JMIR Publications

ABSTRACT

Background:

During all the pandemia period one of the main challenges with COVID-19 is 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 ICU care.

Objective:

This study aimed to develop, in the first medical patients evaluation, a personalized Artificial Intelligence (AI) model for predicting patient risk of death and/or ICU admission related to SARS-COV-2 infection, before receiving any kind of treatment.

Methods:

This is a population-based, observational, and retrospective study covering three years of evolution (Feb 20th, 2020, and Jan 20th, 2023), including all patients attended by the only reference hospital in the city of Fuenlabrada (Madrid, Spain). The models employed Random Forest technique. Feature importance was evaluated using SHAP method. All processing was performed using Python 3.10.0 and scikit-learn 1.3.0. The model was also applied to different epidemic SARS-COV-2 infection waves to reflect the different circulating SARS-COV-2 viruses. Data were collected from 11,975 patients at Hospital de Fuenlabrada. Of all evaluated patients, 4,998 required hospitalizations, and 6,737 were discharged. Predictive models were built with records from 4,758 patients and validated with 6,977 patients (hospitalized and non-hospitalized after evaluation in the emergency department). Variables recorded were epidemiological data (age, sex, place of birth, underlying health status), clinical data (including Charlson index), laboratory results (CPR, IL6, ferritin, hematologic data, kidney and liver function), vaccination status, and radiologic data at admission. Using these variables, predictive models of ICU admission and/or death evolution have been created to help clinicians personalize medical services.

Results:

This is a population-based, observational, and retrospective study covering three years of evolution (Feb-20th, 2020, and Jan-20th, 2023), including all patients attended by the only reference hospital in the city of Fuenlabrada (Madrid, Spain). The models employed Random Forest technique. Feature importance was evaluated using SHAP method. All processing was performed using Python 3.10.0 and scikit-learn 1.3.0. The model was also applied to different epidemic SARS-COV-2 infection waves to reflect the different circulating SARS-COV-2 viruses. Data were collected from 11,975 patients . 4,998 required hospitalizations, and 6,737 were discharged. Predictive models were built with records from 4,758 patients and validated with 6,977 patients (hospitalized and non-hospitalized after evaluation in the emergency department). Variables recorded were epidemiological data (age, sex, place of birth, ), clinical data (including Charlson index), laboratory results (CPR, IL6, ferritin, hematologic data, kidney and liver function), vaccination status, and radiologic data at admission.

Conclusions:

The AI model demonstrated stability across pandemic waves, indicating its potential to assist in personal health services during the three-year pandemic, particularly in early preventive and predictive clinical situations.