Heart failure is the leading cause of hospitalization in the US, with high readmission and mortality rates. In the US alone, there are 1 million hospital visits related to heart failure every year. Most of the acute heart failure patients present with pulmonary edema (fluid overload in their lungs) and exhibit heterogeneous responses to treatment. This heterogeneity precludes effective treatment and leads to long hospital stays. Inadequate fluid management leads to high readmission rates.
The treatment success in acute heart failure cases depends crucially on effective management of patient fluid status, which in turn requires accurate pulmonary edema assessment and prior response analysis. Clinicians order chest x-rays for heart failure patients every day in routine clinical practice, but their assessments of the severity of pulmonary edema based on the x-ray images are inconsistent across practitioners and even across different reads by the same physician. Other surrogates for patient fluid status and response to treatment are either noisy (such as body weight and urine output) or require an invasive procedure (Swan-Ganz catheterization).
To support better clinical decision making for heart failure patients and enable quantitative research on efficacy of treatments, we are developing machine learning algorithms for automatic assessment of pulmonary edema severity from chest x-ray images. Non-invasive and quantitative tracking and monitoring of patient fluid status enables personalized treatment and ultimately shortens hospital stays.
In addition to supporting heart failure treatment, the quantitative assessment of pulmonary edema severity on chest x-ray images is useful throughout clinical medicine. Pulmonary edema is a manifestation of fluid status in sepsis and renal failure, just as it is in heart failure. Managing fluid status is critical in those clinical contexts as well. Automatic and accurate assessment of pulmonary edema in chest x-ray images will support treatment planning of sepsis and other diseases where fluid status is critical. This quantitative phenotyping can also improve clinical trial management where patient fluid status is a biomarker, because the measurements from images are more reproducible and accurate.
Representative chest x-ray images of pulmonary edema: