[Value of a deep learning-based visual model for predicting postoperative upper limb functional recovery after severe acute cervical spinal cord injury].

Objective: To assess the value of a deep learning-based visual model for predicting postoperative upper limb functional recovery after severe acute cervical spinal cord injury. Methods: A retrospective study was conducted involving 162 patients with severe acute cervical spinal cord injury treated at the First Affiliated Hospital of Nanjing Medical University between 2020 and 2025. The Upper Extremity Motor Score (UEMS) was used to assess prognosis before treatment and at 6 months post-treatment. Patients with an improvement of ≥10 points were assigned to the good outcome group, and those with an improvement of <10 points to the poor outcome group. Variables showing statistical significance in univariate analysis were taken as initial inputs, and the least absolute shrinkage and selection operator (LASSO) was applied to identify variables associated with prognosis. The patients were divided into a training set (n=97), a validation set (n=32), and a test set (n=33) at a ratio of 6∶2∶2. A predictive model was constructed based on a multilayer perceptron (MLP) and validated on the independent test set. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. Results: Among the 162 patients, 129 were male and 33 were female, with a mean age of (55.5±9.8) years. Based on univariate analysis and LASSO regression, the four variables, i.e., the BASIC score, SCC>50%, preoperative UEMS, and light touch score, were identified. An MLP prediction model was constructed using these four variables. The distributions of risk factors among the training, validation, and test sets were balanced (all P>0.05). During model training, the training loss approached 0, while the validation loss gradually decreased and stabilized. In the training set, the AUC (95%CI) was 0.987 (0.969-0.999), with a sensitivity of 93.33% and a specificity of 89.55%. In the validation set, the AUC (95%CI) was 0.955 (0.868-1.000), with a sensitivity of 90.00% and a specificity of 95.45%. In the test set, the AUC (95%CI) was 0.930 (0.830-0.991), with a sensitivity of 80.00% and a specificity of 95.65%. Neural network weight analysis showed that preoperative UEMS was the strongest positive predictor, followed by the light touch score, SCC, and BASIC score; among them, the BASIC score and SCC were negative predictors, while preoperative UEMS and the light touch score were positive predictors. The calibration curve demonstrated good calibration ability of the model. Conclusion: The deep learning prediction model based on MLP demonstrates good predictive performance for predicting upper limb functional outcomes in patients with SACSCI,with preoperative UEMS being the most important predictor.