Artificial Intelligence Neural Network Consistently Interprets Lung Ultrasound Artifacts in Hospitalized Patients: A Prospective Observational Study

Background Interpretation of lung ultrasound artifacts by clinicians can be inconsistent. Artificial intelligence (AI) may perform this task more consistently. Research Question Can AI characterize lung ultrasound artifacts similarly to humans, and can AI interpretation be corroborated by clinical data? Study Design and Methods Lung sonograms (n=665) from a convenience sample of 172 subjects were prospectively obtained using a pre-specified protocol and matched to clinical and radiographic data. Three investigators scored sonograms for A-lines and B-lines. AI was trained using 142 subjects and then tested on a separate dataset of 30 patients. Three radiologists scored similar anatomic regions of contemporary radiographs for interstitial and alveolar infiltrates to corroborate sonographic findings. The ratio of oxyhemoglobin saturation:fraction of inspired oxygen (S/F) was also used for comparison. The primary outcome was the intraclass correlation coefficient (ICC) between the median investigator scoring of artifacts and AI interpretation. Results In the test set, the correlation between the median investigator score and the AI score was moderate to good for A lines (ICC 0.73, 95% CI [0.53-0.89]), and moderate for B lines (ICC 0.66, 95% CI [0.55-0.75]). The degree of variability between the AI score and the median investigator score for each video was similar to the variability between each investigator’s score and the median score. The correlation among radiologists was moderate (ICC 0.59, 95% CI [0.52-0.82]) for interstitial infiltrates and poor for alveolar infiltrates (ICC 0.33, 95% CI [0.07-0.58]). There was a statistically significant correlation between AI scored B-lines and the degree of interstitial opacities for five of six lung zones. Neither AI nor human-scored artifacts were consistently associated with S/F. Interpretation Using a limited dataset, we showed that AI can interpret lung ultrasound A-lines and B-lines in a fashion that could be clinically useful.