External validation of ARDS mortality prediction model using combined clinical data and chest radiograph analysis using deep neural networks. The advantge of a scalable predictive model
作者
Ana Santos Dornellas Cysneiros,Tiago Galvão,Nuno Domingues,Rachael A.F. Cusack,Pompeo Costantino Kouroupis,Luís Filipe Nunes Bento,Ignacio Martín‐Loeches
出处
期刊:The European respiratory journal [European Respiratory Society] 日期:2025-09-27卷期号:: PA4194-PA4194
标识
DOI:10.1183/13993003.congress-2025.pa4194
摘要
Background: ARDS remains a leading cause for intensive care (ICU) admission with high mortality particularly for severe ARDS. Material and Methods: An initial retrospective analysis of 110 COVID ARDS (C-ARDS) patients was used to create a binary logistic regression model of mortality prediction based on PaO2/FiO2 ratio (P/F) at day one and day three of invasive mechanical ventilation, convoluted neural network (CNN) extracted chest x-ray features and age. Early model performance evaluation on test data (23 patients out of 110) indicated an area under the receiver operating characteristic (ROC) curve of 0.862 CI (0.654-0.969). External validation was conducted on patients with ARDS from an external ICU population. This included a total of 142 ARDS patients, 66 with C-ARDS and 76 with other forms of primary ARDS. Results: Using the original training set of 87 patients from the original ICU and 66 C-ARDS patients from the external ICU as the external test AUC was 0.741 95% CI ( 0.619-0.841) however when using from the original training set but with the 76 non COVID ARDS patients the AUC was only 0.611 95% CI (0.493-0.721).A new model was developed using the total 110 patients from the original model (training set and test set) and using as the test set the 66 C-ARDS patients from the externa ICU. AUC was 0.855 95% CI (0.747-0.930). Conclusion: The mortality prediction model was only valid for C-ARDS which indicates that C-ARDS is not interchangeable with other forms of primary ARDS. Likely because imaging patterns are different. Importantly the mortality prediction model was scalable.