关键词: 机器学习, 腰椎间盘突出症, 预后, 预测模型, 偏倚风险, 系统评价

Abstract: OBJECTIVE: Based on different algorithms of machine learning, the prediction model of lumbar disc herniation has become a trend and hot spot in the development of precision medicine. However, there is limited evidence on the reporting quality and methodological quality of prediction models of lumbar disc herniation outcomes using machine learning. This article is aimed to explore the performance of machine learning algorithms in predicting the prognosis of lumbar disc herniation by comprehensively analyzing the report quality and risk of bias of previous studies that developed and validated prognosis prediction models based on machine learning through a comprehensive literature search, in order to explore the performance of machine learning algorithms in predicting the prognosis of lumbar disc herniation.
METHODS: The databases of CNKI, WanFang, VIP, SinOMED, PubMed, Web of Science, Embase, and The Cochrane Library were searched by computer. Studies on the use of machine learning to develop (and/or validate) prognostic prediction models for lumbar disc herniation were collected from the inception of the database to December 31, 2023. Two researchers independently screened the literature, extracted data, and assessed the risk of bias of the included studies. The reporting quality and risk of bias of the included studies were assessed by the Multivariable Transparent Reporting of Predictive Models (TRIPOD) statement and the Predictive Model Risk of Bias Assessment Tool (PROBAST). The results of the evaluation were analyzed using descriptive statistics and visual charts. 
RESULTS: (1) A total of 23 articles were included, and the TRIPOD compliance of each study ranged from 11% to 87%, with a median compliance of 54%. The quality of reporting of titles, detailed descriptions of treatment measures, blinding of predictors, handling of missing data, details of risk stratification, specific procedures for enrollment, model interpretation, and model performance was mostly poor, with TRIPOD adherence rates ranging from 4% to 35%. (2) Of all included studies, 61% had a high risk of bias and 39% had an unclear overall risk of bias. The area under the curve, accuracy, sensitivity and specificity were used to evaluate the performance of the model. The areas under the curve of 20 models were reported, ranging from 0.561 to 0.999. Three models reported the accuracy of the model, ranging from 82.07% to 89.65%. (3) Among all included studies, the statistical analysis domain was most often assessed as having a high risk of bias, mainly due to the small number of valid samples, the selection of predictors based on univariate analysis and the lack of calibration and discrimination assessment of the model in the study. 
CONCLUSION: These results indicate that machine learning can achieve good predictive ability in the development and validation of prognostic models for lumbar disc herniation. The commonly used algorithms include regression algorithm, support vector machine, decision tree, random forest, artificial neural network, naive Bayes and other algorithms. Reasonable algorithms combined with clinical practice can improve the accuracy of prognosis prediction of lumbar disc herniation. However, the reporting and methodological quality of prognosis prediction models based on machine learning are poor, the prediction performance of different models varies greatly, and the generalization and extrapolation of research models are unclear. There is an urgent need to improve the design, implementation and reporting of such studies. To promote the application of machine learning in the clinical practice of lumbar disc herniation prediction models, it is necessary to comprehensively consider various predictors related to the prognosis of the disease before modeling, and strictly follow the relevant standards of PROBAST tool during modeling.

Key words: machine learning, lumbar disc herniation, prognosis, prediction model, risk of bias, systematic review