文献速递医学人工智能与信息学流行病学与生物统计

骨科机器学习模型报告不完整:TRIPOD与TRIPOD+AI报告完整性系统综述

Incomplete reporting persists in orthopedic machine learning models: a systematic review of TRIPOD and TRIPOD+AI reporting completeness.
PubMed 文献 · J Clin Epidemiol · 2026-07-31

结果显示模型数量快速增长但报告缺口显著,TRIPOD+AI完整性中位数仅45%,样本量论证与性能报告尤其不足。

要点

  • 骨科ML预后模型研究快速增长,近半数发表于2023-2024年。
  • TRIPOD+AI报告完整性中位数仅45%(IQR 40-50)。
  • 仅36%开发研究报告校准,13%提供可访问模型。

核心结论

  • 尽管骨科ML预后预测模型超过400个且快速增长,但报告完整性存在显著缺口,样本量论证和性能报告方面尤需改进。

研究设计与样本

  • 从433项符合条件的研究提取特征与趋势,并在25本高影响力骨科期刊的93项研究子集中评估TRIPOD+AI报告完整性。
  • TRIPOD 2015完整性由TRIPOD+AI条目推导,并与既往系统综述比较。

主要发现

  • 近半数研究发表于2023-2024年,27%的开发研究存在高样本量担忧。
  • 36%的研究报告了校准,仅13%提供了可访问的模型。
  • TRIPOD+AI完整性中位数为45%(IQR 40-50)。
  • 开发研究的TRIPOD 2015完整性中位数与既往综述相似(54% [IQR 28-84] vs 50% [IQR 27-79]),评估研究则提高(80% [IQR 45-95] vs 61% [IQR 43-90])。

临床怎么用

  • 骨科ML预后模型研究者应优先遵循TRIPOD+AI报告标准,尤其完善样本量论证与性能报告。

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

BACKGROUND AND OBJECTIVES: Machine learning (ML) prognostic models in orthopedic surgery are published at an accelerating pace, yet whether reporting meets the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) + artificial intelligence (AI; 2024) transparency standards remains unclear. We examined (1) characteristics and trends of preoperative ML prognostic models for perioperative outcomes through December 31, 2024; (2) TRIPOD+AI reporting completeness among the 25 highest-impact orthopedic journals; and (3) how derived TRIPOD 2015 completeness compares with prior systematic reviews. METHODS: We performed a systematic review of studies published through December 31, 2024, developing or externally evaluating preoperative ML prognostic models predicting any intra- or post-operative outcome in orthopedic surgery. Study characteristics and trends were extracted from 433 eligible studies. TRIPOD+AI reporting completeness was assessed in a subset of 93 studies (80 development-applicable and 20 evaluation-applicable; not mutually exclusive as 7 studies performed both) published in the 25 highest-impact orthopedic journals. TRIPOD 2015 reporting completeness was derived from TRIPOD+AI items. RESULTS: Nearly half of all studies were published in 2023-2024, a total of 27% of development studies had high sample size concern, 36% reported calibration, and only 13% provided an accessible model. The median TRIPOD+AI completeness was 45% (IQR 40-50). The median TRIPOD 2015 completeness was similar in development studies (54% [IQR 28-84] vs 50% [IQR 27-79]) but improved in evaluation studies (80% [IQR 45-95] vs 61% [IQR 43-90]) compared with prior systematic reviews. CONCLUSION: Despite rapid growth with more than 400 ML orthopedic prognostic prediction models, this baseline assessment reveals substantial gaps in reporting completeness. There is considerable room for improvement, particularly in sample size justification and performance reporting. Future work should prioritize TRIPOD+AI implementation, journal enforcement, and periodic reevaluation of reporting completeness as post-2024 publications mature.

来源:J Clin Epidemiol,查看原文。版权归原作者所有。

引用

  1. Kuijten RH, de Groot TM, van Weezenbeek MA, et al. Incomplete reporting persists in orthopedic machine learning models: a systematic review of TRIPOD and TRIPOD+AI reporting completeness. J Clin Epidemiol. 2026. PMID: 42537858. doi:10.1016/j.jclinepi.2026.112438

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。