Abstract
ABSTRACT
Introduction
Accurate diagnosis and assessment of hip involvement are important for the clinical management of ankylosing spondylitis (AS). This study aimed to develop and internally validate an artificial intelligence–based Ensemble Scoring System (ESS) for AS diagnosis and hip‐disease severity assessment and to explore its association with heterogeneous immune‐cell infiltration and age at first documented severe hip involvement.
Methods
A total of 5721 patients with pelvic radiographs and linked clinical records were retrospectively included. Participants were divided into training and held‐out validation cohorts at a 7:3 ratio strictly at the patient level, with no patient or image overlap. Deep‐learning features extracted by a convolutional neural network were integrated with clinical variables using XGBoost to construct task‐specific ensemble deep‐learning (EDL) models. Exploratory proteomic, transcriptomic, single‐cell RNA‐sequencing, immunohistochemical, and peripheral‐blood analyses were performed to investigate AS‐related immune infiltration. An exploratory immune ensemble (IE) score was constructed using the monocyte proportion and EDL score. Kaplan–Meier analysis was used to examine the association between the IE score and age at severe hip involvement.
Results
The EDL model achieved AUCs of 0.978 for distinguishing AS from non‐AS, 0.990 for identifying hip arthropathy among patients with AS, and 0.996 for identifying BASRI‐Hip grade ≥ 3 among patients with hip arthropathy. Monocyte percentage was an influential feature across the three EDL classification tasks. Exploratory multiomics and immunohistochemical analyses showed higher LYN expression in AS than in the corresponding non‐AS fracture‐control groups and suggested that LYN expression was concentrated predominantly in monocytes. A high IE score was associated with earlier severe hip involvement, with estimated median ages at first documented severe hip involvement of 36 and 50 years in the high‐ and low‐IE‐score groups, respectively. This finding represents exploratory risk stratification and does not establish independent prediction of biological disease progression.
Conclusions
The ESS showed strong internal validation performance for AS diagnosis and hip‐disease severity classification. Exploratory analyses linked LYN dysregulation in monocytes with heterogeneous immune infiltration, while the IE score was associated with earlier severe hip involvement. Prospective multicenter external validation is required before clinical application.