Fernando García-García
Research Subjects
Fernando García-García has a total of 42 co-authors. Here are the top co-authors by number of shared publications:
- Dae-Jin Lee
School of Science and Technology IE University Madrid Spain
5 shared publications - Maria Elena Hernando
Biomedical Engineering and Telemedicine Centre, Center for Biomedical Technology, Universidad Politécnica de Madrid, Spain
3 shared publications - Susana Garcia-Gutierrez
Osakidetza Basque Health Service, Usansolo University Hospital, Research Unit, GaldakaoGaldakao, Spain
3 shared publications Antoni Torres
Pneumology Department, Hospital Clinic of Barcelona, Barcelona, Spain
2 shared publications- Francisco J. Mendoza-Garcés
Galdakao-Usansolo University Hospital, Anaesthesia & Resuscitation Service - Galdakao (Basque Country), Spain
2 shared publications Isabel Urrutia Landa
Respiratory Service at Galdakao-Usansolo University Hospital, Galdakao, Spain
2 shared publications
Latest Publications
- Open Access
- 26 September 2026
- Published by Wiley in Expert Systems
- Vol. 43 (11) , e70440
A Semi‐Interpretable Multimodal Model Combining Electronic Health Records and Chest X‐Rays for COVID ‐19‐Related Death Prediction
Purpose: Prognostic models can benefit from integrating heterogeneous data sources. While deep learning‐based multimodal models often improve predictive performance, their limited interpretability remains a barrier to clinical adoption. Moreover, achieving robust generalization with such models typically requires large and diverse datasets, which are not always available in clinical research. This study evaluated whether a late fusion model combining traditional statistical approaches with deep learning models could rival deep learning‐only multimodal models while preserving interpretability to support clinical use. Methods: We conducted a retrospective single‐center cohort study of 965 hospitalized patients with laboratory‐confirmed SARS‐CoV‐2 infection who underwent frontal chest X‐ray imaging during hospitalization between March 2020 and January 2022. Of these, 64 (6.6%) died. Structured clinical data and chest X‐ray images were used to predict COVID‐19‐related death. We developed unimodal baseline models and two multimodal approaches: a deep learning‐based joint fusion model and a late fusion model leveraging logistic regression to combine predictions from the best‐performing models for every data modality. Given the limited sample size and class imbalance, performance was evaluated using repeated five‐fold cross validation. Results: The late fusion model achieved the best overall performance, with the highest median test AUPRC of 0.363 (IQR 0.271–0.384) across cross validation folds, outperforming unimodal baselines based on structured clinical data or images alone. The late fusion model also outperformed the joint fusion model, achieving, respectively, a 3.1% and 3.6% relative improvement in AUPRC and Brier score on pooled out‐of‐sample predictions and yielding the highest balanced F1‐score. Moreover, it preserved partial interpretability by linking predictions to clinically meaningful variables such as age and comorbidities and enabling image‐based predictions to be traced back to radiographic abnormalities. Conclusions: Late fusion models integrating statistical models for clinical data with deep learning for image classification can enhance prognostic performance over unimodal models without fully sacrificing interpretability. In settings with limited data, they may also outperform deep learning‐based joint fusion models. These findings support late fusion as a promising strategy that warrants prospective evaluation and external validation before clinical implementation.
- 26 September 2026
- Published by Wiley in Expert Systems
- Vol. 43 (11) , e70440
Open AccessA Semi‐Interpretable Multimodal Model Combining Electronic Health Records and Chest X‐Rays for COVID ‐19‐Related Death Prediction
Purpose: Prognostic models can benefit from integrating heterogeneous data sources. While deep learning‐based multimodal models often improve predictive performance, their limited interpretability remains a barrier to clinical adoption. Moreover, achieving robust generalization with such models typically requires large and diverse datasets, which are not always available in clinical research. This study evaluated whether a late fusion model combining traditional statistical approaches with deep learning models could rival deep learning‐only multimodal models while preserving interpretability to support clinical use. Methods: We conducted a retrospective single‐center cohort study of 965 hospitalized patients with laboratory‐confirmed SARS‐CoV‐2 infection who underwent frontal chest X‐ray imaging during hospitalization between March 2020 and January 2022. Of these, 64 (6.6%) died. Structured clinical data and chest X‐ray images were used to predict COVID‐19‐related death. We developed unimodal baseline models and two multimodal approaches: a deep learning‐based joint fusion model and a late fusion model leveraging logistic regression to combine predictions from the best‐performing models for every data modality. Given the limited sample size and class imbalance, performance was evaluated using repeated five‐fold cross validation. Results: The late fusion model achieved the best overall performance, with the highest median test AUPRC of 0.363 (IQR 0.271–0.384) across cross validation folds, outperforming unimodal baselines based on structured clinical data or images alone. The late fusion model also outperformed the joint fusion model, achieving, respectively, a 3.1% and 3.6% relative improvement in AUPRC and Brier score on pooled out‐of‐sample predictions and yielding the highest balanced F1‐score. Moreover, it preserved partial interpretability by linking predictions to clinically meaningful variables such as age and comorbidities and enabling image‐based predictions to be traced back to radiographic abnormalities. Conclusions: Late fusion models integrating statistical models for clinical data with deep learning for image classification can enhance prognostic performance over unimodal models without fully sacrificing interpretability. In settings with limited data, they may also outperform deep learning‐based joint fusion models. These findings support late fusion as a promising strategy that warrants prospective evaluation and external validation before clinical implementation.
