Lessons from the data: Applying machine learning for clinical decision support

machine learning clinical decision support

Mauricio Santillana, PhD, faculty member in the Computational Health Informatics Program at Boston Children’s Hospital, had an idea as he witnessed the volume of continuous real-time data generated in the pediatric intensive care unit (PICU). He realized that tapping the data on patients’ ever-changing vital signs, with the help of machine-learning algorithms, could support clinical decision-making and predict (and help head off) up-coming health issues.

He started a dialogue with the hospital’s Innovation & Digital Health Accelerator, and now collaborates closely with clinicians in the PICU to create machine-learning algorithms that can help them provide the highest level of care.

“It’s fairly recent that clinicians realized people with backgrounds in math and statistics can be very helpful in a clinical context,” says Santillana“We follow a philosophy of working together to help doctors perform better. That requires developers to come together with clinicians. We want to spread high quality care across the board.”

Data processing in the PICU

Mauricio Santillana clinical decision support
Santillana at a recent Innovators’ Showcase

Clinicians in the PICU often intubate and extubate patients. The decision to do either can be tricky. “An endotracheal tube left in a mechanically ventilated patient for too long can lead to lung damage or infection,” notes Santanilla, who is also affiliated with Harvard Medical School and the Harvard Institute for Applied and Computational Sciences.

Protocols exist to determine the best time to intubate and extubate patients, based on routine checks to determine whether they meet certain criteria. “If the patient passes those tests, the tube can be removed,” Santanilla says. However, extubation is not always successful, requiring re-intubation with often detrimental consequences.

When Santillana learned that the PICU continuously collects patient data, he started to think about applying machine learning to improve the rate of successful extubations. “What if we used a computational algorithm to predict when to extubate a patient, based on respiratory rate, heart rate and other vital signs?”

pediatric ICU clinical decision supportThis question led to a computational algorithm created to predict ideal time for extubation—and ultimately to a pilot study showing an increased rate of successful extubation procedures in the PICU.

Santillana is also applying machine learning techniques to predict how long a patient will stay in the PICU. “We’ve shown that you get a stronger and more accurate prediction of how long a patient will stay in the PICU if you combine vital sign information with demographic information,” he says. “In our early research focused on epidemic surveillance, combining multiple data sources creates a more robust and accurate computational algorithm.”

Diagnosing critical congenital heart defects

Congenital heart defects are among the most fatal birth defects, often requiring immediate intervention. Efficient and accurate diagnosis of congenital heart defects is therefore critical.

Enter Douglas Perrin, PhD, a staff scientist at Boston Children’s Cardiac Surgery Research Center and an instructor in surgery at Harvard Medical School.

“We were curious if we could apply machine learning to diagnostics with cardiac issues that are not compatible with life,” he says. “These are issues that you have to catch within the first 72 hours of life.”

Current diagnostic processes for critical congenital heart defects have shortcomings. Pulse oximetry has less than ideal accuracy, and echocardiography requires significant human expertise to interpret.

clinical decision support echocardiography congenital heart disease

To diagnose defects more efficiently and effectively on two-dimensional echocardiograms, Perrin is creating an image classification framework — “an automated screening technique by which imaging can be escalated to further expert review,” he explains. (The image at left is predicted to be normal.)

In the future, Perrin also wants to look at moving heart images. The current system “classifies disease on a single static frame,” he says. “It’s not looking at motion, which is how a cardiologist would diagnose disease. Video analysis is more cutting-edge, but that’s the next step — to have the Machine Learning system evaluate data like a human would.”

Pushing reference brain images to clinicians’ desktops

Reading and interpreting magnetic resonance (MR) images of the brain poses a special challenge in children. As Richard Robertson, MD, radiologist-in-chief at Boston Children’s Hospital, recently told the Boston Globe, “If you look at the brain, it changes pretty rapidly in the first two to three years [of life]. If can be really hard to decide whether what you’re seeing is the process of normal development or if something has gone wrong.”

Many pediatric MR images are not interpreted by pediatric specialists, and only 3 percent of U.S. radiologists practice in pediatrics. In contrast, a high-volume facility like Boston Children’s can read 30,000 to 50,000 images per day.

brain MRI scans clinical decision support
These MRI images demonstrate age-specific differences in brain myelination, helping clinicians distinguish normal from abnormal.

To help radiologists, Boston Children’s has teamed with GE Healthcare to create clinical decision support tools to increase diagnostic accuracy and reduce time to diagnosis. The initial focus is on helping distinguish abnormal from normal brain development, by creating an application that provides a set of normative reference images for radiologists to compare against a child’s scan.

“If a radiologist had access to normative data at the point of care, appearing alongside the patient’s information and corresponding to the age of the patient they are assessing, they would be able to confidently diagnose and point to next steps,” says Sanjay Prabhu, MBBS, a pediatric neuroradiologist at Boston Children’s and director of the Advanced Image Analysis Lab. “From there, the clinician can determine if additional scanning is required – saving the child and their family’s time and unnecessary stress.” In the future, the team envisions incorporating machine learning algorithms to provide additional, automated support.

Looking to the future

Innovators throughout Boston Children’s Hospital are looking to develop clinical decision support platforms. To encourage them, IDHA has launched a Clinical Decision Support Innovation Challenge. Winners can receive up to $50,000 in direct costs, R&D software development support and strategic project management support.

Boston Children’s staff with ideas for transformative decision support technologies can apply here. (After the April 28 deadline, applications will be accepted on a rolling basis.)