Detecting abnormal excessive buildup of fluid in the lungs, or pulmonary edema, is crucial in preventing conditions such as heart failure, kidney failure, and acute respiratory distress syndrome (ARDS). Most existing methods for measuring fluid accumulation in lungs are either expensive and invasive, thus unsuitable for continuous monitoring, or inaccurate and unreliable. To provide continuous and non-invasive monitoring of lung water status, Hawaii Advanced Wireless Technologies Institute (HAWTI) invented the Cardio-Pulmonary Stethoscope (CPS), a low-cost device with chest patch radio frequency (RF) sensors that was proven to be able to detect heart rate, respiration rate, and changes in lung water content from a single RF measurement. The CPS measurement procedure and the accuracy of results have been verified in a National Institute of Health (NIH) sponsored clinical trial conducted in collaboration with The Queen’s Medical Center in Honolulu.

This thesis presents recent advances in expanding the capability of the CPS for assessing lung water status, in addition to monitoring the change in lung water, using artificial intelligence (AI). An important first step in our AI pipeline is to build a database of a diverse patient population. To this end, we utilize an NIH dataset consisting of CT-scans of patients of various genders, ages, and body fat compositions. We then develop an automatic workflow that reads the CT-scans and creates 3-D models for high-fidelity simulation in Ansys High Frequency Structure Simulator (HFSS). From HFSS, we obtain scattering parameters (S-parameters) measured by the CPS at various lung water levels. Compared to data collection from clinical trials, this “Virtual Clinical Trial” approach is low-cost, less time-consuming, and risk-free.

Using the database we built, we develop AI models which use the patient metadata, namely gender, age, fat thickness, and S-parameters from the CPS as input, and output its assessment of the lung water status (i.e., normal, edematous, and severely edematous statuses). For a cohort of over 200 diverse individuals, our AI models achieve above 70% accuracy in assessing the lung water status. Furthermore, our AI models are interpretable and simple to explain.