- Open Access
- 8 September 2026
- Published by Elsevier in SSRN Electronic Journal
Losses based on ordinal binary decomposition for deep ordinal classification and segmentation
Ordinal classification (OC) is a supervised classification task where the ground truth has a natural class ordering. Ordinal segmentation (OS) is semantic segmentation, also with ordered labels. Traditional approaches to OC include naı̈ve nominal classification and regression, whereas recent works exploit unimodal regularisation (UR), but these force quantitative distances among the classes –which may not hold–, alongside regularisation strength ηR as a key extra hyperparameter, which may impact performance critically. In this work, two novel differentiable losses for OC and OS are presented, which leverage the principle of ordinal binary decompositions (OBD) with no need for per-subtask models. These losses have a nominal term, plus four OBD-related terms pertaining to the four types of ordinal decomposition that exist. We propose two weighting strategies for those OBD loss terms: either based on the frequency of data, or via differential entropy – which accounts for class imbalance. Unlike in UR with ηR, these avoid the need for any loss-related hyperparameter. To illustrative our proposal, four real-world datasets are tackled (three OC, one OS). Eight state-of-the-art baseline losses are used for comparison. Experimental results –regarding the behaviour of the DL models trained with each loss– varied notably across datasets and across performance metrics, indicating trade-offs. Whereas ’naı̈ve’ cross-entropy ranked the best overall, it was sub-par in error distance-based metrics. An UR and an OBD loss functions were the subsequent best, and safe in the sense of no sub-par results in any facet. In all, OBD proved to be a convenient option for OC/OS, without inductive biasesimplicit in its formulation (unlike UR).
- 8 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessLosses based on ordinal binary decomposition for deep ordinal classification and segmentation
Ordinal classification (OC) is a supervised classification task where the ground truth has a natural class ordering. Ordinal segmentation (OS) is semantic segmentation, also with ordered labels. Traditional approaches to OC include naı̈ve nominal classification and regression, whereas recent works exploit unimodal regularisation (UR), but these force quantitative distances among the classes –which may not hold–, alongside regularisation strength ηR as a key extra hyperparameter, which may impact performance critically. In this work, two novel differentiable losses for OC and OS are presented, which leverage the principle of ordinal binary decompositions (OBD) with no need for per-subtask models. These losses have a nominal term, plus four OBD-related terms pertaining to the four types of ordinal decomposition that exist. We propose two weighting strategies for those OBD loss terms: either based on the frequency of data, or via differential entropy – which accounts for class imbalance. Unlike in UR with ηR, these avoid the need for any loss-related hyperparameter. To illustrative our proposal, four real-world datasets are tackled (three OC, one OS). Eight state-of-the-art baseline losses are used for comparison. Experimental results –regarding the behaviour of the DL models trained with each loss– varied notably across datasets and across performance metrics, indicating trade-offs. Whereas ’naı̈ve’ cross-entropy ranked the best overall, it was sub-par in error distance-based metrics. An UR and an OBD loss functions were the subsequent best, and safe in the sense of no sub-par results in any facet. In all, OBD proved to be a convenient option for OC/OS, without inductive biasesimplicit in its formulation (unlike UR).
