Oral Presentation Sydney Spinal Symposium 2025

Predicting complications after elective spinal surgery for lumbar disc herniation and spinal stenosis: development and validation of multivariable prediction models (124864)

Giovanni Ferreira 1 , Bjornar Berg 2 , Steven Hicks 2 , Are Pripp 2 , Lars Christian Haugli Bråten 3 , Tarjei Rysstad 4 , Christopher Maher 1 , Tore Solberg 5 , Margreth Grotle 2
  1. Institute for Musculoskeletal Health, Institute for Musculoskeletal Health, Sydney Local Health District & The University of Sydney, Sydney, NSW, Australia
  2. Centre for Intelligent Musculoskeletal Health, Oslo Metropolitan University, Oslo, Norway
  3. Division of Clinical Neuroscience, Oslo University Hospital, Oslo, Norway
  4. Department of Physiotherapy, Oslo Metropolitan University, Oslo, Norway
  5. The Norwegian Registry for Spine Surgery, Tromsø, Norway

Background: Surgical interventions are likely more effective than non-surgical interventions in improving pain and disability in the short-term in people with lumbar disc herniation and lumbar spine stenosis. However, there are risks associated with those procedures, such as complications. Existing models to predict the risk of complication have been developed, but have several limitations. 

Aim: To develop and validate multivariable models to predict the risk of complication in people undergoing surgery for lumbar disc herniation and lumbar spine stenosis. 

Methods: Using data from the Norwegian Registry for Spine Surgery (The Registry; 100% coverage of surgical units in Norway), we identified patients who underwent first-time elective surgery in public/private hospitals. We developed the model using 30 candidate predictors including patient demographics, clinical features, and healthcare use prior to surgery. The outcome of interest was 90-day complication based on ICD-10 codes as recorded in The Registry. We developed three models: a logistic regression and two machine learning (XGBoost and neural network) and validated it using internal-external cross validation. We assessed key performance measures: overall fit (Cox-Snell R2; Brier Scores), calibration (calibration-in-the-large, calibration slope), discrimination (C-statistic). Clinical utility was assessed using decision curve analysis.

Results: We identified 29,566 patients, of whom 1,356 (4.6%) experienced a complication. The models showed evidence of some overprediction (eg calibration-in-the-large for logistic regression model: -0.19, 95% CI -1.13 to 0.76). The logistic regression model had good discrimination (C-statistic: 0.64, 95% CI 0.62 to 0.66). The logistic regression may have clinical utility at lower predicted risks of complication (~6%-15%) compared to treat-all/treat-none strategies. Machine learning analyses are underway; preliminary analyses suggest substantially better performance compared to the logistic regression model.

Conclusion: Our models may have some clinical utility and assist in the early identification of patients undergoing spinal surgery at high risk of post-surgical complications.