When members of the Center for Pediatric Data Science and Analytics Methodology in the Johns Hopkins All Children’s Institute for Clinical & Translational Research traveled to Glasgow, Scotland, in July to attend the 44th Annual International Conference of the Institute of Electrical and Electronics Engineers (IEEE) Engineering in Medicine and Biology Society, they presented the results of three important studies aimed at improving the delivery of health care. 

The IEEE conference, themed “Biomedical Engineering transforming the provision of health care — Promoting wellness through personalized and predictable provision at the point of care,” attracted researchers from around the world. Attendees shared their research in areas such as biomedical signal processing, neural and rehabilitation engineering, biomedical sensors and wearable systems, translational engineering at the point of care, computational systems, and biorobotics and biomechanics. 

The Machine Learning and Predictive Analytics Team 

“We find predictive analytics solutions for developing clinical decision-making support systems, for aiding in disease management, tracking hospital readmissions, or gathering information regarding adverse events and preventing them,” explains Luis Ahumada, Ph.D., director of the Center for Pediatric Data Science and Analytics Methodology in the Johns Hopkins All Children’s Institute for Clinical & Translational Research. “Data science techniques can be applied to any type of unknown — whether in the past, present or future.”  

To solve many problems, the data science center’s team often develops algorithms — computerized procedures used for solving a problem or performing computations. They can be designed to discover and visualize patterns, as well as make predictions. 

According to Ahumada, the topics they selected for study and ultimately presented at the conference are important areas of inquiry for the clinicians with whom the data science center’s team works closely. 
For example, the data science center team recently worked with Anthony A. Sochet, M.D., MHSc, of the Pediatric Critical Care Medicine Division at Johns Hopkins All Children’s, to build data-driven artificial intelligence (AI) tools to predict which pediatric patients suffering from bronchiolitis — a common yet serious respiratory infection among children — might respond to non-invasive ventilation, and which patients may need to be intubated.  

ML and AI-based Studies Aim to Improve Health Care 

In Glasgow, the data science center’s team presented three research projects addressing ways to identify patient social needs using natural language processing, predicting which patients may be at-risk for hospital readmission within seven days of previous discharge, and analyzing survival probabilities for pediatric heart transplantation patients. 

Applying natural language processing to identify patient social needs 

“Electronic health records (EHRs) are a rich source of data, but most of the data is in unstructured medical notes making it difficult to identify patient social needs,” says Johns Hopkins All Children’s data scientist Geoffrey Gray, Ph.D. “In this study we applied natural language processing (NLP), a subfield of linguistics and a type of field inquiry that uses computer science and artificial intelligence to analyze large amounts of text data embedded in medical notes to identify patients’ social needs, which are important to both treatment and health care costs.”  

Working with the Johns Hopkins Center for Health Disparities Solutions in Baltimore, Gray, who built the program for using NLP, explained that they used NLP in processing medical records notes to reveal patient social needs in three areas — residential instability, food insecurity and transportation. Algorithms were employed to match the topical keywords to the 1.5 million patient records and more than 65 million medical notes contained in the Johns Hopkins Health System (JHHS) database. 

“Using NLP techniques to extract data from unstructured EHRs can result in the identification of patients who are at-risk and assist providers in focusing their resources on assessing the needs of patients who are medically underserved,” explains Gray. 

The algorithm for residential instability (homelessness) performed best, says Gray, probably because the keyword “homeless” appeared more often in the medical notes. Transportation issues were less likely to be identified. However, the algorithm for food insecurity revealed that 28.97 percent of the patients in the database said they experienced food insecurity. 

Building a machine learning model to predict seven-day hospital readmissions 

According to Ahumada, ML is a subset of AI that uses a set of algorithms to analyze and interpret structured data and can facilitate a “precision medicine” approach to health care problem solving. 

“Hospital readmissions occurring within seven days of hospital discharge represent a burden for the health care system and also serve as a measure for the quality of hospital care,” Ahumada says. “In this study, we developed an ML model to help predict seven-day readmissions. It is technically easier to track 30-day readmissions than to track seven-day readmissions, but the seven-day data is clinically more useful.”  

To conduct this study, the researchers — who included John Morrison, M.D., Ph.D., of the Johns Hopkins All Children’s Division of Hospital Medicine — analyzed four years of EMRs and developed algorithms to build predictive models focusing on a number of features. The features included demographics such as age, sex, race and median income from ZIP code. Payer information, patient history, previous admissions and readmissions were also included. Clinical information included length of hospital stay, medication counts, unique diagnoses and admitting unit. 

The study showed that stratifying risk and integrating predictive models of risk into clinical decisions support systems can present opportunities for modifying the seven-day hospital readmission risk factors. The data they gathered will soon be implemented into a clinical decision support solution for predicting which patients might be at risk for a seven-day hospital readmission, says Ahumada. 

Predicting survival for pediatric heart transplantation patients 

Because of limited organ supply for pediatric heart transplantation, it is important to be able to predict probability of survival for pediatric heart transplantation patients. Accordingly, the data science center team embarked on a research project using ML techniques for predicting pediatric heart transplantation patient survival.  

To conduct their research, the team — which included Awais Ashfaq, MBBS, and Alfred Asante-Korang, M.D., of the Johns Hopkins All Children’s Heart Institute — used data from United Network for Organ Sharing (UNOS), the nation’s largest available registry for organ transplant, and subsequently employed a range of ML tools to build algorithms and predictive models.  

Patients were selected for the model if they received only a heart transplant, if they had not had a previous transplant, and if their transplant had occurred within the previous 10 years. Having been on a ventilator before transplantation and a patient’s functional status were also important criteria.  

The predictive values they found indicated that their model performed in such a way that it could be used as a clinical decision-making support tool to “optimize” the use of scarce organ availability. 

“This research demonstrates that development of high-performance ML models for survival analysis of pediatric heart transplant patients can guide clinical decision making,” says Ahumada. “The highest performing features selected were a combination of both automatic feature selection, as well as expert clinical input from surgeons and cardiologists.” 

The Future of ML and AI in Health Care 

According to Ahumada, IEEE is one of the world’s oldest and largest technical professional organizations, and the IEEE conferences are among the most highly rated.  

Attending IEEE’s society for engineering in medicine and biology conferences is important for the data science center team. Attendance offers not only the opportunity to share their research, but also helps them stay abreast of the latest biomedical engineering research, particularly in the area of ML and AI systems as applied to health care.  

Ahumada adds that the introduction of ML and AI into medical research, and the advent of innovative computer programs and software to develop tools such as clinical decision-making support systems, are “not an attempt to replace your doctor with a computer.” 

“The use of ML and AI can provide additional data points that clinicians can use in their decision making,” Gray says. “For example, our research into patient social needs using NLP to review 65 million medical notes was accomplished in several hours — an impossible task for humans.”  

Ahumada also noted that while the use of ML and AI in clinical decision making is revolutionizing health care, certain “biases” built into some ML, AI and algorithms have been detected. The problem is global and, because biases can lead to health care disparities, the problem has received international attention. Consequently, uncovering and mitigating biases has become an important issue for the AI industry and efforts to reduce bias represent an important task in biomedical engineering.