- Open Access
- 26 August 2025
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 20 (8) , e0322101
Modelling in-hospital length of stay: A comparison of linear and ensemble models for competing risk analysis
Length of Stay (LoS) for in-hospital patients is a relevant indicator of efficiency in healthcare. Moreover, it is often related to the occurrence of hospital-acquired complications. In this work, we aim to explore time-to-event analysis for modelling LoS. We employed competing risk models (CR), as we considered two mutually exclusive outcomes: favorable discharge and deterioration. The explanatory variables included the patient’s sex, age, and longitudinal vital signs collected from a dataset comprising admissions. To address sparse measurements, we transformed longitudinal vital signs into cross-sectional statistics. Our approach involves data pre-processing, imputation of missing data, and variable selection. We proposed four types of CR models: Cause-specific Cox, Sub-distribution hazard, and two variants of Random Survival Forests, with both generalised Log-Rank test (cause-specific hazard estimates) and Gray’s test (cumulative incidences estimations) as node splitting rules. Performance in LoS CR models was evaluated over a time frame from 2 to 15 days. Additionally, we considered baselines with two well-established clinical early warning scores the National Early Warning Score (NEWS) and the Modified Early Warning Score (MEWS). The best model was Random Survival Forest using Gray’s test split, with Integrated Brier Score[×100] of 0.386, C-Index above 99%, and Brier Score below 0.006, along the entire time frame. Employing cross-sectional statistics derived from vital signs, along with rigorous data pre-processing, outperformed the degree of correctness of modelling LoS, compared to NEWS and MEWS.- 26 August 2025
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 20 (8) , e0322101
Open AccessModelling in-hospital length of stay: A comparison of linear and ensemble models for competing risk analysis
Length of Stay (LoS) for in-hospital patients is a relevant indicator of efficiency in healthcare. Moreover, it is often related to the occurrence of hospital-acquired complications. In this work, we aim to explore time-to-event analysis for modelling LoS. We employed competing risk models (CR), as we considered two mutually exclusive outcomes: favorable discharge and deterioration. The explanatory variables included the patient’s sex, age, and longitudinal vital signs collected from a dataset comprising admissions. To address sparse measurements, we transformed longitudinal vital signs into cross-sectional statistics. Our approach involves data pre-processing, imputation of missing data, and variable selection. We proposed four types of CR models: Cause-specific Cox, Sub-distribution hazard, and two variants of Random Survival Forests, with both generalised Log-Rank test (cause-specific hazard estimates) and Gray’s test (cumulative incidences estimations) as node splitting rules. Performance in LoS CR models was evaluated over a time frame from 2 to 15 days. Additionally, we considered baselines with two well-established clinical early warning scores the National Early Warning Score (NEWS) and the Modified Early Warning Score (MEWS). The best model was Random Survival Forest using Gray’s test split, with Integrated Brier Score[×100] of 0.386, C-Index above 99%, and Brier Score below 0.006, along the entire time frame. Employing cross-sectional statistics derived from vital signs, along with rigorous data pre-processing, outperformed the degree of correctness of modelling LoS, compared to NEWS and MEWS. - Open Access
- 25 June 2024
- Published by Springer Nature in Pneumonia
- Vol. 16 (1) , 1-16
Obtaining patient phenotypes in SARS-CoV-2 pneumonia, and their association with clinical severity and mortality
Background: There exists consistent empirical evidence in the literature pointing out ample heterogeneity in terms of the clinical evolution of patients with COVID-19. The identification of specific phenotypes underlying in the population might contribute towards a better understanding and characterization of the different courses of the disease. The aim of this study was to identify distinct clinical phenotypes among hospitalized patients with SARS-CoV-2 pneumonia using machine learning clustering, and to study their association with subsequent clinical outcomes as severity and mortality. Methods: Multicentric observational, prospective, longitudinal, cohort study conducted in four hospitals in Spain. We included adult patients admitted for in-hospital stay due to SARS-CoV-2 pneumonia. We collected a broad spectrum of variables to describe exhaustively each case: patient demographics, comorbidities, symptoms, physiological status, baseline examinations (blood analytics, arterial gas test), etc. For the development and internal validation of the clustering/phenotype models, the dataset was split into training and test sets (50% each). We proposed a sequence of machine learning stages: feature scaling, missing data imputation, reduction of data dimensionality via Kernel Principal Component Analysis (KPCA), and clustering with the k-means algorithm. The optimal cluster model parameters –including k, the number of phenotypes– were