Artificial Intelligence for Clinical Interpretation of Bedside Chest Radiographs.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Intensive Care Medicine

Background:

• Bedside chest radiography is a frequent imaging modality for intensive care unit (ICU) patients.
• Accurate interpretation is crucial for timely clinical decision-making.
• Existing interpretation methods can be time-consuming and subject to variability.

Purpose of the Study:

• To assess the diagnostic performance of a neural network (NN) model.
• The NN was trained on structured, semiquantitative radiologic reports of bedside chest radiographs.
• To evaluate the NN's ability to identify key thoracic abnormalities.

Main Methods:

• Retrospective single-center study involving over 193,000 chest radiographs from 45,000+ ICU patients (2009-2020).
• Radiographs were semiquantitatively rated by 98 radiologists for disease severity.
• A neural network was trained to detect cardiomegaly, pulmonary congestion, pleural effusion, opacities, and atelectasis.

Main Results:

• The neural network achieved higher agreement (κ = 0.86) with expert panel consensus than individual radiologists (κ = 0.81-0.84).
• Non-radiologist physicians showed improved interpretations (κ = 0.87 aided vs. 0.79 unaided) when assisted by the NN's preliminary readings (P < .001).
• The model demonstrated robust performance in identifying common ICU-related thoracic pathologies.

Conclusions:

• A neural network trained on structured radiologic reports can enhance diagnostic accuracy for bedside chest radiographs.
• AI-assisted interpretation significantly improves the performance of non-radiologist physicians in the ICU.
• This technology holds promise for improving patient care through more efficient and accurate image interpretation.