tetano
Editor, Senior Moderator
J Med Internet Res
. 2025 Jul 10:27:e70674.
doi: 10.2196/70674. AI Predictive Model of Mortality and Intensive Care Unit Admission in the COVID-19 Pandemic: Retrospective Population Cohort Study of 12,000 Patients
Jose Manuel Ruiz Giardin[SUP] #[/SUP][SUP] 1 2 [/SUP], Óscar Garnica[SUP] #[/SUP][SUP] 3 [/SUP], Nieves Mesa Plaza[SUP] 4 [/SUP], Juan Víctor SanMartín López[SUP] 1 2 [/SUP], Ana Farfán Sedano[SUP] 4 [/SUP], Elena Madroñal Cerezo[SUP] 4 [/SUP], Miguel Ángel Duarte Millán[SUP] 4 [/SUP], Aida Izquierdo Martínez[SUP] 4 [/SUP], Luis Rivas[SUP] 4 [/SUP], Marta Rivilla[SUP] 4 [/SUP], Alejandro Morales Ortega[SUP] 4 [/SUP], Begoña Frutos Pérez[SUP] 4 [/SUP], Cristina De Ancos Aracil[SUP] 4 [/SUP], Ruth Calderón[SUP] 1 [/SUP], Guillermo Soria Fernandez[SUP] 1 [/SUP], Jorge Marrero Francés[SUP] 4 [/SUP], David Bernal Bello[SUP] 4 [/SUP], Jose Ángel Satué Bartolomé[SUP] 4 [/SUP], María Toledano Macías[SUP] 4 [/SUP], Sara Piedrabuena García[SUP] 4 [/SUP], Marta Guerrero Santillán[SUP] 4 [/SUP], Rafael Cristóbal[SUP] 4 [/SUP], Belen Mora[SUP] 4 [/SUP], Laura Velázquez Ríos[SUP] 4 [/SUP], Vanesa García de Viedma[SUP] 4 [/SUP], Paula Cuenca Ruiz[SUP] 1 [/SUP], Ibone Ayala Larrañaga[SUP] 4 [/SUP], Lorena Carpintero[SUP] 1 [/SUP], Celia Lara[SUP] 4 [/SUP], Alvaro Ricardo Llerena[SUP] 4 [/SUP], Virginia García Bermúdez[SUP] 4 [/SUP], Gema Delgado Cárdenas[SUP] 5 [/SUP], Paloma Pardo Rovira[SUP] 5 [/SUP], Elena Tejero Sánchez[SUP] 5 [/SUP], Maria Jesús Domínguez García[SUP] 5 [/SUP], Carolina Mariño[SUP] 6 [/SUP], Cristina Bravo[SUP] 6 [/SUP], Ana Ontañon[SUP] 6 [/SUP], Mario García[SUP] 6 [/SUP], Jose Ignacio Hidalgo Pérez[SUP] 3 [/SUP], Santiago Prieto Menchero[SUP] 7 [/SUP], Natalia González Pereira[SUP] 7 [/SUP], Sonia Gonzalo Pascua[SUP] 4 [/SUP], Jorge Tarancón Rey[SUP] 8 [/SUP], Luis Antonio Lechuga Suárez[SUP] 8 [/SUP]; FUENCOVID[SUP] 9 [/SUP]
Affiliations
Background: One of the main challenges with COVID-19 has been that although there are known factors associated with a worse prognosis, clinicians have been unable to predict which patients, with similar risk factors, will die or require intensive care unit (ICU) care.
Objective: This study aimed to develop a personalized artificial intelligence model to predict the patient risk of mortality and ICU admission related to SARS-CoV-2 infection during the initial medical evaluation before any kind of treatment.
Methods: It is a population-based, observational, retrospective study covering from February 1, 2020, to January 24, 2023, with different circulating SARS-CoV-2 viruses, vaccinated status, and reinfections. It includes patients attended by the reference hospital in Fuenlabrada (Madrid, Spain). The models used the random forest technique, Shapley Additive Explanations method, and processing with Python (version 3.10.0; Python Software Foundation) and scikit-learn (version 1.3.0). The models were applied to different epidemic SARS-CoV-2 infection waves. Data were collected from 11,975 patients (4998 hospitalized and 6737 discharged). Predictive models were built with records from 4758 patients and validated with 6977 patients after evaluation in the emergency department. Variables recorded were age, sex, place of birth, clinical data, laboratory results, vaccination status, and radiologic data at admission.