chosen automatically, by maximizing the average Silhouette score across the training set. Results: We enrolled 1548 patients, each of them characterized by 92 clinical attributes (d=109 features after variable encoding). Our clustering algorithm identified k=3 distinct phenotypes and 18 strongly informative variables: Phenotype A (788 cases [50.9% prevalence] – age $$\sim$$ 57, Charlson comorbidity $$\sim$$ 1, pneumonia CURB-65 score $$\sim$$ 0 to 1, respiratory rate at admission $$\sim$$ 18 min-1, FiO2 $$\sim$$ 21%, C-reactive protein CRP $$\sim$$ 49.5 mg/dL [median within cluster]); phenotype B (620 cases [40.0%] – age $$\sim$$ 75, Charlson $$\sim$$ 5, CURB-65 $$\sim$$ 1 to 2, respiration $$\sim$$ 20 min-1, FiO2 $$\sim$$ 21%, CRP $$\sim$$ 101.5 mg/dL); and phenotype C (140 cases [9.0%] – age $$\sim$$ 71, Charlson $$\sim$$ 4, CURB-65 $$\sim$$ 0 to 2, respiration $$\sim$$ 30 min-1, FiO2 $$\sim$$ 38%, CRP $$\sim$$ 152.3 mg/dL). Hypothesis testing provided solid statistical evidence supporting an interaction between phenotype and each clinical outcome: severity and mortality. By computing their corresponding odds ratios, a clear trend was found for higher frequencies of unfavourable evolution in phenotype C with respect to B, as well as more unfavourable in phenotype B than in A. Conclusion: A compound unsupervised clustering technique (including a fully-automated optimization of its internal parameters) revealed the existence of three distinct groups of patients – phenotypes. In turn, these showed strong associations with the clinical severity in the progression of pneumonia, and with mortality.- 25 June 2024
- Published by Springer Nature in Pneumonia
- Vol. 16 (1) , 1-16
Open AccessObtaining patient phenotypes in SARS-CoV-2 pneumonia, and their association with clinical severity and mortality
Background: There exists consistent empirical evidence in the literature pointing out ample heterogeneity in terms of the clinical evolution of patients with COVID-19. The identification of specific phenotypes underlying in the population might contribute towards a better understanding and characterization of the different courses of the disease. The aim of this study was to identify distinct clinical phenotypes among hospitalized patients with SARS-CoV-2 pneumonia using machine learning clustering, and to study their association with subsequent clinical outcomes as severity and mortality. Methods: Multicentric observational, prospective, longitudinal, cohort study conducted in four hospitals in Spain. We included adult patients admitted for in-hospital stay due to SARS-CoV-2 pneumonia. We collected a broad spectrum of variables to describe exhaustively each case: patient demographics, comorbidities, symptoms, physiological status, baseline examinations (blood analytics, arterial gas test), etc. For the development and internal validation of the clustering/phenotype models, the dataset was split into training and test sets (50% each). We proposed a sequence of machine learning stages: feature scaling, missing data imputation, reduction of data dimensionality via Kernel Principal Component Analysis (KPCA), and clustering with the k-means algorithm. The optimal cluster model parameters –including k, the number of phenotypes– were chosen automatically, by maximizing the average Silhouette score across the training set. Results: We enrolled 1548 patients, each of them characterized by 92 clinical attributes (d=109 features after variable encoding). Our clustering algorithm identified k=3 distinct phenotypes and 18 strongly informative variables: Phenotype A (788 cases [50.9% prevalence] – age $$\sim$$ 57, Charlson comorbidity $$\sim$$ 1, pneumonia CURB-65 score $$\sim$$ 0 to 1, respiratory rate at admission $$\sim$$ 18 min-1, FiO2 $$\sim$$ 21%, C-reactive protein CRP $$\sim$$ 49.5 mg/dL [median within cluster]); phenotype B (620 cases [40.0%] – age $$\sim$$ 75, Charlson $$\sim$$ 5, CURB-65 $$\sim$$ 1 to 2, respiration $$\sim$$ 20 min-1, FiO2 $$\sim$$ 21%, CRP $$\sim$$ 101.5 mg/dL); and phenotype C (140 cases [9.0%] – age $$\sim$$ 71, Charlson $$\sim$$ 4, CURB-65 $$\sim$$ 0 to 2, respiration $$\sim$$ 30 min-1, FiO2 $$\sim$$ 38%, CRP $$\sim$$ 152.3 mg/dL). Hypothesis testing provided solid statistical evidence supporting an interaction between phenotype and each clinical outcome: severity and mortality. By computing their corresponding odds ratios, a clear trend was found for higher frequencies of unfavourable evolution in phenotype C with respect to B, as well as more unfavourable in phenotype B than in A. Conclusion: A compound unsupervised clustering technique (including a fully-automated optimization of its internal parameters) revealed the existence of three distinct groups of patients – phenotypes. In turn, these showed strong associations with the clinical severity in the progression of pneumonia, and with mortality. - 12 March 2024
- Published by Elsevier in Computer Methods and Programs in Biomedicine
- Vol. 248, 108118
Reliable prediction of difficult airway for tracheal intubation from patient preoperative photographs by machine learning methods
- 12 March 2024
- Published by Elsevier in Computer Methods and Programs in Biomedicine
- Vol. 248, 108118
Reliable prediction of difficult airway for tracheal intubation from patient preoperative photographs by machine learning methods
