A team of researchers from UTHealth Houston School of Biomedical Informatics (SBMI) co-authored an article that was published in the Lancet Digital Health in late April. The group includes SBMI students, staff, faculty, and alumni, who focused on developing “accurate and transferrable predictive models of outcomes on hospital admission for patients with COVID-19.”
Assistant Professor and Alumna Laila Rasmy Bekhet, PhD, ‘21 and Associate Professor Degui Zhi, PhD conceptualized this research in 2020 during the onset of the COVID-19 pandemic. They recognized that accurate COVID-19 patient prognoses prediction was a common concern for clinicians, health care researchers, and data scientists. To remedy that issue, the researchers built models called recurrent neural network-based models for COVID-19 outcome prediction, or CovRNN for short.
“We developed a method that can be used to train accurate and generalizable AI models,” stated Bekhet. “By using existing electronic health record (EHR) data, the CovRNN can predict which COVID-19 patients will have severe cases that may need access to hospital beds or mechanical ventilators, or are at high-risk of mortality.”
CovRNN was designed to forecast three different outcomes: in-hospital mortality, mechanical ventilation need, and a hospital stay lasting more than seven days. The deep learning-based models receive relevant big data and are trained to achieve state-of-the-art prediction accuracy. This advanced use of artificial intelligence significantly influences how clinicians can treat patients who have COVID-19 as the models provide physicians with keen insights into their care decisions.
SBMI MS student and McGovern Medical School at UTHealth Houston Assistant Professor Masayuki Nigo, MD is a frontline infectious disease physician at Memorial Hermann and part of the research team. “We have experienced multiple patients with COVID-19 who rapidly deteriorated after hospitalization. It has been challenging for frontline physicians to predict the course at the time of admission,” said Nigo. “However, our model offers clinicians empirical predictions that are very precise. For example, it can predict the need for mechanical ventilation with approximately 93% accuracy which allows physicians to decide the appropriate resource allocation and the disposition of patients.”
Not only is CovRNN an accurate model, but it is also a generalizable one that can serve countless hospitals across the country. “Unlike most existing models that rely heavily on data preprocessing, CovRNN takes the EHR data directly, almost in their native form. This design makes CovRNN very generalizable,” stated Bekhet.
In order to serve a wide range of health care facilities, the team needed to train CovRNN by using a large data set collected from a cross-section of U.S. hospitals. In April of 2020, Cerner provided academic research centers with complimentary access to critical, de-identified COVID-19 patient data to help fight the pandemic. UTHealth Houston researchers where fortunate to receive that data, which also included quarterly updates. CovRNN was extensively validated using data from Cerner as well as Optum’s de-identified COVID-19 electronic health record dataset.
Once the data was acquired, Bekhet and Zhi recruited an interdisciplinary team comprised of various members from the SBMI Community as having a diverse group of researchers maximizes the efficacy of the models. “A collaborative approach enriches the quality of the predictive models as each individual’s unique expertise brings valuable contributions to the study,” said Zhi.
The team included close to one dozen experts. Dr. Nigo described the current needs for the models from the practice perspective and MS Alumnus Bijun Sai Kannadath, MBBS, MS, ’18 who works at the University of Arizona College of Medicine, contributed in a similar fashion while providing practical feedback. Assistant Professor Angela Ross, DNP assessed both the bias of the study and the transparency of the reported methods and results. SBMI researchers Ziqian Xie, PhD and Yujia Zhou, MS, who is also a 2017 alumna of SBMI, offered technical and data support. Several students, including PhD student Bingyu Mao, MA and Dual MS/MPH student Khush Patel, MD, ran experiments and created evaluation results and visualizations. Professor and Associate Dean for Innovation Hua Xu, PhD supported the study and provided insightful recommendations to enhance the models.
While the U.S. is described as being out of the pandemic phase, it is still critically important to continue developing best practices for treating this highly mutable virus. The team also shared the source code and trained model at GitHub to enable global efforts to fight this virus. Bekhet and Zhi plan to “continue collaborating with UTHealth Houston researchers while evaluating the models on more recent COVID-19 patient data.”
The researchers have already presented their findings at recent American Medical Informatics Association (AMIA) conferences to share this knowledge with the informatics community. Follow-up implementation work on CovRNN will also be presented at the AMIA 2022 Clinical Informatics Conference, which will be hosted in Houston in later this month.