Spinal morphology-based multimodal AI for predicting pulmonary dysfunction in adolescent idiopathic scoliosis.

PURPOSE: This study aimed to develop a multimodal model that integrates morphological and functional features to address pulmonary function decline associated with three-dimensional structural changes in Adolescent Idiopathic Scoliosis (AIS). The proposed model predicts pulmonary dysfunction by combining clinical indicators with X-rays, which directly depict the morphological deformities characteristic of AIS. METHODS: This study included 178 patients with AIS who underwent standing posteroanterior radiography and pulmonary function testing as part of preoperative evaluation. A dual-stream deep learning model was developed, employing an EfficientNet backbone to extract features from radiographs and a multilayer perceptron (MLP) for structured clinical data. These inputs were integrated using Feature-wise Linear Modulation (FiLM) within a multi-task learning framework to jointly predict binary abnormalities (< 80% of the predicted value) in FVC and FEV1. Model performance was evaluated using AUC through five-fold cross-validation and benchmarked against conventional machine learning models trained exclusively on tabular data. RESULTS: The optimal multimodal configuration (EfficientNet-B0 combined with MLP using FiLM fusion) achieved AUCs of 0.814 ± 0.031 for FVC and 0.841 ± 0.033 for FEV1, yielding a macro-AUC of 0.827 ± 0.022. This configuration outperformed logistic regression models trained on tabular features (0.719 ± 0.051 for FVC and 0.710 ± 0.057 for FEV1). CONCLUSION: The multimodal deep learning model that integrates holistic radiographic analysis with clinical data demonstrated superior accuracy in predicting pulmonary dysfunction in patients with AIS compared with conventional methods. This data-driven approach shows promise for improving clinical risk assessment and supporting personalized decision-making.