**Authors and Affiliations**
Malins, J.G., Luoma, C.E., Anisuzzaman, D.M., Jackson, J.I., Lopez-Jimenez, F., Attia, Z., Pislaru, S.V., Kane, G.C.
Department of Cardiovascular Medicine, Mayo Clinic, Rochester, MN

**Background**
In recent years, lung ultrasound has emerged as a portable, safe, and available means of assessing various lung pathologies, such as fibrosis and edema. Yet, existing approaches are mostly qualitative and subjective and require advanced training; as a result, there is a need for standardized and reproducible approaches for scoring and categorizing lung ultrasound images. Artificial intelligence (AI) techniques have the potential to help achieve this aim.

**Methods**
The current dataset consists of lung ultrasound clips collected during transthoracic echocardiography exams performed either entirely at rest or before and after exercise stress. These clips were divided into datasets for model training (963 clips), internal validation (380 clips), and testing on a holdout dataset (417 clips). Frames from lung ultrasound clips were provided as input to two AI models: one which classifies an input clip as either normal or abnormal, and one which provides a continuous lung density score.

**Results**
Performance for both models was strong on the holdout testing dataset, with the normal/abnormal classification model achieving an AUC of 0.909 (95% confidence interval of 0.880 to 0.938), and the lung density regression model achieving a coefficient of determination (R2) of 0.496.

**Conclusion**
Although further refinement and external validation is required for these algorithms, future deployment could optimize imaging resources by assisting providers in making clinical decisions regarding the management of heart failure patients and the potential need for additional lung imaging such as chest computed tomography.



**Authors and Affiliations**
_Malins, J.G., Luoma, C.E., Anisuzzaman, D.M., Jackson, J.I., Lopez-Jimenez, F., Attia, Z., Pislaru, S.V., Kane, G.C._
_Department of Cardiovascular Medicine, Mayo Clinic, Rochester, MN_

**Background**
In recent years, lung ultrasound has emerged as a portable, safe, and available means of assessing various lung pathologies, such as fibrosis and edema. Yet, existing approaches are mostly qualitative and subjective and require advanced training; as a result, there is a need for standardized and reproducible approaches for scoring and categorizing lung ultrasound images. Artificial intelligence (AI) techniques have the potential to help achieve this aim.

**Methods**
The current dataset consists of lung ultrasound clips collected during transthoracic echocardiography exams performed either entirely at rest or before and after exercise stress. These clips were divided into datasets for model training (963 clips), internal validation (380 clips), and testing on a holdout dataset (417 clips). Frames from lung ultrasound clips were provided as input to two AI models: one which classifies an input clip as either normal or abnormal, and one which provides a continuous lung density score.

**Results**
Performance for both models was strong on the holdout testing dataset, with the normal/abnormal classification model achieving an AUC of 0.909 (95% confidence interval of 0.880 to 0.938), and the lung density regression model achieving a coefficient of determination (_R^2^_) of 0.496.

**Conclusion**
Although further refinement and external validation is required for these algorithms, future deployment could optimize imaging resources by assisting providers in making clinical decisions regarding the management of heart failure patients and the potential need for additional lung imaging such as chest computed tomography.



Using AI to detect abnormal lung pathologies and estimate lung density scores from lung ultrasound video clips collected during echocardiographic examinations