adequately powered, multi-centre validation is essential before clinical implementation.
- 16 January 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
Open AccessFoundation models enable wearable signal screening for cardiovascular disease among people living with HIV
Background: Cardiovascular disease screening faces significant challenges in resource-limited settings, where infrastructure and computational constraints preclude advanced assessment. These constraints are particularly acute for people living with human immunodeficiency virus (HIV), who experience elevated cardiovascular risk yet often receive care in clinics without specialist diagnostic capacity. Pretrained physiological foundation models offer potential for low-cost screening using wearable sensors, though their applicability in resource-constrained settings remains unclear. Methods: We evaluate pretrained physiological embeddings from foundation models for cardiovascular disease detection using photoplethysmography signals from 80 people living with HIV in Ho Chi Minh City, Vietnam. Of 80 participants, 13 (16%) had cardiologist-confirmed cardiovascular disease. We compare strictly zero-shot deployment (NormWear without local training) with frozen PaPaGei embeddings plus locally trained classifier, alongside traditional approaches. Results: Here we show that the PaPaGei-embedding approach achieves area under the receiver operating characteristic curve 0.769 (95% confidence interval: 0.70, 0.84) and average precision 0.489 (0.37, 0.61) in this pilot cohort, numerically higher than zero-shot NormWear (0.610; 0.226), principal component analysis features (0.651; 0.208), and supervised clinical models (0.744; 0.433). This approach requires local labels for classifier training but avoids computationally intensive foundation model fine-tuning. However, given the small positive class size (13 cases), these findings require validation in larger cohorts. PaPaGei embeddings capture clinically coherent structure: patients on dolutegravir-based regimens cluster in low-risk regions, while those with high cholesterol variability occupy high-risk areas. Conclusions: These preliminary findings provide a potential methodological framework for deploying foundation models in resource-constrained settings, though adequately powered, multi-centre validation is essential before clinical implementation.
- Open Access
Foundation model embeddings enable cardiovascular screening for people living with HIV in Vietnam using wearable signals
Cardiovascular disease (CVD) screening faces significant challenges in resource-limited settings, where infrastructure and computational constraints preclude the use of advanced remote assessment. These constraints are particularly acute for people living with HIV (PLWH), who experience elevated CVD risk yet often receive care in clinics without the capacity for specialist diagnostics. We evaluate pretrained physiological embeddings from foundation models for CVD detection using low-cost wearable photoplethysmography (PPG) signals from 80 PLWH outpatients in Ho Chi Minh City, Vietnam. We compare a strictly zero-shot approach (NormWear applied without any local training) with a more practical pipeline that uses frozen PaPaGei embeddings plus a locally trained classifier. The PaPaGei-embedding approach achieved superior discrimination (AUROC 0.769) compared with zero-shot NormWear (0.610), traditional PCA features (0.651), and established clinical scores, including the Framingham score (0.551) and the D:A:D-modified Framingham score (0.462). Without fine-tuning the foundation model itself, PaPaGei embeddings captured clinically coherent structure: patients on dolutegravir-based regimens clustered in low-risk regions, while those with high cholesterol variability occupied high-risk areas, consistent with cardiometabolic pathophysiology. These results show that pretrained physiological embeddings can enable accurate screening when combined with lightweight local calibration, reducing reliance on extensive feature engineering while preserving pathophysiological plausibility and actionable triage behaviour in the learned embeddings. This provides a practical approach for deploying foundation models in resource-constrained settings where training deep models may be infeasible.
Open AccessFoundation model embeddings enable cardiovascular screening for people living with HIV in Vietnam using wearable signals
Cardiovascular disease (CVD) screening faces significant challenges in resource-limited settings, where infrastructure and computational constraints preclude the use of advanced remote assessment. These constraints are particularly acute for people living with HIV (PLWH), who experience elevated CVD risk yet often receive care in clinics without the capacity for specialist diagnostics. We evaluate pretrained physiological embeddings from foundation models for CVD detection using low-cost wearable photoplethysmography (PPG) signals from 80 PLWH outpatients in Ho Chi Minh City, Vietnam. We compare a strictly zero-shot approach (NormWear applied without any local training) with a more practical pipeline that uses frozen PaPaGei embeddings plus a locally trained classifier. The PaPaGei-embedding approach achieved superior discrimination (AUROC 0.769) compared with zero-shot NormWear (0.610), traditional PCA features (0.651), and established clinical scores, including the Framingham score (0.551) and the D:A:D-modified Framingham score (0.462). Without fine-tuning the foundation model itself, PaPaGei embeddings captured clinically coherent structure: patients on dolutegravir-based regimens clustered in low-risk regions, while those with high cholesterol variability occupied high-risk areas, consistent with cardiometabolic pathophysiology. These results show that pretrained physiological embeddings can enable accurate screening when combined with lightweight local calibration, reducing reliance on extensive feature engineering while preserving pathophysiological plausibility and actionable triage behaviour in the learned embeddings. This provides a practical approach for deploying foundation models in resource-constrained settings where training deep models may be infeasible.
