Machine Learning for Predictive Modeling of 90-day Readmission, Major Medical Complication, and Discharge to a Facility in Patients Undergoing Long Segment Posterior Lumbar Spine Fusion

AI IllustrationLong segment lumbar posterior spinal fusion (LSLPSF) is fraught with high complication rates and readmission rates, contributing to the large costs associated with surgery.  In a recent article published in Spine, Sigurd Berven, MD, and colleagues shared their findings from comprehensive models they created using administrative claims data (State Inpatient Data) and machine learning to predict discharge to facility, 90-day readmissions, and 90-day major medical complications after long segment lumbar spine fusion. 

Readmission rates have become an important metric of quality of care, and furthermore, under the Patient Protection and Affordable Care Act, are directly tied to reimbursement. As the healthcare industry moves toward value-based payment programs such as the Bundled Payment Care Initiative, hospitals and surgeons become directly financially responsible for the entire episode of care. 

In order to target costs and optimize outcome, it is critical to understand drivers of discharge disposition, readmission, and medical complications. Tools to predict these outcomes preoperatively can be used to guide both informed decision making for the patient and surgeon as well reimbursement reform. Big data sets and machine learning allow for complex, computer based algorithms to find relationships between variables. While their use is widespread in other areas, such as banking and commercial advertising, it is quite limited in the medical field.  

Berven and colleagues created and validated a predictive model for postoperative outcomes in patients undergoing LSLSF by assessing various machine learning algorithms, including more traditional logistic regression and other techniques not widespread in orthopaedic literature. 

Prediction models for postoperative outcomes rely on the integration of large amounts of clinical and demographic data. The importance of developing predictive models is two-fold. First, they serve a clinical purpose, both in shared decision making and in developing targets for preoperative optimization. Second, they provide information when considering the development of reimbursement models.  

The researchers examined three different machine learning techniques—logistic regression, random forest, and elastic net regression, and found that logistic regression modestly outperformed the other techniques for all three outcome variables. This information can be used to guide decision making between the surgeon and patient, as well as provide information for structuring value-based payment models. 

Sigurd Berven, MD

Dr. Berven’s research interests include assessment of clinical outcomes of surgery, and minimally invasive techniques in spine surgery.  He is also is studying cellular and molecular techniques for the biological regeneration of components of the spine including the intervertebral disc. Dr. Berven has been an invited speaker at national and international conferences, speaking on topics including measurement of outcomes in spine surgery, evaluation and management of spinal disorders, and advanced techniques in spine surgery.

Machine Learning for Predictive Modeling of 90-day Readmission, Major Medical Complication, and Discharge to a Facility in Patients Undergoing Long Segment Posterior Lumbar Spine Fusion. Jain D, Durand W, Burch S, Daniels A, Berven S. Spine (Phila Pa 1976). 2020 Aug 15;45(16):1151-1160. doi: 10.1097/BRS.0000000000003475.