BMC infectious diseases · 2025
Unraveling relevant cross-waves pattern drifts in patient-hospital risk factors among hospitalized COVID-19 patients using explainable machine learning methods.
СтатьяМногоцентровое исследованиеНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Lana Fernanda Cristina Barbosa, Marinho Carolina Coimbra, de Paiva Bruno Barbosa Miranda, Valle Lucas Rocha, do Nascimento Guilherme Fonseca, da Rocha Leonardo Chaves Dutra, Carneiro Marcelo, Batista Joanna d'Arc Lyra, Anschau Fernando, Paraiso Pedro Gibson, Bartolazzi Frederico, Cimini Christiane Corrêa Rodrigues, Schwarzbold Alexandre Vargas, Rios Danyelle Romana Alves, Gonçalves Marcos André, Marcolino Milena Soriano
Аннотация
BACKGROUND: Several studies explored factors related to adverse clinical outcomes among COVID-19 patients but lacked analysis of the impact of the temporal data shifts on the strength of association between different predictors and adverse outcomes. This study aims to evaluate factors related to patients and hospitals in the prediction of in-hospital mortality, need for invasive mechanical ventilation (IMV), and intensive care unit (ICU) transfer throughout the pandemic waves.
METHODS: This multicenter retrospective cohort included COVID-19 patients from 39 hospitals, from March/2020 to August/2022. The pandemic was divided into waves: 10/03/2020-14/11/2020 (first), 15/11/2020-25/12/2021 (second), 26/12/2021-03/08/2022 (third). Patient-related factors included clinical, demographic, and laboratory data, while hospital-related factors covered funding sources, accreditation, academic status, and socioeconomic characteristics. Shapley additive explanation (SHAP) values derived from the predictions of a light gradient-boosting machine (LightGBM) model were used to assess potential risk factors for death, IMV and ICU.
RESULTS: Overall, 16,958 adult patients were included (median age 59 years, 54.7% men). LightGBM achieved competitive effectiveness metrics across all periods. Temporal drifts were observed due to a decrease in various metrics, such as the recall for the positive class [ICU: 0.4211 (wave 1) to 0.1951 (wave 3); IMV: 0.2089 (wave 1) to 0.0438 (wave 3); death: 0.2711 (wave 1) to 0.1175 (wave 3)]. Peripheral arterial oxygen saturation to the fraction of inspired oxygen ratio (SatO2/FiO2) at admission had great predictive capacity for all outcomes, with an optimal cut-off value for death prediction of 227.78. Lymphopenia had its association strength increased over time for all outcomes, optimal threshold for death prediction of 643 × 109/L. Thrombocytopenia was the most important feature in wave 2 (ICU); overall, values below 143,000 × 109/L were more related to death.
CONCLUSION: Data drifts were observed in all scenarios, affecting potential predictive capabilities of explainable machine learning methods. Upon admission, SatO2/FiO2 values, platelet and lymphocyte count were significant predictors of adverse outcomes in COVID-19 patients. Overall, inflammatory response markers were more important than clinical characteristics. Limitations included sample representativeness and confounding factors. Integrating the drift's knowledge into models to improve effectiveness is a challenge, requiring continuous updates and monitoring of performance in real-world applications.
CLINICAL TRIAL NUMBER: Not applicable.
METHODS: This multicenter retrospective cohort included COVID-19 patients from 39 hospitals, from March/2020 to August/2022. The pandemic was divided into waves: 10/03/2020-14/11/2020 (first), 15/11/2020-25/12/2021 (second), 26/12/2021-03/08/2022 (third). Patient-related factors included clinical, demographic, and laboratory data, while hospital-related factors covered funding sources, accreditation, academic status, and socioeconomic characteristics. Shapley additive explanation (SHAP) values derived from the predictions of a light gradient-boosting machine (LightGBM) model were used to assess potential risk factors for death, IMV and ICU.
RESULTS: Overall, 16,958 adult patients were included (median age 59 years, 54.7% men). LightGBM achieved competitive effectiveness metrics across all periods. Temporal drifts were observed due to a decrease in various metrics, such as the recall for the positive class [ICU: 0.4211 (wave 1) to 0.1951 (wave 3); IMV: 0.2089 (wave 1) to 0.0438 (wave 3); death: 0.2711 (wave 1) to 0.1175 (wave 3)]. Peripheral arterial oxygen saturation to the fraction of inspired oxygen ratio (SatO2/FiO2) at admission had great predictive capacity for all outcomes, with an optimal cut-off value for death prediction of 227.78. Lymphopenia had its association strength increased over time for all outcomes, optimal threshold for death prediction of 643 × 109/L. Thrombocytopenia was the most important feature in wave 2 (ICU); overall, values below 143,000 × 109/L were more related to death.
CONCLUSION: Data drifts were observed in all scenarios, affecting potential predictive capabilities of explainable machine learning methods. Upon admission, SatO2/FiO2 values, platelet and lymphocyte count were significant predictors of adverse outcomes in COVID-19 patients. Overall, inflammatory response markers were more important than clinical characteristics. Limitations included sample representativeness and confounding factors. Integrating the drift's knowledge into models to improve effectiveness is a challenge, requiring continuous updates and monitoring of performance in real-world applications.
CLINICAL TRIAL NUMBER: Not applicable.