Establishment of seven lung ultrasound phenotypes: a retrospective observational study of an LUS registry

Abstract Background Lung phenotypes have been extensively utilized to assess lung injury and guide precise treatment. However, current phenotypic evaluation methods rely on CT scans and other techniques. Although lung ultrasound (LUS) is widely employed in critically ill patients, there is a lack of comprehensive and systematic identification of LUS phenotypes based on clinical data and assessment of their clinical value. Methods Our study was based on a retrospective database. A total of 821 patients were included from September 2019 to October 2020. 1902 LUS examinations were performed in this period. Using a dataset of 55 LUS examinations focused on lung injuries, a group of experts developed an algorithm for classifying LUS phenotypes based on clinical practice, expert experience, and lecture review. This algorithm underwent validation and refinement with an additional 140 LUS images, leading to five iterative revisions and the generation of 1902 distinct LUS phenotypes. Subsequently, a validated machine learning algorithm was applied to these phenotypes. To assess the algorithm’s effectiveness, experts manually verified 30% of the phenotypes, confirming its efficacy. Using K-means cluster analysis and expert image selection from the 1902 LUS examinations, we established seven distinct LUS phenotypes. To further explore the diagnostic value of these phenotypes for clinical diagnosis, we investigated their auxiliary diagnostic capabilities. Results A total of 1902 LUS phenotypes were tested by randomly selecting 30% to verify the phenotypic accuracy. With the 1902 LUS phenotypes, seven lung ultrasound phenotypes were established through statistical K-means cluster analysis and expert screening. The acute respiratory distress syndrome (ARDS) exhibited gravity-dependent phenotypes, while the cardiogenic pulmonary edema exhibited nongravity phenotypes. The baseline characteristics of the 821 patients included age (66.14 ± 11.76), sex (560/321), heart rate (96.99 ± 23.75), mean arterial pressure (86.5 ± 13.57), Acute Physiology and Chronic Health Evaluation II (APACHE II)score (20.49 ± 8.60), and duration of ICU stay (24.50 ± 26.22); among the 821 patients, 78.8% were cured. In severe pneumonia patients, the gravity-dependent phenotype accounted for 42% of the cases, whereas the nongravity-dependent phenotype constituted 58%. These findings highlight the value of applying different LUS phenotypes in various diagnoses. Conclusions Seven sets of LUS phenotypes were established through machine learning analysis of retrospective data; these phenotypes could represent the typical characteristics of patients with different types of critical illness.