How Machine Learning Experts and Physicians at Johns Hopkins All Children’s are Working Toward a Healthier Future for All Kids
Experts in machine learning and artificial intelligence at Johns Hopkins All Children’s are working with physicians to harness the power of data & analytics.
One of the tenets of Johns Hopkins Medicine is to treat the patient, not the disease. It is a philosophy that traces to William Osler, one of four founding professors of the Johns Hopkins University School of Medicine in Baltimore more than 100 years ago.
Modern medicine increasingly is adopting tools that focus on precision medicine – which takes into account not just the disease a person is diagnosed with but also how his or her individual genetics may impact their condition – which builds upon Osler’s legacy.
At Johns Hopkins All Children's Hospital in St. Petersburg, Florida, experts in machine learning and artificial intelligence are working with physicians in a range of medical specialties on new tools that aim to continuously improve patient outcomes by harnessing the power of data and analytics.
The team recently published four papers in medical journals and has presented their work at national conferences, to broaden the reach of their discoveries. The team often collaborates not only with clinical and research experts at Johns Hopkins All Children’s, but also with colleagues at other leading children’s hospitals across North America.
Machine learning and artificial intelligence have been used to create models that can help care providers identify patients who are at risk for certain medical issues. But these have been primarily for adult patients.
While the same concept can be applied to pediatric care, the models themselves should not be. Because pediatric health experts know that children aren’t just small adults (differing in terms of physiology, developmental stages, types of chronic conditions and acute illnesses, etc.), pediatric care requires pediatric-centered models.
A Powerful Tool
Machine learning is a type of artificial intelligence in which a computer is given the capability to receive information and learn and improve on its knowledge of the topic without additional programming.
An example of this work is detailed in a June 2020 edition of the journal Anesthesia-Analgesia, in a paper entitled “Machine Learning Applied to Registry Data: Development of a Patient-Specific Prediction Model for Blood Transfusion Requirements During Craniofacial Surgery Using the Pediatric Craniofacial Perioperative Registry Dataset.”
The Johns Hopkins All Children’s team included:
- Ali Jalali, Ph.D., data scientist
- Hannah Lonsdale, MBChB, FRCA, MFCI, research associate
- Lillian V. Zamora, M.D., pediatric anesthesiologist
- Luis Ahumada, MSCS, Ph.D., director of the Machine Learning and Predictive Analytics Unit
- Anh Thy H. Nguyen, MSPH, biostatistician
- Mohamed Rehman, M.D., chair of the Department of Anesthesia and Pain Medicine
- Allison M. Fernandez, M.D., M.B.A., pediatric anesthesiologist
Collaborators beyond Johns Hopkins All Children’s were:
- James Fackler, M.D., associate professor of anesthesiology and critical care medicine and pediatrics at the Johns Hopkins University School of Medicine
- Paul A. Stricker, anesthesiologist in the Department of Anesthesiology and Critical Care Medicine and associate division chief for Academic Affairs, General Anesthesiology, at The Children's Hospital of Philadelphia
- Pediatric Craniofacial Collaborative Group, a group of providers from institutions interested in the care of children with craniofacial defects such as craniosynostosis
The team used data from pediatric patients who have undergone craniofacial surgery (which is done to correct congenital and acquired abnormalities of the head and neck), because as the study notes, “Currently, there are no consensus recommendations for guiding blood conservation or transfusion in this patient population.”
The model would help to identify previously unknown factors that could help physicians determine earlier which children may be more at-risk for blood loss during surgery, so they can take appropriate steps to address this more quickly.
“Machine learning helped us identify possible factors that the anesthesiologist could influence on the day of surgery that have the potential to improve patient outcomes,” Fernandez says. “The development and further testing of the prediction model is helpful so that physicians can make more informed decisions in the future regarding intraoperative blood management.”
The eventual goal would be to put such a model into practice in the clinical setting, giving care teams another tool that would complement their clinical expertise and aid in decision-making. But more immediate next steps for this project would involve continued testing of the model with data from additional institutions.
For this initial project, the team used data from a registry comprised of information from more than 30 hospitals around the country. Using data from multiple institutions is key because even when research centers that have access to a large amount of data create a model and that model is put into practice at another institution, the model often doesn’t translate. Using data from multiple institutions allows the team to work with data from patients across a wide range of demographics and outcomes, providing insights that are more representative of the broader population of patients.
On this and other projects, the team at Johns Hopkins All Children’s has worked with experts at institutions like Children’s Hospital of Philadelphia, SickKids hospital in Toronto, and Johns Hopkins Children’s Center in Baltimore, harnessing each institution’s clinical expertise and building on the Johns Hopkins All Children’s team’s expertise in machine learning.
