Abstract
The predictive models developed in this study can enable accurate preoperative estimation of length of stay (LOS) and risk of rehabilitation discharge for adult patients undergoing elective spine surgery. The demonstrated models exhibited better performance than the American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP) for prediction of LOS and equivalent performance to NSQIP for prediction of discharge location.
Introduction
In the current healthcare landscape of "bundled payments," hospitals receive fixed reimbursements for surgical episodes. This financial model places a premium on accurate preoperative planning; unexpected prolonged hospital stays or discharges to skilled nursing facilities can significantly increase costs and lead to financial losses for the healthcare system.
While the American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP) risk calculator is the current industry standard for risk assessment, it often lacks the granularity required for complex spine procedures. This study aimed to develop and validate spine-specific predictive models for length of stay (LOS) and discharge disposition, comparing their performance against the established NSQIP benchmarks.
Methodology
The researchers employed a retrospective analysis of electronic health records (EHR) and compared several statistical and machine learning approaches:
- Generalized Linear Models (GLM): Poisson Regression was specifically chosen for LOS modeling. Since LOS is measured in discrete days (count data) and cannot be negative, the Poisson distribution is mathematically superior to standard linear regression for this application.
- Logistic Regression & LASSO: These techniques were used to predict the binary outcome of discharge disposition (Home vs. Rehab/Skilled Nursing Facility).
- Decision Tree Learning: This was utilized as a non-linear alternative to evaluate potential improvements in predictive accuracy.
A critical component of these models was the inclusion of the Surgical Invasiveness Index. Unlike standard procedure codes, this index quantifies surgical stress by scoring the number of vertebral levels decompressed, fused, or instrumented, providing a more precise measure of the surgical burden.
Data Overview
The study utilized a de-identified EHR dataset from the University of California, San Francisco (UCSF), spanning 2014 to 2019.
- Cohort: 3,678 adult patients undergoing elective spine surgery.
- Stratification: Patients were categorized into three clinical groups: cervical degenerative, lumbar degenerative, and adult spinal deformity (ASD).
- Validation: The data was split into a training set (84%, n=3,091) and a validation set (16%, n=587).
Key Insights
The Value of Spine-Specific Granularity
The primary advantage of the new models over the ACS NSQIP calculator for LOS prediction stems from the inclusion of spine-specific variables. General calculators often treat diverse spine surgeries as similar risks, whereas the difference between a single-level fusion and a multi-level deformity correction is substantial in terms of recovery trajectory.
Surgical Invasiveness as a Primary Predictor
The Surgical Invasiveness Index emerged as a significant predictor across all models. This suggests that the magnitude of the surgical intervention—the "how much"—is often more predictive of recovery than the specific diagnosis or procedure type.
Demographic and Physiological Factors
The models identified that advanced age, higher BMI, and Black race were associated with increased LOS and a higher likelihood of discharge to a rehabilitation facility. These findings underscore the necessity of accounting for socioeconomic and physiological factors during preoperative counseling and resource allocation.
Results
The developed models demonstrated superior or equivalent performance compared to the industry standard:
- Length of Stay (LOS): The machine learning models (Poisson and LASSO) achieved an $R^2$ of 0.29, nearly doubling the predictive power of the ACS NSQIP calculator ($R^2 = 0.16$).
- Discharge Disposition: The Logistic Regression model achieved an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.79, statistically equivalent to the NSQIP’s 0.75.
- Clinical Benchmarks: Within the study cohort, the average LOS was 3.66 days. Approximately 78% of patients were discharged home, while 22% required post-acute care in a rehabilitation or skilled nursing facility.
Limitations and Future Directions
Despite the improved performance, several limitations remain. The study was conducted at a single high-volume academic center, which may limit the generalizability of the models to community hospital settings with different patient demographics or discharge protocols. Additionally, the reliance on ICD-10 codes for comorbidities can lead to under-reporting of condition severity.
Future research should focus on incorporating social determinants of health, such as the Risk Assessment and Prediction Tool (RAPT) for social support, and expanding the dataset to include multiple institutions to further refine and validate these predictive tools.
Summary updated