One of the most common surgical complications is postoperative pain that lasts long after the surgical incision has healed, colliding between 10% and 35% of the estimated 300 million people worldwide who undergo surgery each year.
The reason for the pain after this surgery remains unknown. Analyzing entanglements of risk factors can be difficult. Pain manifests not only surgical trauma, but also a complex combination of interactions between the peripheral nervous system and the central nervous system, the immune system, and the emotional and cognitive abilities of the person who handles the pain.
That's where machine learning comes in. As data was collected prior to surgery, machine learning algorithms can tear through many factors present and predict who could strain persistent surgical pain.
Previous clinical trials to prevent this pain have failed when attempting to alleviate individual risk factors in a highly diverse population of surgical patients.
“We've seen a lot of trouble with our patients,” said Simon Hallowtunian, professor of anesthesiology at Washington University School of Medicine in St. Louis. There is no single formula for determining individual risk, he added.
“It's not a simple 1+1 type. Here we collect some measurements and build an accurate risk profile,” says Haroutonian. “This really wants machine learning to bring benefits and teases some of those small contributors at personal risk.”
Haroutonian is part of Washu's interdisciplinary team investigating the issue, including Chenyang Lu, AI Director at the Health Institute and Fullgraf Professor of Computer Science and Engineering at the McKelvey School of Engineering.
In a study published in ACM's Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technology, LU and the team share how machine learning can help guide physicians to prevent postoperative pain. Most importantly, the system not only predicts who might develop that pain, but also provides estimates of the uncertainty for each prediction.
Being able to effectively communicate uncertainty can make all the difference in physician decisions. Lu and his team developed a “uncertain” machine learning model because they wanted to include not only their ability to predict patient risk, but also how confident AI is about their risk estimates.
“It gives the model the ability to say 'I don't know' and quantifies that uncertainty,” said Ziqi Xu, a doctoral student at LU Lab and the first author of the paper.
A common problem with AI-driven clinical decision support systems is to provide a YES or NO answer, but there is no details about how confident the machine is in the answer, Lu said. He compared it using generative AI programs like ChatGpt:Machine. The machine can give you “confidence” in its responses and responses to prompts, even hallucinations.
However, clinicians need to know the level of uncertainty in their predictions, so they can use their knowledge to make the best decisions. Humans and machine learning systems are intended to work as teams, and “if they don't communicate uncertainty in a calibrated way, it can cause problems,” Lu added.
To provide these estimates, the team enrolled 782 patients and participated in the study. They asked people to fill out a series of daily research questions delivered to their smartphones a few days or weeks before the surgery. Data missing from uncertainty estimates were factored because not all patients took the time to complete the survey.
Lu then combined the findings with clinical information such as patient health history, lab results, and more. His team developed a new model that provides estimates of uncertainty based in part on the amount of data provided by patients and the individual factors in risk assessment.
You might say to the model: Patient X has a 30% chance of developing persistent pain, but that estimate has a 50% chance of “uncertainty.” In such cases, doctors need to research more and lean on clinical knowledge to help patients make the best choice to manage their pain.
In another example, the model might say that patient Y has a 10% chance of developing persistent pain, and the model is 80% certain of that estimate. In that case, the doctor can more safely assume the predictability of the risk of persistent pain.
When testing models against other prediction algorithms, the team found that performance was improved and the model was optimal for “calibration performance.”
From data to doctors
Incorporating the model into the clinical decision support process is the next step in the research, Lu said.
Doctors hope that using the data can be used to predict who develop sustained postoperative pain, but importantly, “we want to understand why,” Lu added. “Understanding causal relationships and developing interventions is important.”
Machine learning helps to support its discovery process and identify variables most relevant to persistent pain, information that can guide better clinical trials.
In some patients, drivers at risk of postoperative pain are more behavioral, and cognitive behavioral therapy (CBT) interventions may provide a solution.
However, other patients may experience pain due to dysregulation of the immune response to surgery, and in such cases the CBT approach may not be sufficient. The focus may need to shift towards interventions that may alter the immune or inflammatory response to surgery, Lu said.
This ongoing work aims to refine the model and identify the causes of persistent postoperative pain, but is supported by a $5 million grant from the National Institutes of Health (NIH). As teams continue to test predictive algorithms, the next step is to develop personalized interventions based on each patient's risk profile.
What contributes to vulnerability or resistance to surgical pain, and testing approaches to address these risks, could ultimately make a big difference in the number of people suffering from pain, Hallowtunian added.
Ziqi Xu, Jingwen Zhang, Simon Haroutonian, Hanyang Liu, Zihan Cao, Gabrielle Rose Messner, Harutyun B Alaverdyan, Saivee Ahuja, Rahual Koshy, Joel Hanns, Madelyn Frumkin, Thomas L. Rodebaug, and Chenyang Lu. Incorporating uncertainty into predictive models using mobile sensing and clinical data: a case study of persistent surgical pain. Proc. ACM interaction. Mob. Wearable ubiquitous technol. 9, 2, Article 58 (June 2025) https://doi.org/10.1145/3729488.
This study was supported by a CDMRP grant from the US Department of Defense to Dr. Simon Halooutunian and support from NIH grant 1RM1NS135283-01 by Dr. Simon Halooutunian and Dr. Chenyan Lu (and Dr. Megan Creed, Prachik Sinha and Dr. Thomas Lord Singer).
Originally published on the McKelvey Engineering website
