使用机器学习预测脊柱重建手术患者的骨骼健康状况
本研究基于211例接受成人脊柱重建手术且具有双能X线吸收测定扫描的30岁以上患者数据,利用Weka软件构建机器学习模型对健康、骨量减少和骨质疏松进行多分类预测。随机森林模型在训练集上平均敏感度0.81、特异度0.95、AUC 0.96,在测试集上平均敏感度0.64、特异度0.78、AUC 0.69,模型对骨质疏松的预测表现最佳。作者认为机器学习方法可用于预测脊柱重建手术患者的骨骼健康状况,并能发现未识别的风险因素。
要点
- 骨质疏松患病率为23.22%,骨量减少患病率为52.61%。
- 随机森林模型训练集平均敏感度0.81、特异度0.95、AUC 0.96。
- 测试集平均敏感度0.64、特异度0.78、AUC 0.69,模型对骨质疏松预测最佳。
核心结论
- 作者认为,使用机器学习方法预测成人脊柱重建手术患者的骨骼健康状况是可能的。
- 作者认为,使用机器学习进行数据挖掘可以发现成人脊柱重建手术患者中未被识别的骨骼健康风险因素。
研究设计与样本
- 回顾性研究,数据为前瞻性收集。
- 纳入211例年龄超过30岁、具有双能X线吸收测定扫描、接受脊柱重建手术的受试者。
- 通过手动和自动方式从电子健康记录中收集数据,使用Weka软件开发预测模型,对健康、骨量减少和骨质疏松进行多分类。
- 骨骼状态依据世界卫生组织(WHO)标准使用双能X线吸收测定T值进行标记。
- 计算准确度、敏感度、特异度和受试者工作特征曲线下面积(AUC),并在未见数据的测试集上评估模型的泛化能力。
主要发现
- 骨质疏松患病率为23.22%,骨量减少患病率为52.61%。
- 随机森林模型在训练集上达到最佳性能,平均敏感度为0.81,特异度为0.95,AUC为0.96。
- 模型在测试集上平均敏感度为0.64,特异度为0.78,AUC为0.69。
- 模型对骨质疏松患者的预测表现最佳。
- 体重指数、保险类型、血清钠水平、血清肌酐水平、减重手术史以及选择性5-羟色胺再摄取抑制剂等药物的使用等多种患者特征表现出预测价值。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
STUDY DESIGN: Retrospective study of data collected prospectively. OBJECTIVE: The goal of this study is to create a predictive model of preoperative bone health status in adult patients undergoing adult spinal reconstructive (ASR) surgery using machine learning (ML). SUMMARY OF BACKGROUND DATA: Despite understanding that bone health impacts spine surgery outcomes, spine surgeons lack the tools to risk stratify patients preoperatively to determine who should undergo bone health screening. An ML approach mines patterns in data to determine the risk for poor bone health in ASR patients. MATERIALS AND METHODS: Two hundred and eleven subjects over the age of 30 with dual energy X-ray absorptiometry scans, who underwent spinal reconstructive surgery were reviewed. Data was collected by manual and automated collection from the electronic health records. The Weka software was used to develop predictive models for multiclass classification of healthy, osteopenia, and osteoporosis (OPO) bone status. Bone status was labeled according to the World Health Organization (WHO) criteria using dual energy X-ray absorptiometry T scores. The accuracy, sensitivity, specificity, and area under the receiver operating curve (AUC) were calculated. The model was evaluated on a test set of unseen data for generalizability. RESULTS: The prevalence of OPO was 23.22% and osteopenia was 52.61%. The random forest model achieved optimal performance with an average sensitivity of 0.81, specificity of 0.95, and AUC of 0.96 on the training set. The model yielded an averaged sensitivity of 0.64, specificity of 0.78, and AUC of 0.69 on the test set. The model was best at predicting OPO in patients. Numerous patient features exhibited predictive value, such as body mass index, insurance type, serum sodium level, serum creatinine level, history of bariatric surgery, and the use of medications such as selective serotonin reuptake inhibitors. CONCLUSION: Predicting bone health status in ASR patients is possible with an ML approach. Additionally, data mining using ML can find unrecognized risk factors for bone health in ASR surgery patients.
引用
- Shen Y, Sardar ZM, Chase H, et al. Predicting Bone Health Using Machine Learning in Patients undergoing Spinal Reconstruction Surgery. Spine (Phila Pa 1976). 2022. PMID: 36302158. doi:10.1097/BRS.0000000000004511
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。