594 - Prediction of Alignment after Surgical Treatment of Adult Spinal Defor...

Oral Posters: Innovative Technologies

Presented by: V. Lafage - View Audio/Video Presentation (Members Only)


B. Moal(1), O. Moal(2), R. Lafage(3), F. Schwab(3), J. Smith(4), C. Shaffrey(4), A. Jain(5), S. Bess(6), J. Gum(7), G. Mundis(8), C. Ames(9), E. Klineberg(10), M. Gupta((1)(1)), V. Lafage(3), International Spine Study Group (ISSG)

(1) Centre Hospitalier, Universitaire de Bordeaux, Bordeaux, France
(2) Hospital for Special Surgery, New York, NY, United States
(3) Hospital for Special Surgery, Spine Service, New York, NY, United States
(4) University of Virginia Medical Center, Department of Neurosurgery, Charlottesville, VA, United States
(5) Johns Hopkins University School of Medicine, Department of Orthopaedic Surgery, Baltimore, MD, United States
(6) NYU Langone Medical Center, Spine Division, Department of Orthopaedics, New York, NY, United States
(7) Norton Leatherman Spine Center, Louisville, KY, United States
(8) Scripps Clinic Torrey Pines, La Jolla, CA, United States
(9) San Francisco Medical Center, University of California, Department of Neurosurgery, San Francisco, NY, United States
(10) University of California, Davis, Department of Orthopedic Surgery, Sacramento, CA, United States
((1) (1) ) Washington University School of Medicine, Department of Orthopaedic Surgery, St. Louis, MO, United States


Introduction: Sagittal plane malalignment is recognized as a key driver of disability and pain in ASD. However, prediction of postoperative sagittal alignment, including changes in unfused vertebrae and pelvic version (pelvic tilt [PT]) remains challenging. Recent advances in machine learning algorithms and the availability of large spine reconstruction databases may offer new means to predict postoperative spino-pelvic alignment. The hypothesis of this study is that ostoperative spinal alignment can be predicted using machine learning algorithms.

Methods: This study is a retrospective analysis of a prospective database of surgical ASD patients. Radiographic sagittal reconstructions at baseline and 1 yr after surgery for ASD patients fused from the sacrum up to between L1 and T4 were prospectively collected. Point distribution model framework and PPCA were used to build a model based on pre- and postoperative spine reconstruction. The unfused thoracic portion and the PT were predicted using both preoperative and the fused portion of the postoperative reconstructions. Model performance was evaluated based on the root-mean-square error (RMSE) between the prediction and the actual postoperative endplate location and on the absolute error of the following parameters: PT, T1SPI and T1-T12 angles. RMSE was expressed in percentage of L1 endplate length.

Results: The model was trained with 250 patients and tested on 60. The average RMSE related to the prediction of the location of the unfused endplate corners was 26%±20. The prediction of the endplate location led to the following average absolute error in terms of radiographic parameters: 3.6˚±3.0 for PT, 3.1˚±2.1 for T1SPI and 8.5˚±6.3 for T1-T12.

Conclusion: The model developed permits prediction of the location of the vertebral endplates and provided good postoperative PT prediction but higher margin of error for prediction of the unfused thoracic spine. Combining the pre-operative surgical planning/simulation with modeling can offer patient-specific prediction of postoperative alignment. Such prediction offers the potential to substantially improve surgical effectiveness.

Comparison between Prediction (red) vs real (blue)]