- 8 February 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 28 (5) , 2613-2623
Cost-Sensitive Ordinal Classification Methods to Predict SARS-CoV-2 Pneumonia Severity
Objective: To study the suitability of costsensitive ordinal artificial intelligence-machine learning (AIML) strategies in the prognosis of SARS-CoV-2 pneumonia severity. Materials & methods: Observational, retrospective, longitudinal, cohort study in 4 hospitals in Spain. Information regarding demographic and clinical status was supplemented by socioeconomic data and air pollution exposures. We proposed AI-ML algorithms for ordinal classification via ordinal decomposition and for cost-sensitive learning via resampling techniques. For performance-based model selection, we defined a custom score including per-class sensitivities and asymmetric misprognosis costs. 260 distinct AI-ML models were evaluated via 10 repetitions of 5×5 nested cross-validation with hyperparameter tuning. Model selection was followed by the calibration of predicted probabilities. Final overall performance was compared against five well-established clinical severity scores and against a ‘standard’ (non-cost sensitive, non-ordinal) AI-ML baseline. In our best model, we also evaluated its explainability with respect to each of the input variables. Results: The study enrolled n = 1548 patients: 712 experienced low, 238 medium, and 598 high clinical severity. d = 131 variables were collected, becoming d ′ = 148 features after categorical encoding. Model selection resulted in our best-performing AI-ML pipeline having: a) no imputation of missing data, b) no feature selection (i.e. using the full set of d ′ features), c) ‘Ordered Partitions’ ordinal decomposition, d) cost-based reimbalance, and e) a Histogram-based Gradient Boosting classifier. This best model (calibrated) obtained a median accuracy of 68.1% [67.3%, 68.8%] (95% confidence interval), a balanced accuracy of 57.0% [55.6%, 57.9%], and an overall area under the curve (AUC) 0.802 [0.795, 0.808]. In our dataset, it outperformed all five clinical severity scores and the ‘standard’ AI-ML baseline. Discussion & conclusion: We conducted an exhaustive exploration of AI-ML methods designed for both ordinal and cost-sensitive classification, motivated by a real-world application domain (clinical severity prognosis) in which these topics arise naturally. Our model with the best classification performance exploited successfully the ordering information of ground truth classes, coping with imbalance and asymmetric costs. However, these ordinal and cost-sensitive aspects are seldom explored in the literature.
- 8 February 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 28 (5) , 2613-2623
Cost-Sensitive Ordinal Classification Methods to Predict SARS-CoV-2 Pneumonia Severity
Objective: To study the suitability of costsensitive ordinal artificial intelligence-machine learning (AIML) strategies in the prognosis of SARS-CoV-2 pneumonia severity. Materials & methods: Observational, retrospective, longitudinal, cohort study in 4 hospitals in Spain. Information regarding demographic and clinical status was supplemented by socioeconomic data and air pollution exposures. We proposed AI-ML algorithms for ordinal classification via ordinal decomposition and for cost-sensitive learning via resampling techniques. For performance-based model selection, we defined a custom score including per-class sensitivities and asymmetric misprognosis costs. 260 distinct AI-ML models were evaluated via 10 repetitions of 5×5 nested cross-validation with hyperparameter tuning. Model selection was followed by the calibration of predicted probabilities. Final overall performance was compared against five well-established clinical severity scores and against a ‘standard’ (non-cost sensitive, non-ordinal) AI-ML baseline. In our best model, we also evaluated its explainability with respect to each of the input variables. Results: The study enrolled n = 1548 patients: 712 experienced low, 238 medium, and 598 high clinical severity. d = 131 variables were collected, becoming d ′ = 148 features after categorical encoding. Model selection resulted in our best-performing AI-ML pipeline having: a) no imputation of missing data, b) no feature selection (i.e. using the full set of d ′ features), c) ‘Ordered Partitions’ ordinal decomposition, d) cost-based reimbalance, and e) a Histogram-based Gradient Boosting classifier. This best model (calibrated) obtained a median accuracy of 68.1% [67.3%, 68.8%] (95% confidence interval), a balanced accuracy of 57.0% [55.6%, 57.9%], and an overall area under the curve (AUC) 0.802 [0.795, 0.808]. In our dataset, it outperformed all five clinical severity scores and the ‘standard’ AI-ML baseline. Discussion & conclusion: We conducted an exhaustive exploration of AI-ML methods designed for both ordinal and cost-sensitive classification, motivated by a real-world application domain (clinical severity prognosis) in which these topics arise naturally. Our model with the best classification performance exploited successfully the ordering information of ground truth classes, coping with imbalance and asymmetric costs. However, these ordinal and cost-sensitive aspects are seldom explored in the literature.
