- Orthopedic surgeons at Massachusetts General Hospital have developed a machine learning algorithm that preoperatively predicts the probability of discharge to a rehabilitation or nursing facility after surgery for degenerative spondylolisthesis
- Variables considered are age, sex, diabetes, elective surgery, body mass index, procedure, number of surgical levels, ASA class, white blood cell count and creatinine
- The methodology used to develop the algorithm might be useful to develop discharge prediction models for many other conditions and elective treatments
Recent studies show that a substantial proportion of extended hospital stays—47% in one study of trauma patients—are attributable to having patients wait for placement in a rehabilitation or nursing facility. Prolonged hospitalization increases not only costs but also the risk of adverse events such as venous thromboembolism and hospital-acquired infections.
Subscribe to the latest updates from Orthopaedics Advances in Motion
In response, Orthopedic Surgeon Joseph H. Schwab, MD, MS, chief of the Orthopaedic Spine Center at Massachusetts General Hospital, and colleagues have developed the first machine learning algorithm that predicts discharge placement. In the European Spine Journal, they explain it's designed to be used preoperatively, so healthcare personnel can arrange non-home discharge well in advance.
A Proof-of-Principle Study
The researchers selected 9,338 patients who underwent decompression, fusion, or fixation for acquired spondylolisthesis between 2009 and 2016 for a proof-of-principle study. Their rationale was that these patients represent a sizeable portion of the patients who undergo spine surgery, usually have elective surgery and are relatively older and thereby more at risk of needing non-home discharge.
A large dataset is needed for developing predictive algorithms, and the researchers used the American College of Surgeons National Surgical Quality Improvement Program database.
The median age of the patients was 63, and 63% were female. The non-home discharge rate was 18.6%.
Developing and Testing the Model
Based on a stepwise backward logistic regression analysis, the independently significant variables were age, sex, diabetes, elective surgery, body mass index, procedure, number of surgical levels, American Society of Anesthesiologists class, preoperative white blood cell count and preoperative creatinine.
Four types of machine learning algorithms (boosted decision tree, support vector machine, Bayes point machine and neural network) were trained with these variables to predict which patients were not discharged home. The Bayes point machine proved to be best in terms of discrimination, calibration and overall performance.
The researchers developed the Bayes point machine model into a web-based application. The user simply inputs the necessary variables to determine the probability of non-home discharge. The calculator can be accessed here.
Short-term and Long-term Uses
The researchers believe the calculator would be most useful during a preoperative visit with a nurse practitioner or case manager, after the visit to the surgeon and anesthesiologist when all variables are known. This would allow for patient education, and discharge planning might be started at that time.
The researchers emphasize that rigorous testing of the model's predictive ability is still needed. Data were unavailable on some variables known to affect discharge placement, such as insurance status, employment status and preoperative patient-reported function. In addition, the model in its present form may not be applicable outside the U.S.
The research group does believe, though, that their methodology could be used to develop a discharge prediction model for many other conditions and elective treatments. External validation will be crucial.
Learn more about the Department of Orthopaedics at Mass General
Refer a patient to the Department of Orthopaedics at Mass General