Introduction: Recent studies have highlighted adverse outcomes of fluid overload in critically ill patients. Therefore, its early recognition is essential for the management of these patients.
Objectives: Our aim was to propose a deep learning (DL) model using data from noninvasive chest X‑ray (CXR) imaging associated with the fluid overload status.
Patients and methods: We collected data from the Medical Information Mart for Intensive Care IV (MIMIC‑IV, v. 1.0) and MIMIC Chest X‑Ray (v. 2.0.0) databases for modeling, and from our hospital database for testing. The extravascular lung water index (ELWI) greater than 10 ml/kg and the global end-diastolic volume index (GEDI) greater than 700 ml/m2 were used as threshold values for the fluid overload status. A DL model with a transfer learning strategy was proposed to predict the fluid overload status based on CXR images, and compared with clinical and semantic label models. Additionally, a visualization technique was adopted to determine the important areas of features in the input images.
Results: The DL model showed a relatively strong performance for predicting the ELWI (area under the curve [AUC] = 0.896; 95% CI, 0.819-0.972 and AUC = 0.718; 95% CI, 0.594-0.822, respectively) and the GEDI status (AUC = 0.814; 95% CI, 0.699-0.930 and AUC = 0.649; 95% CI, 0.510-0.787, respectively) in both the primary and the test cohort. The performance was better than that of the clinical and semantic label models.
Conclusions: As CXR is routinely used in the intensive care unit, a simple, fast, low‑cost, and noninvasive DL model based on this modality can be regarded as an effective supplementary tool for identifying fluid overload, and should be widely adopted in the clinical setting, especially when invasive hemodynamic monitoring is not available.