评估机器学习在脊柱畸形管理中的效用与挑战:系统综述
该研究为遵循PRISMA指南并在PROSPERO注册的系统综述,检索PubMed、Embase及Google Scholar自2019年以来将机器学习模型用于脊柱侧凸管理的研究,共纳入63篇全文。最常研究的疾病为青少年特发性脊柱侧凸(n=48)和成人脊柱畸形(n=15),机器学习用途以诊断/分类(n=38)最多,最常用模型为卷积神经网络(n=38);仅16项研究进行外部验证,无研究将模型用于实践。作者认为机器学习在脊柱侧凸管理中前景可期,但仍存在显著挑战,需更大规模、多中心、更多验证及有效实施策略的研究。
要点
- 综述纳入63篇文献,研究最多的疾病为青少年特发性脊柱侧凸(n=48)和成人脊柱畸形(n=15)。
- 机器学习用途依次为诊断/分类(n=38)、手术结局(n=11)、预后(n=7)和风险评估(n=7)。
- 仅16项研究报告外部验证,无研究将模型用于临床实践。
核心结论
- 作者认为机器学习在脊柱侧凸管理中的现状是前景可期的,显示出改善诊断和优化治疗的潜力。
- 作者指出重大挑战仍然存在;为充分理解机器学习的益处,需要更大规模、多中心、更多验证及有效实施策略的研究。
研究设计与样本
- 系统综述,遵循PRISMA指南并在PROSPERO数据库注册。
- 检索PubMed、Embase和Google Scholar数据库,纳入2019年以来将机器学习模型用于脊柱侧凸管理的研究。
- 共纳入63篇经全文审查的文章;研究最多的疾病为青少年特发性脊柱侧凸(n=48)和成人脊柱畸形(n=15),5项研究包含其他疾病,1项未明确说明。
主要发现
- 机器学习用途为诊断/分类(n=38)、手术结局(n=11)、预后(n=7)和风险评估(n=7)。
- 最常用模型包括卷积神经网络(n=38)、随机森林(n=11)和支持向量机(n=9)。
- 仅16项研究报告外部验证,无研究将模型用于实践。
关键数值与推荐
- 纳入标准为2019年以来使用机器学习模型进行脊柱侧凸管理的研究,不符合纳入标准的研究被排除。
- 数据提取聚焦于脊柱侧凸类型、样本量、模型效用、机器学习模型架构、模型开发、训练数据、交叉验证、单中心或多中心研究设计及研究/模型局限性。
临床怎么用
- 原文未据此给出推荐。
局限性
- 常见局限性包括样本量小、伪影遮挡图像、潜在选择偏倚、疾病严重程度多样性有限、排除参数、缺乏患者多样性、准确性随弯曲部位和严重程度变化、回顾性研究及缺乏外部验证。
- 仅16项研究报告外部验证,无研究将模型用于实践。
- 作者指出为充分理解机器学习的益处,需要更大规模、多中心、更多验证及有效实施策略的研究。
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: Systematic review. OBJECTIVE: To review the application of machine learning (ML) in scoliosis management, while addressing the limitations and challenges in implementing ML in scoliosis practice. SUMMARY OF BACKGROUND DATA: ML has revolutionized medical research, particularly in scoliosis management, enabling automated assessment of diagnosis, disease progression, and treatment response. METHODS: A systematic review of the literature was conducted following the PRISMA guidelines and was registered in the PROSPERO database. The literature search included the PubMed, Embase, and Google Scholar databases. Studies since 2019 were included if they used ML models in scoliosis management. Studies that did not meet the inclusion criteria were excluded. Data extraction focused on the scoliosis type, sample size, model utility, ML model architecture, model development, training data, cross-validation, single- or multi-center study design, and study/model limitations. RESULTS: The study included 63 articles that underwent full-text review. The most commonly studied disorders were adolescent idiopathic scoliosis (n=48) and adult spinal deformity (n=15), with five studies including other disorders and one not specifying. The uses of ML were diagnosis/classification (n=38), operative outcomes (n=11), prognosis (n=7), and risk assessment (n=7). The most used models included convolutional neural networks (n=38), random forest (n=11), and support vector machines (n=9). Only 16 studies reported external validation, while none implemented their model in practice. Common limitations reported were small sample size, artifacts obscuring images, potential selection bias, limited variety in disease severity, exclusion of parameters, lack of patient diversity, varying accuracy with curve site and severity, retrospective study, and lack of external validation. CONCLUSION: The current state of machine learning in scoliosis management is promising, showing potential for improving diagnosis and optimizing treatments. However, significant challenges remain. To fully understand the benefits of ML, larger, multi-center studies with more validation and effective implementation strategies are needed.
引用
- Khan JA, Houdane A, Hajja A, et al. Assessing the Utility and Challenges of Machine Learning in Spinal Deformity Management: A Systematic Review. Spine (Phila Pa 1976). 2025. PMID: 39973111. doi:10.1097/BRS.0000000000005236
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。