- Open Access
- 13 April 2023
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 18 (4) , e0284150
Extracting relevant predictive variables for COVID-19 severity prognosis: An exhaustive comparison of feature selection techniques
With the COVID-19 pandemic having caused unprecedented numbers of infections and deaths, large research efforts have been undertaken to increase our understanding of the disease and the factors which determine diverse clinical evolutions. Here we focused on a fully data-driven exploration regarding which factors (clinical or otherwise) were most informative for SARS-CoV-2 pneumonia severity prediction via machine learning (ML). In particular, feature selection techniques (FS), designed to reduce the dimensionality of data, allowed us to characterize which of our variables were the most useful for ML prognosis. We conducted a multi-centre clinical study, enrollingn= 1548 patients hospitalized due to SARS-CoV-2 pneumonia: where 792, 238, and 598 patients experienced low, medium and high-severity evolutions, respectively. Up to 106 patient-specific clinical variables were collected at admission, although 14 of them had to be discarded for containing ⩾60% missing values. Alongside 7 socioeconomic attributes and 32 exposures to air pollution (chronic and acute), these becamed= 148 features after variable encoding. We addressed this ordinal classification problem both as a ML classification and regression task. Two imputation techniques for missing data were explored, along with a total of 166 unique FS algorithm configurations: 46 filters, 100 wrappers and 20 embeddeds. Of these, 21 setups achieved satisfactory bootstrap stability (⩾0.70) with reasonable computation times: 16 filters, 2 wrappers, and 3 embeddeds. The subsets of features selected by each technique showed modest Jaccard similarities across them. However, they consistently pointed out the importance of certain explanatory variables. Namely: patient’s C-reactive protein (CRP), pneumonia severity index (PSI), respiratory rate (RR) and oxygen levels –saturation Sp O2, quotients Sp O2/RR and arterial Sat O2/Fi O2–, the neutrophil-to-lymphocyte ratio (NLR) –to certain extent, also neutrophil and lymphocyte counts separately–, lactate dehydrogenase (LDH), and procalcitonin (PCT) levels in blood. A remarkable agreement has been founda posterioribetween our strategy and independent clinical research works investigating risk factors for COVID-19 severity. Hence, these findings stress the suitability of this type of fully data-driven approaches for knowledge extraction, as a complementary to clinical perspectives.
- 13 April 2023
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 18 (4) , e0284150
Open AccessExtracting relevant predictive variables for COVID-19 severity prognosis: An exhaustive comparison of feature selection techniques
With the COVID-19 pandemic having caused unprecedented numbers of infections and deaths, large research efforts have been undertaken to increase our understanding of the disease and the factors which determine diverse clinical evolutions. Here we focused on a fully data-driven exploration regarding which factors (clinical or otherwise) were most informative for SARS-CoV-2 pneumonia severity prediction via machine learning (ML). In particular, feature selection techniques (FS), designed to reduce the dimensionality of data, allowed us to characterize which of our variables were the most useful for ML prognosis. We conducted a multi-centre clinical study, enrollingn= 1548 patients hospitalized due to SARS-CoV-2 pneumonia: where 792, 238, and 598 patients experienced low, medium and high-severity evolutions, respectively. Up to 106 patient-specific clinical variables were collected at admission, although 14 of them had to be discarded for containing ⩾60% missing values. Alongside 7 socioeconomic attributes and 32 exposures to air pollution (chronic and acute), these becamed= 148 features after variable encoding. We addressed this ordinal classification problem both as a ML classification and regression task. Two imputation techniques for missing data were explored, along with a total of 166 unique FS algorithm configurations: 46 filters, 100 wrappers and 20 embeddeds. Of these, 21 setups achieved satisfactory bootstrap stability (⩾0.70) with reasonable computation times: 16 filters, 2 wrappers, and 3 embeddeds. The subsets of features selected by each technique showed modest Jaccard similarities across them. However, they consistently pointed out the importance of certain explanatory variables. Namely: patient’s C-reactive protein (CRP), pneumonia severity index (PSI), respiratory rate (RR) and oxygen levels –saturation Sp O2, quotients Sp O2/RR and arterial Sat O2/Fi O2–, the neutrophil-to-lymphocyte ratio (NLR) –to certain extent, also neutrophil and lymphocyte counts separately–, lactate dehydrogenase (LDH), and procalcitonin (PCT) levels in blood. A remarkable agreement has been founda posterioribetween our strategy and independent clinical research works investigating risk factors for COVID-19 severity. Hence, these findings stress the suitability of this type of fully data-driven approaches for knowledge extraction, as a complementary to clinical perspectives.
