RESUMO
ABSTRACT Background: Artificial intelligence (AI) in radiology has improved diagnostic performance and shortened reading times of coronavirus disease 2019 (COVID-19) patients studies. Objectives: The objectives pf the study were to analyze the performance of a chest computed tomography (CT) AI quantitative algorithm for determining the risk of mortality/mechanical ventilation (MV) in hospitalized COVID-19 patients and explore a prognostic multivariate model in a tertiary-care center in Mexico City. Methods: Chest CT images of 166 COVID-19 patients hospitalized from April 1 to 20, 2020, were retrospectively analyzed using AI algorithm software. Data were collected from their medical records. We analyzed the diagnostic yield of the relevant CT variables using the area under the ROC curve (area under the curve [AUC]). Optimal thresholds were obtained using the Youden index. We proposed a predictive logistic model for each outcome based on CT AI measures and predetermined laboratory and clinical characteristics. Results: The highest diagnostic yield of the assessed CT variables for mortality was the percentage of total opacity (threshold >51%; AUC = 0.88, sensitivity = 74%, and specificity = 91%). The AUC of the CT severity score (threshold > 12.5) was 0.88 for MV (sensitivity = 65% and specificity = 92%). The proposed prognostic models include the percentage of opacity and lactate dehydrogenase level for mortality and troponin I and CT severity score for MV requirement. Conclusion: The AI-calculated CT severity score and total opacity percentage showed good diagnostic accuracy for mortality and met MV criteria. The proposed prognostic models using biochemical variables and imaging data measured by AI on chest CT showed good risk classification in our population of hospitalized COVID-19 patients.
RESUMO
BACKGROUND: Artificial intelligence (AI) in radiology has improved diagnostic performance and shortened reading times of coronavirus disease 2019 (COVID-19) patients' studies. OBJECTIVES: The objectives pf the study were to analyze the performance of a chest computed tomography (CT) AI quantitative algorithm for determining the risk of mortality/mechanical ventilation (MV) in hospitalized COVID-19 patients and explore a prognostic multivariate model in a tertiary-care center in Mexico City. METHODS: Chest CT images of 166 COVID-19 patients hospitalized from April 1 to 20, 2020, were retrospectively analyzed using AI algorithm software. Data were collected from their medical records. We analyzed the diagnostic yield of the relevant CT variables using the area under the ROC curve (area under the curve [AUC]). Optimal thresholds were obtained using the Youden index. We proposed a predictive logistic model for each outcome based on CT AI measures and predetermined laboratory and clinical characteristics. RESULTS: The highest diagnostic yield of the assessed CT variables for mortality was the percentage of total opacity (threshold >51%; AUC = 0.88, sensitivity = 74%, and specificity = 91%). The AUC of the CT severity score (threshold > 12.5) was 0.88 for MV (sensitivity = 65% and specificity = 92%). The proposed prognostic models include the percentage of opacity and lactate dehydrogenase level for mortality and troponin I and CT severity score for MV requirement. CONCLUSION: The AI-calculated CT severity score and total opacity percentage showed good diagnostic accuracy for mortality and met MV criteria. The proposed prognostic models using biochemical variables and imaging data measured by AI on chest CT showed good risk classification in our population of hospitalized COVID-19 patients.