“A tool like this – when advanced through additional phases of applied research – will not make decisions about a patient’s care on its own, but it will enable physicians to make even better and more-informed decisions,” Ahumada says. “The physician and the model together will be powerful.”
“Artificial Intelligence will not replace excellent patient care teams, but it will provide tools to validate and improve their patients’ care,” Rehman says. “Patients will benefit when these models are thoughtfully and strategically integrated into patient care within and across health systems.”
A team at the hospital has also been working on a study investigating key factors associated with hospital readmission in children with respiratory conditions, such as asthma or complications from bronchiolitis (a lung infection that causes inflammation and congestion). Respiratory conditions were targeted by the study because children with these conditions have the second highest rates of readmission, only surpassed by children with cancer. Patients were considered “readmitted” to the hospital if they return within seven days of being discharged.
The team includes pediatric hospitalists John Morrison, M.D., Ph.D., and Paola Dees, M.D.; data scientists Ali Jalali, Ph.D. and Luis Ahumada, MSCS, Ph.D.; and pediatric hospital medicine fellow Brittany Casey, M.D. Dees presented on this innovative work at the Machine Learning in Healthcare virtual conference in August.
Research like this, if it is able to demonstrate a safe reduction in the number of children who are readmitted after leaving the hospital, has the ultimate goals of both improving care and decreasing health care costs.
Building an effective model takes a wide range of expertise. The team engaged physicians, nurses, respiratory therapists, case managers, and others involved in the patients’ care to offer their nuanced perspectives on what can contribute to readmission rates.
“From there, we evaluated the data that we had available, matching the important clinical factors to what the data scientists could capture,” says Morrison, who is also assistant professor in the Department of Pediatrics at Johns Hopkins University School of Medicine and the director of the Center for Training, Education, Engagement and Mentorship within the Johns Hopkins All Children’s Institute for Clinical and Translational Research. “We’ve worked through a lot of different iterations of the model to determine what data matters most to clinicians, to develop a tool that they would eventually be able to use as part of their daily rounds in the inpatient setting.”
So far, as it has been tested, the model has worked differently from how the team first hypothesized. Rather than identifying the patients most at-risk for readmission, evidence so far suggests it may more accurately identify the patients least at-risk for readmission.
The project has given the team the opportunity to work on the cutting edge of machine learning in a way that supports the hospital’s mission and contributes to the field. “Machine learning has proven in so many different arenas to be a powerful tool to drive innovative solutions in health care,” Dees says.
Robust data collection has long been a key component of the hospital’s research work, and is an important factor in machine learning. The hospital has been collecting high-frequency vital sign data from the intensive care units for a number of years, and has been continually expanding in this arena, using the newly developed Precision Medicine Analytics Platform (PMAP). For example, within the last few years, the hospital began collecting continuous vital sign data from intensive care units, using state-of-the-art monitors to collect nearly 1,000 continuous data inputs from an individual patient. Previously data were collected every 45 seconds.
“As part of the hospital’s long-term data collection and analysis strategy for precision medicine projects, we are building a framework that will allow us to develop novel machine learning algorithms for specific conditions or diseases. This will eventually help our clinicians learn even more meaningful insights about our patients and their conditions in near-real time,” Ahumada says.
Other recent publications and presentations include:
- “Deep Anesthesia: Toward Improving Post-Operative Anesthesia Complications Using Deep Learning,” presented at the Machine Learning for Healthcare conference Aug. 7-8.
- “Machine Learning and Artificial Intelligence in Pediatric Research: Current State, Future Prospects, and Examples in Perioperative and Critical Care,” published June 2020 in The Journal of Pediatrics, by Hannah Lonsdale, MBChB; Ali Jalali, MSME, Ph.D.; Luis Ahumada, MSCS, Ph.D.; and Clyde Matava, M.D.
- “Artificial Intelligence in Anesthesiology: Hype, Hope, and Hurdles,” published May 2020 in Anesthesia-Analgesia, by Hannah Lonsdale, MBChB; Ali Jalali, MSME, Ph.D.; Jorge A. Gálvez, M.D., MBI; Luis Ahumada, MSCS, Ph.D.; and Allan F. Simpao, M.D., MBI.
- “Artificial intelligence, machine learning and the pediatric airway,” published December 2019 in Pediatric Anesthesia, by Clyde Matava, M.D., Evelina Pankiv, Luis Ahumada, MSCS, Ph.D., Benjamin Weingarten, and Allan F. Simpao, M.D., MBI.