An MRI-based radiomics model for precision subtyping of hereditary spastic paraplegia: discriminating SPG4 from SPG5

Differentiating spastic paraplegia type 4 (SPG4) from type 5 (SPG5) is clinically relevant with differing therapeutic strategies, yet genetic testing remains limited necessitate alternative approaches. This study aimed to develop and validate an interpretable radiomics model based on spinal cord MRI for distinguishing SPG4 from SPG5.

In this prospective study, SPG4 (N = 40) and SPG5 (N = 30) were randomly assigned to training (n = 56) and test (n = 14) sets. Radiomic features were extracted from the whole spinal cord (WSC) and cervical spinal cord functional subregions (SCFS) on three-dimensional T1- and T2-weighted images. Predictive models were constructed using four machine learning algorithms and evaluated by receiver operating characteristic curve (ROC) analysis. Model robustness was assessed by permutation testing, and interpretability was provided through SHapley Additive exPlanations (SHAP).

Radiomics models based on the WSC consistently outperformed those based on SCFS. The optimal WSC model, constructed with a linear support vector classifier on combined T1- and T2-weighted sequences, achieved an AUC of 1.00 in the test set. SHAP analysis identified five consensus features contributed significantly to discrimination between SPG4 and SPG5. All five features also showed significant differences between the two groups (p < 0.001), with the most influential feature (T1_original_shape_SurfaceVolumeRatio) correlating with disease severity as measured by the Spastic Paraplegia Rating Scale (r = 0.434, p < 0.01).

Spinal cord radiomics enables accurate and interpretable differentiation between SPG4 and SPG5, providing a non-invasive adjunct to genetic testing. This approach offers objective imaging biomarkers that may support precision diagnosis, improve patient stratification in hereditary spastic paraplegia.