CDS Members Continue Evolve AI That Predicts COVID-19 Patient Deterioration

Last year, CDS PhD student Artie (Yiqiu) Shen and colleagues including Assistant Professor and Emerging Scholar of Computer Engineering at NYU Abu Dhabi Farah Shamont, CDS PhD students Nan Wu and Aakash Kaku, CDS alum Jungkyu Park (now a PhD student at NYU School of Medicine), and several others, began developing an AI system that could potentially help to predict the deterioration of emergency room COVID-19 patients. The project was under the supervision of CDS Assistant Professor of Mathematics and Data Science Carlos Fernandez-Granda and CDS affiliated professor Krzysztof J. Geras, who also is Assistant Professor in the Department of Radiology at the NYU Grossman School of Medicine. The paper, “an artificial intelligence system for predicting the deterioration of COVID-19 patients in the emergency department”, was recently published in npj Digital Medicine. It has also been featured in online science news outlets EurekAlert! and SciTechDaily.

Overview of the AI system and the architecture of its deep learning component from “an artificial intelligence system for predicting the deterioration of COVID-19 patients in the emergency department”.

CDS featured the project earlier in its journey in an initial blog post in August of last year. The project came into conception in April 2020 when Krzysztof called for volunteers to work on a chest X-ray dataset NYU Langone collected from COVID patients. Following a discussion with what later became the project’s team, there was a realization that they could potentially use the dataset to develop a model to assist in determining which COVID-19 patients were more likely to experience clinical deterioration.

The team developed the AI system using a dataset of 19,957 chest X-ray exams collected from 4,722 patients who tested positive for COVID-19 (via the polymerase chain reaction, PCR, test) at NYU Langone Health between March 3, 2020 and May 13, 2020. The system leverages deep convolutional neural networks to perform risk evaluation from chest X-ray images. They designed their imaging-based classifier in particular based on the Globally-Aware Multiple Instance Classifier (GMIC), denoted as COVID-GMIC. The AI also learns from routinely collected clinical variables using a gradient boosting model (GBM), denoted as COVID-GBM. (Gradient boosting is a machine learning technique used to address regression analysis and classification problems.) They combined output predictions from COVID-GMIC and COVID-GBM to predict each patient’s overall risk of deterioration over varying time horizons, ranging from 24 to 96 hours. The system also included another model that predicts how the risk of deterioration is expected to evolve over time by computing deterioration risk curves (DRC) denoted as COVID-GMIC-DRC.

During the first wave of COVID-19, the team silently deployed a preliminary version of the AI at NYU Langone Health to demonstrate that it can produce predictions in real-time within a real clinical setting. Ultimately, the results strongly suggested that the AI is a viable tool to inform triage of COVID-19 patients.

To read the paper in its entirety, please visit its npj Digital Medicine page. The team’s code and trained models are available on their Github profile.

By Ashley C. McDonald

Official account of the Center for Data Science at NYU, home of the Undergraduate, Master’s, and Ph.D. programs in Data Science.