Department of Orthopedics, Division of Spine Surgery, Brown University Warren Alpert Medical School, Providence, RI.
Spine (Phila Pa 1976). 2018 Aug 1;43(15):1058-1066. doi: 10.1097/BRS.0000000000002515.
Retrospective cohort study.
Blood transfusion is frequently necessary after adult spinal deformity (ASD) surgery. We sought to develop predictive models for blood transfusion after ASD surgery, utilizing both classification tree and random forest machine-learning approaches.
Past models for transfusion risk among spine surgery patients are disadvantaged through use of single-institutional data, potentially limiting generalizability.
This investigation was conducted utilizing the American College of Surgeons National Surgical Quality Improvement Program dataset years 2012 to 2015. Patients undergoing surgery for ASD were identified using primary-listed current procedural terminology codes. In total, 1029 patients were analyzed. The primary outcome measure was intra-/postoperative blood transfusion. Patients were divided into training (n = 824) and validation (n = 205) datasets. Single classification tree and random forest models were developed. Both models were tested on the validation dataset using area under the receiver operating characteristic curve (AUC), which was compared between models.
Overall, 46.5% (n = 479) of patients received a transfusion intraoperatively or within 72 hours postoperatively. The final classification tree model used operative duration, hematocrit, and weight, exhibiting AUC = 0.79 (95% confidence interval 0.73-0.85) on the validation set. The most influential variables in the random forest model were operative duration, surgical invasiveness, hematocrit, weight, and age. The random forest model exhibited AUC = 0.85 (95% confidence interval 0.80-0.90). The difference between the classification tree and random forest AUCs was nonsignificant at the validation cohort size of 205 patients (P = 0.1551).
This investigation produced tree-based machine-learning models of blood transfusion risk after ASD surgery. The random forest model offered very good predictive capability as measured by AUC. Our single classification tree model offered superior ease of implementation, but a lower AUC as compared to the random forest approach, although this difference was not statistically significant at the size of our validation cohort. Clinicians may choose to implement either of these models to predict blood transfusion among their patients. Furthermore, policy makers may use these models on a population-based level to assess predicted transfusion rates after ASD surgery.
回顾性队列研究。
成人脊柱畸形（ASD）手术后经常需要输血。我们试图利用分类树和随机森林机器学习方法为 ASD 手术后的输血建立预测模型。
过去用于脊柱手术患者输血风险的模型存在使用单一机构数据的劣势，这可能限制了其普遍性。
本研究利用美国外科医师学会国家手术质量改进计划数据集 2012 年至 2015 年的数据进行。使用主要列出的当前程序术语代码识别接受 ASD 手术的患者。共分析了 1029 名患者。主要观察指标是术中/术后输血。患者被分为训练（n=824）和验证（n=205）数据集。建立了单分类树和随机森林模型。使用验证数据集上的接收器工作特征曲线（AUC）下面积来测试这两个模型，然后比较模型之间的 AUC。
总体而言，46.5%（n=479）的患者在手术期间或术后 72 小时内接受了输血。最终的分类树模型使用手术时间、血细胞比容和体重，在验证数据集上的 AUC 为 0.79（95%置信区间 0.73-0.85）。随机森林模型中最具影响力的变量是手术时间、手术侵袭性、血细胞比容、体重和年龄。随机森林模型的 AUC 为 0.85（95%置信区间 0.80-0.90）。在验证队列大小为 205 名患者时，分类树和随机森林 AUC 之间的差异无统计学意义（P=0.1551）。
本研究产生了 ASD 手术后输血风险的基于树的机器学习模型。随机森林模型的 AUC 表明其具有很好的预测能力。我们的单分类树模型在实施方面具有优势，但与随机森林方法相比，AUC 较低，尽管在我们的验证队列大小上，这种差异没有统计学意义。临床医生可以选择在他们的患者中使用这些模型中的任何一个来预测输血。此外，决策者可以在基于人群的水平上使用这些模型来评估 ASD 手术后的预测输血率。
3。
