The establishment of a decision tree model for the individualized treatment of spinal metastases based on RPA

Abstract

Objective To evaluate the prognostic factors of spinal metastases by recursive partitioning analysis (RPA) and establish a decision tree model that can guide clinicians to select individualized treatment. Methods Between January 2011 and December 2015, three institutional databases were searched to identify 169 patients with metastatic spinal tumors underwent surgery. The ratio of male and female was 1.48∶1 (102 males and 67 females). The average age was 59.2±11.1 years. One-hundred eighteen cases of patients were randomly selected as training samples and the remaining 51 cases were verified samples. Preoperative factors were collected and analyzed by RPA methods, including primary tumor, KPS score, Frankel grade, gender, age, visceral metastasis, bone metastasis, spinal metastasis, blood glucose, blood pressure, surgery site, symptoms, surgery interval, serum albumin level and other risk factors. Pearson's Chi-square test was performed to judge the segmentation point of the decision tree model. The decision tree model is built using the Chi-square automatic interaction detector (CHIAD) algorithm and sensitivity and specificity was automatically calculated. Results The median postoperative survival time was 12.6±1.2 months [95%CI(10.1, 15.0)]. 70% were randomly selected as the experimental group (118 cases), and 30% were the verification group (51 cases). The sensitivity was 96.9%. The specificity was 89.8%. The Kappa coefficient was 0.874 in the experimental group. The sensitivity was 95.4%. The specificity was 90.8%. The Kappa coefficient was 0.810 in the test group. The prognostic factors (weight from high to low) based on RPA were Frankel grade (F=8.132, P=0.005), the primary tumor and the KPS score (Equal, F=9.871, P=0.000 and F=11.945, P=0.003), serum albumin and movement time (Equal, F=7.566, P=0.018 and F=9.966, P=0.008). The decision tree model consists of 7 types. Survival time was 51 months, 18 months, 13 months, 8 months, 4 months, 5 months, 9 months in the class I~VII, respectively. According to the difference of RPA survival time, the operation was classified as 3 grade. Total spinal resection was regarded as grade 1, including class I and class II, with an average survival time of over 18 months. Limited operation was regarded as grade 2, including class III, V and VII with survival time in 6-18 months. The conservative treatment was regarded as grade 3, including class Ⅳ and Ⅵ with the survival time was less than 6 months. Conclusion The decision tree model based on RPA for predicting the survival time of spinal metastases can not only identify the prognostic factors, but also classify and grade various prognostic factors; the decision tree model is simple and can guide clinicians to choose the best surgical plan by predicting the survival time. Key words: Methods; Decision trees; Multicenter study; Individualized medicine