OBJECTIVE: To train and validate an algorithm mimicking decision making of experienced surgeons regarding upper instrumented vertebra (UIV) selection in surgical correction of thoracolumbar adult spinal deformity. METHODS: A retrospective review was conducted of patients with adult spinal deformity who underwent fusion of at least the lumbar spine (UIV > L1 to pelvis) during 2013-2018. Demographic and radiographic data were collected. The sample was stratified into 3 groups: training (70%), validation (15%) and performance testing (15%). Using a deep learning algorithm, a neural network model was trained to select between upper thoracic (T1-T6) and lower thoracic (T7-T12) UIV. Parameters used in the deep learning algorithm included demographics, coronal and sagittal preoperative alignment, and postoperative pelvic incidence-lumbar lordosis mismatch. RESULTS: The study included 143 patients (mean age 63.3 ± 10.6 years, 81.8% women) with moderate to severe deformity (maximum Cobb angle: 43° ± 22°; T1 pelvic angle: 27° ± 14°; pelvic incidence-lumbar lordosis mismatch: 22° ± 21°). Patients underwent a significant change in lumbar alignment (Δpelvic incidence-lumbar lordosis mismatch: 21° ± 16°, P < 0.001); 35.0% had UIV in the upper thoracic region, and 65.0% had UIV in the lower thoracic region. At 1 year, revision rate was 11.9%, and rate of radiographic proximal junctional kyphosis was 29.4%. Neural network comprised 8 inputs, 10 hidden neurons, and 1 output (upper thoracic or lower thoracic). After training, results demonstrated an accuracy of 81.0%, precision of 87.5%, and recall of 87.5% on testing. CONCLUSIONS: An artificial neural network successfully mimicked 2 lead surgeons' decision making in the selection of UIV for adult spinal deformity correction. Future models integrating surgical outcomes should be developed.
OBJECTIVE: To train and validate an algorithm mimicking decision making of experienced surgeons regarding upper instrumented vertebra (UIV) selection in surgical correction of thoracolumbar adult spinal deformity. METHODS: A retrospective review was conducted of patients with adult spinal deformity who underwent fusion of at least the lumbar spine (UIV > L1 to pelvis) during 2013-2018. Demographic and radiographic data were collected. The sample was stratified into 3 groups: training (70%), validation (15%) and performance testing (15%). Using a deep learning algorithm, a neural network model was trained to select between upper thoracic (T1-T6) and lower thoracic (T7-T12) UIV. Parameters used in the deep learning algorithm included demographics, coronal and sagittal preoperative alignment, and postoperative pelvic incidence-lumbar lordosis mismatch. RESULTS: The study included 143 patients (mean age 63.3 ± 10.6 years, 81.8% women) with moderate to severe deformity (maximum Cobb angle: 43° ± 22°; T1 pelvic angle: 27° ± 14°; pelvic incidence-lumbar lordosis mismatch: 22° ± 21°). Patients underwent a significant change in lumbar alignment (Δpelvic incidence-lumbar lordosis mismatch: 21° ± 16°, P < 0.001); 35.0% had UIV in the upper thoracic region, and 65.0% had UIV in the lower thoracic region. At 1 year, revision rate was 11.9%, and rate of radiographic proximal junctional kyphosis was 29.4%. Neural network comprised 8 inputs, 10 hidden neurons, and 1 output (upper thoracic or lower thoracic). After training, results demonstrated an accuracy of 81.0%, precision of 87.5%, and recall of 87.5% on testing. CONCLUSIONS: An artificial neural network successfully mimicked 2 lead surgeons' decision making in the selection of UIV for adult spinal deformity correction. Future models integrating surgical outcomes should be developed.