Study cohort
We conducted a prospective, multicenter nonrandomized controlled trial involving five hospitals in Beijing: 1) Peking University People’s Hospital, 2) Peking University Third Hospital, 3) Beijing Friendship Hospital Affiliated to Capital Medical University, 4) Chaoyang Hospital Affiliated to Capital Medical University, and 5) Chinese People’s Liberation Army (PLA) General Hospital. All the hospitals recorded patient information in a database specifically created for SCI cases. Prior to the start of the study, the protocol involving all five hospitals was approved by the ethics committee. Methylprednisolone was administered at the discretion of the treating team according to the recommendations of the NASCIS-2 study9.
Inclusion criteria
1) Age: 16–85 years old, irrespective of sex;
2) Final diagnosis by spine magnetic resonance (MR) imaging;
3) Cervical and thoracic fracture dislocation or without fracture dislocation but combined with spinal cord injury;
4) No other injury involving life, injury severity score < 1618;
5) Receiving surgical decompression
Exclusion criteria
1) History of mental illness and metal allergy;
2) Long-term alcohol abuse and drug abuse;
3) Did not agree to participate in this trial/the legal representative of the patient refuses to sign informed consent;
4) Poor compliance, could not be followed up as required
A total of 249 patients met all inclusion criteria and were included in the study from June 1, 2016, to June 1, 2020. At the same time, we retrospectively included patients with acute SCI at Tianjin Binhai Hospital from June 1, 2016, to June 1, 2020, as the validation sample set. The inclusion and exclusion criteria and data collection were consistent with the prospective study.
Predictor variables
The determination of our predictor variables were based on four main principles: 1) the literature proves that the selected variables are related to the patient's functional outcome; 2) the selected variables are easy to obtain in clinical work; 3) the selected variables have good reliability among doctors; 4) the selected variables cover the three aspects including patient's clinical characteristics, MR imaging and treatment plan. Based on these four principles, we identified a total of 6 predictors with three aspects: 1) clinical characteristics, including age, American Spinal Injury Association (ASIA) Impairment Scale (AIS) at admission, level of injury and baseline ASIA motor
score (AMS); 2) MR imaging, mainly including Brain and Spinal Injury Center (BASIC) score; 3) surgical timing, specifically comparing whether surgical decompression was received within 24 hours or not (Table 1A). All six predictors have demonstrated prognostic significance in relation to long-term functional outcome after SCI6,7,10-14,19-21. A professional orthopedic surgeon conducted physical examinations to identify the patients’ neurologic level of injury, AMS and AIS at admission. MR imaging was the earliest recorded MR result for patients. The MR imaging examinations were performed with a 1.5-Tesla MR scanner (Signa CV/I, GE Healthcare, Milwaukee, WI). We assessed sagittal T2 FSE, sagittal T1, and axial T2 FSE sequences to calculate the BASIC score. Two authors individually and independently assessed the imaging data twice to eliminate intra- and inter-observer bias. The timing of the operation was to truthfully record the time between injury and the operation.
Outcome and follow-up
We assessed the SCIM score at 1 year after the operation as the functional outcome index. The SCIM score is composed of 19 items, with three main domains (Table 1B): self-care (six items, scores range from 0–20); respiration and sphincter management (four items, scores range from 0–40); and mobility (nine items, scores range from 0–40). The SCIM score was first proposed by Cate et al. in 199722 and has now been revised in a third edition23. An international multicenter study found that SCIM has good reliability, validity and practicability in people with SCI24 and is superior to FIM25.
Statistical analysis
XGBoost builds a nonlinear regression prediction model through the method of boosting integrated learning. Compared with other boosting ensemble learning, XGBoost can be used to construct predictive models more efficiently and accurately by performing second-order Taylor expansion, regularization term, and optimizing greedy algorithms on the objective function. We implemented XGBoost through Python 3.9. Since our sample data are relatively small, we choose n_estimators=10000 (CART) in the XGBoost prediction model. A paired-sample T test was used to test the accuracy of the prediction set by IBM SPSS Statistics for Windows, version 26.0 (IBMCorp., Armonk, N.Y., USA).