Results: The best mortality predictor achieved an area under the receiver operating characteristic curve (AUC) of 0.92, sensitivity of 0.89, specificity of 0.82, positive predictive value (PPV) of 0.35, and mean negative predictive value (NPV) of 0.98. The ICU admission predictor had an AUC of 0.89, sensitivity of 0.75, specificity of 0.88, PPV of 0.37, and NPV of 0.98. During validation, the mortality model exhibited good performance for the nonhospitalized group, achieving an AUC of 0.95, sensitivity of 0.88, specificity of 0.98, PPV of 0.21, and NPV of 0.99, predicting the death of 30 of 34 patients who were not hospitalized. For the hospitalized patients, the mortality model achieved an AUC of 0.85, sensitivity of 0.86, specificity of 0.74, PPV of 0.24, and NPV of 0.98. The model for predicting ICU admission had an AUC of 0.82, sensitivity of 1.00, specificity of 0.59, PPV of 0.05, and NPV of 1.00. The models' metrics presented stability along all pandemic waves. Key mortality predictors included age, Charlson value, and tachypnea. The worse prognosis was linked to high values in urea, erythrocyte distribution width, oxygen demand, creatinine, procalcitonin, lactate dehydrogenase, heart failure, D-dimer, oncological and hematological diseases, neutrophil, and heart rate. A better prognosis was linked to higher values of lymphocytes and systolic and diastolic blood pressures. Partial or no vaccination provided less protection than full vaccination.
Conclusions: The artificial intelligence models demonstrated stability across pandemic waves, indicating their potential to assist in personal health services during the 3-year pandemic, particularly in early preventive and predictive clinical situations.
Keywords: COVID-19; SARS-CoV-2; artificial intelligence; death; intensive care unit; mortality; population study; predictive model; random forest.
. 2025 Jul 10:27:e70674.
doi: 10.2196/70674. AI Predictive Model of Mortality and Intensive Care Unit Admission in the COVID-19 Pandemic: Retrospective Population Cohort Study of 12,000 Patients
Jose Manuel Ruiz Giardin[SUP] #[/SUP][SUP] 1 2 [/SUP], Óscar Garnica[SUP] #[/SUP][SUP] 3 [/SUP], Nieves Mesa Plaza[SUP] 4 [/SUP], Juan Víctor SanMartín López[SUP] 1 2 [/SUP], Ana Farfán Sedano[SUP] 4 [/SUP], Elena Madroñal Cerezo[SUP] 4 [/SUP], Miguel Ángel Duarte Millán[SUP] 4 [/SUP], Aida Izquierdo Martínez[SUP] 4 [/SUP], Luis Rivas[SUP] 4 [/SUP], Marta Rivilla[SUP] 4 [/SUP], Alejandro Morales Ortega[SUP] 4 [/SUP], Begoña Frutos Pérez[SUP] 4 [/SUP], Cristina De Ancos Aracil[SUP] 4 [/SUP], Ruth Calderón[SUP] 1 [/SUP], Guillermo Soria Fernandez[SUP] 1 [/SUP], Jorge Marrero Francés[SUP] 4 [/SUP], David Bernal Bello[SUP] 4 [/SUP], Jose Ángel Satué Bartolomé[SUP] 4 [/SUP], María Toledano Macías[SUP] 4 [/SUP], Sara Piedrabuena García[SUP] 4 [/SUP], Marta Guerrero Santillán[SUP] 4 [/SUP], Rafael Cristóbal[SUP] 4 [/SUP], Belen Mora[SUP] 4 [/SUP], Laura Velázquez Ríos[SUP] 4 [/SUP], Vanesa García de Viedma[SUP] 4 [/SUP], Paula Cuenca Ruiz[SUP] 1 [/SUP], Ibone Ayala Larrañaga[SUP] 4 [/SUP], Lorena Carpintero[SUP] 1 [/SUP], Celia Lara[SUP] 4 [/SUP], Alvaro Ricardo Llerena[SUP] 4 [/SUP], Virginia García Bermúdez[SUP] 4 [/SUP], Gema Delgado Cárdenas[SUP] 5 [/SUP], Paloma Pardo Rovira[SUP] 5 [/SUP], Elena Tejero Sánchez[SUP] 5 [/SUP], Maria Jesús Domínguez García[SUP] 5 [/SUP], Carolina Mariño[SUP] 6 [/SUP], Cristina Bravo[SUP] 6 [/SUP], Ana Ontañon[SUP] 6 [/SUP], Mario García[SUP] 6 [/SUP], Jose Ignacio Hidalgo Pérez[SUP] 3 [/SUP], Santiago Prieto Menchero[SUP] 7 [/SUP], Natalia González Pereira[SUP] 7 [/SUP], Sonia Gonzalo Pascua[SUP] 4 [/SUP], Jorge Tarancón Rey[SUP] 8 [/SUP], Luis Antonio Lechuga Suárez[SUP] 8 [/SUP]; FUENCOVID[SUP] 9 [/SUP]
