Digital twin validation and predictive value of delta radiomics based on tabular transformer for synergistic muscle imbalance in lumbar disc degeneration.

backgroundCurrent methods struggle to quantify muscle synergistic imbalance and nonlinear interactions in Lumbar Intervertebral Disc Degeneration (LIDD). This study proposes a novel strategy combining "Delta Radiomics" with a Tabular Transformer deep learning architecture, aiming to improve the precision of early risk assessment.

methodsA retrospective analysis was conducted on L4-L5 MRI images from 663 subjects. Texture features of the Psoas Major (PM) and Erector Spinae (ES) were extracted, and (PM-ES)/ES was calculated as the "Delta Feature" to eliminate individual variation. An in silico digital twin simulation system was constructed to validate the physical sensitivity of these features. The predictive performance of the nonlinear Tabular Transformer model was compared with a traditional logistic regression model.

resultsIn silico experiments confirmed that Delta features possess superior "common-mode noise rejection" capabilities. In the early stage of degeneration (Step 20), when original features showed only weak changes, Delta features detected highly significant inter-group differences (P < 0.001). In the independent test set, the Tabular Transformer model demonstrated the best discriminative power with an AUC of 0.848 (95% CI: 0.781-0.916), outperforming the Combined linear model utilizing the same features (AUC = 0.804, P = 0.089) and single-modality models. Although the raw AUC improvement was marginal, Decision Curve Analysis (DCA) and quantitative calibration (Brier score = 0.145) showed that the Transformer model offered higher net clinical benefit and superior absolute risk estimation across a wider threshold range. Additionally, SHAP analysis revealed that the core features identified by the model highly aligned with the biomechanical compensation failure mechanisms set in the simulation, validating the model's decision logic.

conclusionDelta radiomics can sensitively capture microscopic muscle imbalance. Combined with the nonlinear modeling advantages of the Tabular Transformer architecture, this approach provides an objective, interpretable, and quantitative basis for the early non-invasive diagnosis and risk stratification of LIDD.