A model outperformed simpler statistical approaches in predicting which patients would encounter trouble
One of the worst nightmares for people who undergo surgery is the onset of complications that force an emergency room visit soon after their procedure.
Emergency readmissions of surgery patients are troubling for hospitals, too. They incur non-refundable costs, as compensation mechanisms have changed under the Affordable Care Act. Hospitals also risk malpractice lawsuits, and readmissions put additional strain on staff resources. Some hospitals have tried using classic statistical computer models to predict the likelihood of surgery patients’ returning unplanned, but the models’ accuracy has been limited.
That led a team of researchers to propose a new approach: Harness recent advances in machine learning to do a better job of identifying surgery patients likeliest to end up back in the hospital. The longer-term goal, of course, would be to reduce the number of readmissions by pinpointing what’s driving them — and fixing those issues before the patient’s initial discharge from the hospital.
In a study published in the medical journal Anesthesiology, UCLA Anderson’s Velibor V. Mišić and Kumar Rajaram, Ronald Reagan UCLA Medical Center’s Eilon Gabel and Ira Hofer, and University of Pittsburgh’s Aman Mahajan report devising three “machine learning” models that scored significantly higher than standard statistical programs in predicting which surgical patients were at greatest risk of readmission.
They began with data on 34,532 adult surgical admissions to UCLA Medical Center from April 2013 to December 2016. From that total they culled 1,942 cases (5.6% of the total) that reentered the hospital in an emergency within 30 days of their surgery.
The authors then constructed a database of as many as 1,013 patient variables, the first step in identifying patterns that could point to a high likelihood of readmission. The variables included demographic markers such as a patient’s age, ethnicity, race and primary language; duration of surgery and volume of blood loss; prescribed medications; lab test results showing levels of markers such as white blood cells, sodium, glucose and the red blood cell protein hemoglobin; and which of the 170 surgical/medical/consulting teams at the hospital was assigned to a patient at the initial admission.
The large number of variables involved per patient was key to the team’s decision to use machine learning models to assess the data. Unlike classic statistical analysis, machine learning models can process large volumes of data and then use it to learn and improve from the experience (as a human mind would), without being instructed how to do so.
What’s more, two of the three models used in the study are based on so-called classification trees, a predictive model that follows a sequence of true/false queries about individual variables that create a growing “tree.” Classification trees also have the advantage of being able to discover nonlinear relationships and interactions among variables, as opposed to linear (“straight-line”) relationships and interactions.
The study’s overall findings:
- The three machine learning models showed nearly uniform success in predicting patient emergency returns within 30 days of surgery: The accuracy rates ranged from 85% to 87%, while most prior statistical models were 60% to 70% accurate, the authors write.
- There is “virtually no loss in performance if our models are restricted to using data available within 36 hours after the completion of surgery,” the study says. The point being, there’s no need to wait until a patient’s discharge date to gather and test data. “This suggests that our models could be used to identify patients that are at high risk of readmission while still in the hospital and soon after the surgical procedure,” the study says.
- The authors tested their initial methodology in two additional studies, first using data on surgical emergency readmissions at Santa Monica Hospital from April 2013 to December 2016, and then predicting 2017–2018 readmissions from 2013–2016 data at UCLA Medical Center. In both cases, the predictive accuracy was virtually identical to that of the initial study.
In an interview, UCLA Anderson’s Mišić said machine learning “is still a relatively new tool in health care.” But use of the technology may get a boost as more medical institutions consolidate their wealth of data in electronic “warehouses,” he said. “Such database systems are crucial to supporting this type of work,” Mišić said.
His Anderson colleague, Rajaram, said in an interview that machine learning “could be used to predict length of hospital stays and potentially lead to better planning and resource management of various types of beds, such as general, ICU and critical care.” Another use, he said, “could be in predicting surgical durations, allowing better planning and scheduling of operating rooms and staff resources.”