- 24 February 2023
- Published by Elsevier in Computer Methods and Programs in Biomedicine
- Vol. 232, 107428
Automated location of orofacial landmarks to characterize airway morphology in anaesthesia via deep convolutional neural networks
- 24 February 2023
- Published by Elsevier in Computer Methods and Programs in Biomedicine
- Vol. 232, 107428
Automated location of orofacial landmarks to characterize airway morphology in anaesthesia via deep convolutional neural networks
- Open Access
- 7 May 2016
- Published by Springer Nature in Medical & Biological Engineering & Computing
- Vol. 55 (2) , 271-282
Modelling the effect of insulin on the disposal of meal-attributable glucose in type 1 diabetes
The management of postprandial glucose excursions in type 1 diabetes has a major impact on overall glycaemic control. In this work, we propose and evaluate various mechanistic models to characterize the disposal of meal-attributable glucose. Sixteen young volunteers with type 1 diabetes were subject to a variable-target clamp which replicated glucose profiles observed after a high-glycaemic-load (\(n=8\)) or a low-glycaemic-load (\(n=8\)) evening meal. [6,6-\(^{2}\hbox {H}_2\)] and [U-\(^{13}\hbox {C}\);1,2,3,4,5,6,6-\(^{2}\hbox {H}_{7}\)] glucose tracers were infused to, respectively, mimic: (a) the expected post-meal suppression of endogenous glucose production and (b) the appearance of glucose due to a standard meal. Six compartmental models (all a priori identifiable) were proposed to investigate the remote effect of circulating plasma insulin on the disposal of those glucose tracers from the non-accessible compartments, representing e.g. interstitium. An iterative population-based parameter fitting was employed. Models were evaluated attending to physiological plausibility, posterior identifiability of their parameter estimates, accuracy—via weighted fitting residuals—and information criteria (i.e. parsimony). The most plausible model, best representing our experimental data, comprised: (1) a remote effect x of insulin active above a threshold \(x_{C}\) = 1.74 (0.81–2.50) \(\cdot \,10^{-2}\) min\(^{-1}\) [median (inter-quartile range)], with parameter \(x_{C}\) having a satisfactory support: coefficient of variation CV = 42.33 (31.34–65.34) %, and (2) steady-state conditions at the onset of the experiment (\(t=0\)) for the compartment representing the remote effect, but not for the masses of the tracer that mimicked endogenous glucose production. Consequently, our mechanistic model suggests non-homogeneous changes in the disposal rates for meal-attributable glucose in relation to plasma insulin. The model can be applied to the in silico simulation of meals for the optimization of postprandial insulin infusion regimes in type 1 diabetes.- 7 May 2016
- Published by Springer Nature in Medical & Biological Engineering & Computing
- Vol. 55 (2) , 271-282
Open AccessModelling the effect of insulin on the disposal of meal-attributable glucose in type 1 diabetes
The management of postprandial glucose excursions in type 1 diabetes has a major impact on overall glycaemic control. In this work, we propose and evaluate various mechanistic models to characterize the disposal of meal-attributable glucose. Sixteen young volunteers with type 1 diabetes were subject to a variable-target clamp which replicated glucose profiles observed after a high-glycaemic-load (\(n=8\)) or a low-glycaemic-load (\(n=8\)) evening meal. [6,6-\(^{2}\hbox {H}_2\)] and [U-\(^{13}\hbox {C}\);1,2,3,4,5,6,6-\(^{2}\hbox {H}_{7}\)] glucose tracers were infused to, respectively, mimic: (a) the expected post-meal suppression of endogenous glucose production and (b) the appearance of glucose due to a standard meal. Six compartmental models (all a priori identifiable) were proposed to investigate the remote effect of circulating plasma insulin on the disposal of those glucose tracers from the non-accessible compartments, representing e.g. interstitium. An iterative population-based parameter fitting was employed. Models were evaluated attending to physiological plausibility, posterior identifiability of their parameter estimates, accuracy—via weighted fitting residuals—and information criteria (i.e. parsimony). The most plausible model, best representing our experimental data, comprised: (1) a remote effect x of insulin active above a threshold \(x_{C}\) = 1.74 (0.81–2.50) \(\cdot \,10^{-2}\) min\(^{-1}\) [median (inter-quartile range)], with parameter \(x_{C}\) having a satisfactory support: coefficient of variation CV = 42.33 (31.34–65.34) %, and (2) steady-state conditions at the onset of the experiment (\(t=0\)) for the compartment representing the remote effect, but not for the masses of the tracer that mimicked endogenous glucose production. Consequently, our mechanistic model suggests non-homogeneous changes in the disposal rates for meal-attributable glucose in relation to plasma insulin. The model can be applied to the in silico simulation of meals for the optimization of postprandial insulin infusion regimes in type 1 diabetes. - Open Access
