A multimodal machine learning model integrating clinical and MRI data for predicting neurological outcomes following surgical treatment for cervical spinal cord injury.

PURPOSE: Predicting the prognosis of cervical spinal cord injury (CSCI) is crucial for patients and healthcare providers, as it informs treatment decisions and rehabilitation planning. This study aimed to develop a multimodal machine learning model integrating clinical and MRI data to predict neurological outcomes in CSCI patients.

METHODS

We conducted a retrospective study of 247 patients with traumatic CSCI who underwent posterior decompression and fusion surgery at a specialized spinal cord injury center between April 2015 and June 2021. Clinical data, including demographics, comorbidities, laboratory data, and neurological findings, were collected. T2-weighted sagittal MRI images were analyzed using a convolutional neural network pre-trained on RadImageNet. Clinical and MRI features were integrated to construct a multimodal predictive model using the Light Gradient Boosting Machine algorithm, validated with 5-fold cross-validation. The primary outcome was defined as achieving American Spinal Injury Association (ASIA) Impairment Scale (AIS) grade D or higher at 6 months post-injury. Shapley Additive Explanations identified key contributing factors in predicting these neurological outcomes.

RESULTS

The model achieved an accuracy of 0.90 and an AUC of 0.94 for predicting recovery to AIS grade D or higher at 6 months post-injury. Key predictors included lower extremity ASIA motor score (AMS), AIS grade at admission, upper extremity AMS, age, HbA1c, and MRI-derived features.

CONCLUSION

This multimodal model demonstrated superior predictive accuracy compared to previous monomodal approaches, emphasizing the value of combining clinical and MRI data. These findings highlight the potential of multimodal predictive models in improving clinical decision-making and outcomes for CSCI patients.