Researchers who have been using Fitbit data to predict surgical outcomes have found a new way to more accurately measure how patients are recovering after spinal surgery.
Using machine learning techniques, researchers set out to develop a way to more accurately predict recovery from lumbar spine surgery.
The results were published in the journal Proceedings of the ACM on Interactive, Mobile, Wearable, and Ubiquitous Technologieshave shown that their model for predicting spine surgery outcomes is superior to previous models.
This is important because outcomes in back surgery, and many other types of orthopedic surgery, vary widely depending on the patient's structural disease as well as the various physical and mental health characteristics of each patient.
Recovery after surgery is influenced by both physical and mental health before surgery. Some people worry excessively in the face of pain, which can worsen their pain and recovery. Others have physiological issues that make pain worse. If doctors can know in advance the various pitfalls patients may face, they can better plan their treatment.
“Predicting outcomes before surgery can help us set some expectations, aid in early intervention and identify high-risk factors,” says lead author Zhiqi Xu, a doctoral student in the lab of Chengyang Lu, a professor in the McKelvey School of Engineering at Washington University in St. Louis.
Previous studies predicting surgical outcomes typically used patient questionnaires administered once or twice in the clinic, capturing static slices over a period of time.
“We were unable to capture the long-term changes in patients' physical and psychological patterns,” Xu says. Previous studies of training machine learning algorithms have focused on only one aspect of surgical outcomes, “ignoring the inherent multidimensional nature of post-surgical recovery,” he adds.
Researchers have used mobile health data from Fitbit devices to monitor and measure recovery and compare activity levels over time, but the new study finds that combining activity data with long-term assessment data can more accurately predict how patients will do after surgery, said Jacob Greenberg, M.D., assistant professor of neurosurgery at the School of Medicine.
“This study is a 'proof of principle' that shows multimodal machine learning can help clinicians get a more accurate 'big picture' of the interrelated factors that influence recovery. Before embarking on this study, the research team first developed statistical methods and protocols to ensure they provided the artificial intelligence system with the right balanced data.”
The team previously published a paper in the journal Neurosurgery This is the first time we have shown that objective patient-reported wearable measurements improve prediction of early recovery compared with traditional patient assessment.
In addition to Greenberg and Xu, Madeline Frumkin, a doctoral student studying psychology and brain sciences in the lab of Thomas Rodebau, is co-first author of the study. Wilson “Zach” Ley, professor of neurosurgery in the School of Medicine, is co-senior author with Rodebau and Lu. Rodebau is now at the University of North Carolina at Chapel Hill.
The study showed that Fitbit data correlated with multiple surveys assessing a person's social and emotional state. The data was collected through an “ecological momentary assessment” (EMA), which uses smartphones to frequently prompt patients to assess their mood, pain levels and behavior multiple times throughout the day.
“We combine wearables, EMA, and clinical records to collect a wide range of information about patients, from physical activity to subjective reports of pain and mental health, and clinical characteristics,” Lu says.
Greenberg added that cutting-edge statistical tools such as “dynamic structural equation modeling,” which Rodebaugh and Frumkin helped advance, played a key role in analysing the complex longitudinal EMA data.
In their latest study, the researchers took all these factors into account and developed a new machine learning technique called “multimodal multitask learning” to effectively combine different types of data to predict multiple recovery outcomes.
With this approach, AI learns to assess the associations between outcomes while capturing differences in outcomes from multimodal data, Lu adds.
According to Xu, the method captures shared information about the interrelated tasks of predicting different outcomes and leverages that shared information to help the model understand how to make accurate predictions.
All of this is compiled into a final package to generate predicted changes in postoperative pain interference and physical function scores for each patient.
Greenberg says the study is ongoing so that researchers can continue to fine-tune the model, conduct more detailed evaluations, predict outcomes, and, most importantly, “understand what factors we might be able to modify to improve long-term outcomes.”
The research was funded by AO Spine North America, the Cervical Spine Research Association, the Scoliosis Research Association, Barnes-Jewish Hospital Foundation, the University of Washington/BJC Healthcare Big Ideas Competition, the Fullgraph Foundation and the National Institute of Mental Health.
Source: Washington University in St. Louis