- 1 January 2016
- Published by Georg Thieme Verlag KG in Methods of Information in Medicine
- Vol. 55 (06) , 533-544
Automatic Identification of Physical Activity Intensity and Modality from the Fusion of Accelerometry and Heart Rate Data
Summary: Background: Physical activity (PA) is essential to prevent and to treat a variety of chronic diseases. The automated detection and quantification of PA over time empowers lifestyle interventions, facilitating reliable exercise tracking and data-driven counseling.Methods: We propose and compare various combinations of machine learning (ML) schemes for the automatic classification of PA from multi-modal data, simultaneously captured by a biaxial accelerometer and a heart rate (HR) monitor. Intensity levels (low / moderate / vigorous) were recognized, as well as for vigorous exercise, its modality (sustained aerobic / resistance / mixed). In to -tal, 178.63 h of data about PA intensity (65.55 % low / 18.96 % moderate / 15.49 % vigorous) and 17.00 h about modality were collected in two experiments: one in free- living conditions, another in a fitness center under controlled protocols. The structure used for automatic classification comprised: a) definition of 42 time-domain signal features, b) dimensionality reduction, c) data clustering, and d) temporal filtering to exploit time redundancy by means of a Hidden Markov Model (HMM). Four dimensionality reduction techniques and four clustering algorithms were studied. In order to cope with class imbalance in the dataset, a custom performance metric was defined to aggregate recognition accuracy, precision and recall.Results: The best scheme, which comprised a projection through Linear Discriminant Ana -lysis (LDA) and k-means clustering, was evaluated in leave-one-subject-out cross-validation; notably outperforming the standard industry procedures for PA intensity classification: score 84.65 %, versus up to 63.60 %. Errors tended to be brief and to appear around transients.Conclusions: The application of ML techniques for pattern identification and temporal filtering allowed to merge accelerometry and HR data in a solid manner, and achieved markedly better recognition performances than the standard methods for PA intensity estimation.
- 1 January 2016
- Published by Georg Thieme Verlag KG in Methods of Information in Medicine
- Vol. 55 (06) , 533-544
Open AccessAutomatic Identification of Physical Activity Intensity and Modality from the Fusion of Accelerometry and Heart Rate Data
Summary: Background: Physical activity (PA) is essential to prevent and to treat a variety of chronic diseases. The automated detection and quantification of PA over time empowers lifestyle interventions, facilitating reliable exercise tracking and data-driven counseling.Methods: We propose and compare various combinations of machine learning (ML) schemes for the automatic classification of PA from multi-modal data, simultaneously captured by a biaxial accelerometer and a heart rate (HR) monitor. Intensity levels (low / moderate / vigorous) were recognized, as well as for vigorous exercise, its modality (sustained aerobic / resistance / mixed). In to -tal, 178.63 h of data about PA intensity (65.55 % low / 18.96 % moderate / 15.49 % vigorous) and 17.00 h about modality were collected in two experiments: one in free- living conditions, another in a fitness center under controlled protocols. The structure used for automatic classification comprised: a) definition of 42 time-domain signal features, b) dimensionality reduction, c) data clustering, and d) temporal filtering to exploit time redundancy by means of a Hidden Markov Model (HMM). Four dimensionality reduction techniques and four clustering algorithms were studied. In order to cope with class imbalance in the dataset, a custom performance metric was defined to aggregate recognition accuracy, precision and recall.Results: The best scheme, which comprised a projection through Linear Discriminant Ana -lysis (LDA) and k-means clustering, was evaluated in leave-one-subject-out cross-validation; notably outperforming the standard industry procedures for PA intensity classification: score 84.65 %, versus up to 63.60 %. Errors tended to be brief and to appear around transients.Conclusions: The application of ML techniques for pattern identification and temporal filtering allowed to merge accelerometry and HR data in a solid manner, and achieved markedly better recognition performances than the standard methods for PA intensity estimation.