An automated system to interpret echocardiography to predict adverse outcomes in patients with right ventricular dysfunction in daily hospital practice

Project Summary Right ventricle (RV) dysfunction is a common and complex form of pediatric heart disease. It is also a common contributor to morbidity and mortality for patients with congenital heart diseases (CHD). Due to the complex geometry of the RV and its relative adaptability to changing physiologic conditions, RV dysfunction is poorly understood and difficult to characterize precisely and accurately, thus diagnosis is often delayed. The most common diagnosis tool is echocardiograms. Manual review of echocardiograms is time consuming, however. Furthermore, there might be uncovered echocardiogram patterns associated with RV dysfunctions. In adult studies, machine learning models (MLM) have been successfully implemented to assess RV functions by echocardiograms. We hypothesize that applying novel MLM to pediatric echocardiograms will allow us to improve the accuracy and reliability of assessment, as well as identify novel markers of RV dysfunction. We propose to develop an automated tool to generate a RV health score to identify RV dysfunction and predict the development and time of adverse outcomes including heart failure, heart and/or lung transplantation, and death. The automated tool will constitute an early warning system module, which will be deployed onto a big-data-based risk prediction platform developed by our small business. The study has three specific aims. First, we will extract echocardiograms and structured electronic medical records from the Stanford Children’s Hospital. Cohorts of children with normal or abnormal RV will be constructed. Second, MLM will be developed and validated to 1) predict the presence of RV dysfunction and the probability of adverse outcomes, and 2) predict the rate of progression to adverse outcomes. A deep learning-based workflow will be established to take input of pediatric echocardiogram and clinical data and generate predictions. Third, we will integrate the models developed in Aim #2 into the HBI Spotlight Solutions. The Spotlight Solutions include a healthcare surveillance platform with high-capacity data infrastructure and risk engines to offer AI solutions to care facilities participating the Healthix, the largest public health information exchange network in the US. This will prepare our algorithms for further clinical validation in other cohorts.Project narrative The machine learning models developed in this study will be applicable to any infant, child or adolescent undergoing an echocardiogram, particularly those requiring serial exams to assess changes from disease progression or treatment effect. It can be adopted by pediatric clinicians to assess cardiac function with a high degree of accuracy and confidence, even for those who have little training in echocardiography, thus improving access to healthcare and potentially reducing costs.