- Open Access
- 31 July 2025
- Published by Springer Nature in Scientific Reports
- Vol. 15 (1) , 1-15
Explainable machine learning for predicting ICU mortality in myocardial infarction patients using pseudo-dynamic data
Myocardial infarction (MI) remains one of the greatest contributors to mortality, and patients admitted to the intensive care unit (ICU) with myocardial infarction are at higher risk of death. In this study, we use two retrospective cohorts extracted from two US-based ICU databases, eICU and MIMIC-IV, to develop an explainable pseudo-dynamic machine learning framework for mortality prediction in the ICU. The method provides accurate prediction for ICU patients up to 24 hours before the event and provides time-resolved interpretability. We compare standard supervised machine learning algorithms with novel tabular deep learning approaches and find that an integrated XGBoost model in our EHR time-series extraction framework (XMI-ICU) performs best. The framework was evaluated on a held-out test set from eICU and externally validated on the MIMIC-IV cohort using the most important features identified by time-resolved Shapley values. XMI-ICU achieved AUROCs of 92.0 (balanced accuracy of 82.3) for a 6-hour prediction of mortality. We demonstrate that XMI-ICU maintains reliable predictive performance across different prediction horizons (6, 12, 18, and 24 hours) during ICU stay while also achieving successful external validation in a separate patient cohort from MIMIC-IV without any previous training on that dataset. We also evaluated the framework for clinical risk analysis by comparing it to the standard APACHE IV system in active use. We show that our framework successfully leverages time-series physiological measurements from ICU health records by translating them into stacked static prediction problems for mortality in heart attack patients and can offer clinical insight from time-resolved interpretability through the use of Shapley values.
- 31 July 2025
- Published by Springer Nature in Scientific Reports
- Vol. 15 (1) , 1-15
Open AccessExplainable machine learning for predicting ICU mortality in myocardial infarction patients using pseudo-dynamic data
Myocardial infarction (MI) remains one of the greatest contributors to mortality, and patients admitted to the intensive care unit (ICU) with myocardial infarction are at higher risk of death. In this study, we use two retrospective cohorts extracted from two US-based ICU databases, eICU and MIMIC-IV, to develop an explainable pseudo-dynamic machine learning framework for mortality prediction in the ICU. The method provides accurate prediction for ICU patients up to 24 hours before the event and provides time-resolved interpretability. We compare standard supervised machine learning algorithms with novel tabular deep learning approaches and find that an integrated XGBoost model in our EHR time-series extraction framework (XMI-ICU) performs best. The framework was evaluated on a held-out test set from eICU and externally validated on the MIMIC-IV cohort using the most important features identified by time-resolved Shapley values. XMI-ICU achieved AUROCs of 92.0 (balanced accuracy of 82.3) for a 6-hour prediction of mortality. We demonstrate that XMI-ICU maintains reliable predictive performance across different prediction horizons (6, 12, 18, and 24 hours) during ICU stay while also achieving successful external validation in a separate patient cohort from MIMIC-IV without any previous training on that dataset. We also evaluated the framework for clinical risk analysis by comparing it to the standard APACHE IV system in active use. We show that our framework successfully leverages time-series physiological measurements from ICU health records by translating them into stacked static prediction problems for mortality in heart attack patients and can offer clinical insight from time-resolved interpretability through the use of Shapley values.
- Top Cited PapersOpen Access
- 17 July 2025
- Published by Springer Nature in Communications Engineering
- Vol. 4 (1) , 128
Explainability in the age of large language models for healthcare
Large language models show remarkable potential in healthcare but face critical explainability challenges that must be addressed before widespread clinical deployment. Here, we examine technical and regulatory solutions needed to develop trustworthy, transparent large language models for responsible healthcare integration.
- 17 July 2025
- Published by Springer Nature in Communications Engineering
- Vol. 4 (1) , 128
Top Cited PapersOpen AccessExplainability in the age of large language models for healthcare
Large language models show remarkable potential in healthcare but face critical explainability challenges that must be addressed before widespread clinical deployment. Here, we examine technical and regulatory solutions needed to develop trustworthy, transparent large language models for responsible healthcare integration.
- Open Access
- 22 June 2025
- Published by JMIR Publications Inc. in JMIR Preprints
Wearable devices for screening of cardiovascular disease in people living with HIV in resource-limited settings: a pilot study (Preprint)
BACKGROUND People living with HIV (PLWH) are at increased risk of cardiovascular disease but in low- and middle-income countries, screening and diagnosing cardiovascular disease requires expertise and equipment unavailable in most HIV treatment centres. OBJECTIVE We aim to evaluate an alternative non-invasive approach using low-cost wearable pulse oximeters and machine learning in asymptomatic PLWH attending routine outpatient clinics in Vietnam. METHODS Eighty asymptomatic PWLH without prior history of cardiovascular disease were screened for cardiovascular disease risk using Framingham and D:A:D modified Framingham scores. Cardiovascular disease was screened for from ECG and echocardiograms performed in all participants at baseline, with ECG repeated at 2-3 follow-up visits over a 12 month period. Continuous pulse oximetry waveforms were collected at baseline using low-cost wearable oximeters from which heart rate variability waveform morphology features were extracted from a 5 minute segment. These features and variables from risk scores were used to develop machine learning models predicting cardiovascular risk and cardiovascular disease. Model features were ranked to deliver further insight. Participants perspectives were collected by questionnaire at a follow up visit. RESULTS Acceptable quality data was available from all participants. Overall 52 (65%) of participants were classified as low risk using the standard Framingham score, whereas 36 (45%) were classified as low risk according to D:A:D modified score. Baseline findings suggestive of cardiovascular disease were seen in 4 (11%) participants in the low risk group and 4 (11%) in the higher risk group. Elastic Net machine learning model achieved the best performance predicting Framingham or D:A:D modified risk groups, with an accuracy of 91% for risk groups alone and 85% and 86% respectively for the composite endpoint of risk or evidence of cardiovascular disease. Feature ranking particularly identified pulse waveform morphology features as important predictors of risk group, and whereas heart rate variability features were included in composite endpoint prediction. Overall participants expressed positive sentiments about wearable devices. CONCLUSIONS Low-cost pulse oximeters linked to machine learning algorithms are a feasible non-invasive approach for diagnosing and screening cardiovascular disease in PLWH in a low resource setting.
- 22 June 2025
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessWearable devices for screening of cardiovascular disease in people living with HIV in resource-limited settings: a pilot study (Preprint)
BACKGROUND People living with HIV (PLWH) are at increased risk of cardiovascular disease but in low- and middle-income countries, screening and diagnosing cardiovascular disease requires expertise and equipment unavailable in most HIV treatment centres. OBJECTIVE We aim to evaluate an alternative non-invasive approach using low-cost wearable pulse oximeters and machine learning in asymptomatic PLWH attending routine outpatient clinics in Vietnam. METHODS Eighty asymptomatic PWLH without prior history of cardiovascular disease were screened for cardiovascular disease risk using Framingham and D:A:D modified Framingham scores. Cardiovascular disease was screened for from ECG and echocardiograms performed in all participants at baseline, with ECG repeated at 2-3 follow-up visits over a 12 month period. Continuous pulse oximetry waveforms were collected at baseline using low-cost wearable oximeters from which heart rate variability waveform morphology features were extracted from a 5 minute segment. These features and variables from risk scores were used to develop machine learning models predicting cardiovascular risk and cardiovascular disease. Model features were ranked to deliver further insight. Participants perspectives were collected by questionnaire at a follow up visit. RESULTS Acceptable quality data was available from all participants. Overall 52 (65%) of participants were classified as low risk using the standard Framingham score, whereas 36 (45%) were classified as low risk according to D:A:D modified score. Baseline findings suggestive of cardiovascular disease were seen in 4 (11%) participants in the low risk group and 4 (11%) in the higher risk group. Elastic Net machine learning model achieved the best performance predicting Framingham or D:A:D modified risk groups, with an accuracy of 91% for risk groups alone and 85% and 86% respectively for the composite endpoint of risk or evidence of cardiovascular disease. Feature ranking particularly identified pulse waveform morphology features as important predictors of risk group, and whereas heart rate variability features were included in composite endpoint prediction. Overall participants expressed positive sentiments about wearable devices. CONCLUSIONS Low-cost pulse oximeters linked to machine learning algorithms are a feasible non-invasive approach for diagnosing and screening cardiovascular disease in PLWH in a low resource setting.