Affiliations
- PMID: 40638909
- DOI: 10.2196/70674
Background: One of the main challenges with COVID-19 has been that although there are known factors associated with a worse prognosis, clinicians have been unable to predict which patients, with similar risk factors, will die or require intensive care unit (ICU) care.
Objective: This study aimed to develop a personalized artificial intelligence model to predict the patient risk of mortality and ICU admission related to SARS-CoV-2 infection during the initial medical evaluation before any kind of treatment.
Methods: It is a population-based, observational, retrospective study covering from February 1, 2020, to January 24, 2023, with different circulating SARS-CoV-2 viruses, vaccinated status, and reinfections. It includes patients attended by the reference hospital in Fuenlabrada (Madrid, Spain). The models used the random forest technique, Shapley Additive Explanations method, and processing with Python (version 3.10.0; Python Software Foundation) and scikit-learn (version 1.3.0). The models were applied to different epidemic SARS-CoV-2 infection waves. Data were collected from 11,975 patients (4998 hospitalized and 6737 discharged). Predictive models were built with records from 4758 patients and validated with 6977 patients after evaluation in the emergency department. Variables recorded were age, sex, place of birth, clinical data, laboratory results, vaccination status, and radiologic data at admission.
Results: The best mortality predictor achieved an area under the receiver operating characteristic curve (AUC) of 0.92, sensitivity of 0.89, specificity of 0.82, positive predictive value (PPV) of 0.35, and mean negative predictive value (NPV) of 0.98. The ICU admission predictor had an AUC of 0.89, sensitivity of 0.75, specificity of 0.88, PPV of 0.37, and NPV of 0.98. During validation, the mortality model exhibited good performance for the nonhospitalized group, achieving an AUC of 0.95, sensitivity of 0.88, specificity of 0.98, PPV of 0.21, and NPV of 0.99, predicting the death of 30 of 34 patients who were not hospitalized. For the hospitalized patients, the mortality model achieved an AUC of 0.85, sensitivity of 0.86, specificity of 0.74, PPV of 0.24, and NPV of 0.98. The model for predicting ICU admission had an AUC of 0.82, sensitivity of 1.00, specificity of 0.59, PPV of 0.05, and NPV of 1.00. The models' metrics presented stability along all pandemic waves. Key mortality predictors included age, Charlson value, and tachypnea. The worse prognosis was linked to high values in urea, erythrocyte distribution width, oxygen demand, creatinine, procalcitonin, lactate dehydrogenase, heart failure, D-dimer, oncological and hematological diseases, neutrophil, and heart rate. A better prognosis was linked to higher values of lymphocytes and systolic and diastolic blood pressures. Partial or no vaccination provided less protection than full vaccination.
Conclusions: The artificial intelligence models demonstrated stability across pandemic waves, indicating their potential to assist in personal health services during the 3-year pandemic, particularly in early preventive and predictive clinical situations.
Keywords: COVID-19; SARS-CoV-2; artificial intelligence; death; intensive care unit; mortality; population study; predictive model; random forest.