Development and validation of a deep learning-powered system for multi-version global alignment and proportion score to predict mechanical complications after adult degenerative scoliosis surgery.

Abstract

OBJECTIVES: The global alignment and proportion (GAP) score was designed to predict mechanical complications (MC) following adult degenerative scoliosis (ADS) surgery. Despite its foundational role, its clinical application has been hindered by the labor-intensive nature of manual calculations, inefficiencies, and the complexity of evolving modified versions. This study aimed to develop an innovative deep-learning-powered tool for vertebrae detection and automated computation of editable, multi-version GAP scores. METHODS: Leveraging a robust dataset of 3,485 spinal X-rays (both screw-containing and screwless images), we trained and tested a vertebrae detection model. Automatic extraction of sagittal parameters was seamlessly integrated into a multi-version GAP scoring system. A retrospective cohort of ADS surgery patients was used to validate various GAP score versions, including a novel ethnicity-age-gender-adjusted C-GAP (EAGA C-GAP) score. RESULTS: Our model demonstrated robust detection performance (mean average precision: 0.780 for screw-containing images and 0.732 for screwless images). The automatic GAP scoring system supports editable versions, with the EAGA C-GAP achieving superior predictive accuracy for MC (area under the curve, AUC = 0.627), outperforming the original and ethnicity-adjusted versions. Among GAP score components, relative lumbar lordosis (RLL) emerged as the most pivotal predictor, while the lordosis distribution index (LDI) contributed the least. CONCLUSIONS: This groundbreaking integration of artificial intelligence into GAP score computation represents a transformative advancement in ADS surgery outcome prediction. By enhancing efficiency and clinical predictive accuracy, the optimized EAGA C-GAP score sets a new standard for individualized prediction tailored to the Chinese population, reflecting the crucial role of ethnicity, age, and gender considerations in surgical outcome predictions.