Smarter Patient Selection for Outpatient Spine Surgery: Can Machine Learning Help?

-Sukanya Rao

As more spine procedures move to the outpatient setting, selecting the right patient becomes increasingly important. The challenge is not simply whether a procedure can be performed outpatient, but whether an individual patient is likely to recover without an unexpected complication, prolonged stay, or readmission.

Outpatient candidacy depends on multiple clinical factors considered together. Could machine learning help surgeons translate those variables into a more individualized assessment of risk?

Can an Algorithm Add to Patient Selection?

A review published in the Journal of Spine Surgery examined how machine-learning models are being applied across cervical and lumbar spine surgery, with particular attention to their potential role in outpatient patient selection.

Rather than examining one model or outcome, the review brought together studies using preoperative data to predict surgical candidacy, complications, length of stay, discharge disposition, readmission, and postoperative outcomes. These are different predictions, but each could provide information relevant to determining whether a patient is suitable for outpatient care.

What Are the Models Actually Predicting?

One study evaluated 12,492 patients undergoing single-level outpatient ACDF to determine whether preoperative patient characteristics could help identify patients at higher versus lower risk of an adverse outcome and, in turn, support selection for outpatient surgery. The model was assessed using the area under the receiver operating characteristic curve (AUC), which measures how well it distinguishes between higher- and lower-risk patients. An AUC of 0.5 means the model cannot distinguish between these groups better than chance, while 1.0 represents perfect separation. The model achieved an AUC of 0.757, indicating a moderate ability to distinguish patients considered more versus less appropriate for outpatient ACDF. It also performed better than two conventional risk-stratification measures used for comparison.

Another study examined whether machine learning could identify patients at risk for 30-day readmission after posterior lumbar fusion. Using data from 23,264 patients, the model achieved a mean AUC of 0.812, indicating good ability to distinguish patients who were more versus less likely to be readmitted. Its mean positive predictive value was 78.5%, meaning that among patients classified by the model as likely to be readmitted, approximately 79% actually experienced the outcome within the study dataset. This suggests that preoperative information may help identify patients who warrant closer attention to postoperative and discharge planning. 

A separate study focused on length of stay after cervical spine fusion, an important consideration when determining whether outpatient care is appropriate. Researchers used electronic medical record data to predict which patients were more likely to require a longer hospitalization. Performance was measured using AUROC, which, like AUC, indicates how well the model separates patients with different outcomes. The gradient-boosted model achieved an AUROC of 0.87 in single-center validation and 0.84 when tested using a nationwide dataset. These values indicate good predictive performance, with the model retaining relatively strong discrimination when evaluated beyond the original center.

Machine learning has also been tested for specific complications. In 184 patients with cervical myelopathy undergoing posterior laminectomy and fusion, a support-vector-machine model used clinical and imaging variables to predict postoperative C5 palsy. It achieved an AUC of 0.923 and an accuracy of 91.8%, demonstrating strong predictive performance within that study population.

Together, these studies show that machine learning is being used not just to assign an outpatient or inpatient label, but to estimate specific outcomes that can inform that decision.

Prediction Is Promising, But Is It Ready for Practice?

The results are encouraging, but predictive performance within a dataset is not the same as demonstrated clinical benefit.

Machine-learning models depend heavily on the populations used to develop them. A model that performs well in one institution or patient population may not perform equally well elsewhere. Although the length-of-stay model maintained good performance in a nationwide dataset, the review identifies external validity as an important limitation across the broader evidence.

The studies also varied considerably in procedures, patient populations, outcomes, datasets, and algorithms, and machine learning was not consistently superior to conventional approaches.

Most importantly, these studies demonstrate that certain postoperative outcomes can be predicted with varying levels of accuracy. They do not establish that using these models to select patients for outpatient surgery reduces complications, readmissions, or length of stay.

From Can We? to For Whom?

As outpatient spine surgery expands, the question is increasingly not only whether a procedure can be performed outpatient, but which patient is appropriate for that pathway.

Machine learning may offer another way to identify risks that matter to that decision by bringing multiple preoperative variables together. For now, its most relevant role may be as a decision-support tool that complements not replacing surgeon judgment.

The real value will depend not simply on how well an algorithm predicts an outcome, but whether that prediction helps surgeons make better-informed decisions about who is ready for outpatient spine surgery.

Interested in Reading the Full Study? https://pubmed.ncbi.nlm.nih.gov/37841781/