- Open Access
- 21 November 2024
- Published by Oxford University Press (OUP) in Journal of the American Medical Informatics Association
- Vol. 33 (1) , 112-122
DySurv: dynamic deep learning model for survival analysis with conditional variational inference
Objective: Machine learning applications for longitudinal electronic health records often forecast the risk of events at fixed time points, whereas survival analysis achieves dynamic risk prediction by estimating time-to-event distributions. Here, we propose a novel conditional variational autoencoder-based method, DySurv, which uses a combination of static and longitudinal measurements from electronic health records to estimate the individual risk of death dynamically. Materials and Methods: DySurv directly estimates the cumulative risk incidence function without making any parametric assumptions on the underlying stochastic process of the time-to-event. We evaluate DySurv on 6 time-to-event benchmark datasets in healthcare, as well as 2 real-world intensive care unit (ICU) electronic health records (EHR) datasets extracted from the eICU Collaborative Research (eICU) and the Medical Information Mart for Intensive Care database (MIMIC-IV). Results: DySurv outperforms other existing statistical and deep learning approaches to time-to-event analysis across concordance and other metrics. It achieves time-dependent concordance of over 60% in the eICU case. It is also over 12% more accurate and 22% more sensitive than in-use ICU scores like Acute Physiology and Chronic Health Evaluation (APACHE) and Sequential Organ Failure Assessment (SOFA) scores. The predictive capacity of DySurv is consistent and the survival estimates remain disentangled across different datasets. Discussion: Our interdisciplinary framework successfully incorporates deep learning, survival analysis, and intensive care to create a novel method for time-to-event prediction from longitudinal health records. We test our method on several held-out test sets from a variety of healthcare datasets and compare it to existing in-use clinical risk scoring benchmarks. Conclusion: While our method leverages non-parametric extensions to deep learning-guided estimations of the survival distribution, further deep learning paradigms could be explored.
- 21 November 2024
- Published by Oxford University Press (OUP) in Journal of the American Medical Informatics Association
- Vol. 33 (1) , 112-122
Open AccessDySurv: dynamic deep learning model for survival analysis with conditional variational inference
Objective: Machine learning applications for longitudinal electronic health records often forecast the risk of events at fixed time points, whereas survival analysis achieves dynamic risk prediction by estimating time-to-event distributions. Here, we propose a novel conditional variational autoencoder-based method, DySurv, which uses a combination of static and longitudinal measurements from electronic health records to estimate the individual risk of death dynamically. Materials and Methods: DySurv directly estimates the cumulative risk incidence function without making any parametric assumptions on the underlying stochastic process of the time-to-event. We evaluate DySurv on 6 time-to-event benchmark datasets in healthcare, as well as 2 real-world intensive care unit (ICU) electronic health records (EHR) datasets extracted from the eICU Collaborative Research (eICU) and the Medical Information Mart for Intensive Care database (MIMIC-IV). Results: DySurv outperforms other existing statistical and deep learning approaches to time-to-event analysis across concordance and other metrics. It achieves time-dependent concordance of over 60% in the eICU case. It is also over 12% more accurate and 22% more sensitive than in-use ICU scores like Acute Physiology and Chronic Health Evaluation (APACHE) and Sequential Organ Failure Assessment (SOFA) scores. The predictive capacity of DySurv is consistent and the survival estimates remain disentangled across different datasets. Discussion: Our interdisciplinary framework successfully incorporates deep learning, survival analysis, and intensive care to create a novel method for time-to-event prediction from longitudinal health records. We test our method on several held-out test sets from a variety of healthcare datasets and compare it to existing in-use clinical risk scoring benchmarks. Conclusion: While our method leverages non-parametric extensions to deep learning-guided estimations of the survival distribution, further deep learning paradigms could be explored.