Senior Design Projects

ECS193 A/B Winter & Spring 2023

Dynamic AI for Critical Congenital Heart Disease Screening

Email **********
Name Chen-Nee Chuah
Affiliation Electrical & Computer Engineering

Project's details

Project title Dynamic AI for Critical Congenital Heart Disease Screening
Background Clinical problem: Congenital heart disease is the most common birth defect. Critical congenital heart defects (CCHD), or those requiring surgical or other invasive intervention in the first month of life, make up 25% of these defects, and account for up to 10% of US infant deaths. This high mortality rate is in part due to difficulty diagnosing CCHD in asymptomatic infants. Currently all US newborns are screened for CCHD with oxygen saturation (SpO2) before discharge from the hospital. SpO2 screening since has improved CCHD detection, but nearly 900 US newborns (115 in California) with CCHD continue to be missed annually. The types of defects that are detected well with SpO2 screening are those with impaired oxygenation, or cyanosis. The types of defects that are most commonly missed by SpO2 screening are defects that have systemic obstruction such as a narrowing or interruption of the aorta (coarctation of the aorta or interrupted aorta arch). Late diagnosis of these lesions is particularly detrimental because the newborns present critically ill when intervention may no longer prevent mortality or morbidity. In fact, some continue to be diagnosed only at autopsy. The addition of other pulse oximetry features such as heart rate, perfusion index and pulse transit delay time, have shown promise to improve CCHD detection. However, as the baby ages through the few days of age, their physiology and thus pulse oximetry features change. Thus, a model that allows for dynamic data incorporation with non-continuous data points may be a more optimal screening test. Furthermore, most of the pulse oximeters do not take into account of the skin color tone. The measurement results could deviate from the actual SpO2 value. One approach to overcome this is to supplement the pulse oximeters with a colorimeter. This could provide a more comprehensive and accurate SpO2 measurement.
Description Available dataset: We currently have pulse oximetry data on over 400 newborns with and without CCHD including varying types of CCHD and non-critical heart defects. Pulse oximetry data includes SpO2, perfusion index, heart rate, pulse transmit. Data was collected up to 3 times for each patient (0-24, 24-48 and >48hr). We currently have 17 patients with measurements up to about 2 weeks of age.

Available code: We already have a working ML-pipeline (based on logistic regression and random forest) to automatically remove artifacts from collected waveforms and to perform off-line CCHD screening. The dynamic model can be built upon this prior work.
Deliverable 1. Develop and implement Android application capable of dual-bluetooth data collection for mobile end-to-end analysis. Incorporate collection into existing Android CCHD data analytic pipeline.
2. Develop dynamic model that incorporates new non-continuous pulse oximetry measurements for a patient and adjusts a patient’s predicted outcome (CCHD vs non-critical CHD vs healthy).
3. Develop an accurate SpO2 measurement by combining pulse oximeters, colorimeters and computational algorithms.
4. Explore inexpensive interventions for drop in PPG accuracy in dark skin (such as camera based). Incorporate intervention into existing CCHD data analytic pipeline.
Skill set desirable N/A
Phone number **********
Client time availability 30-60 min weekly or more
IP requirement Open source project
Attachment Click here
Selected No
Team members N/A
TA N/A