基于临床特征与影像组学构建慢性非特异性腰痛疼痛程度的预测模型及其应用价值

Construction of a predictive model for chronic nonspecific low back pain based on clinical features and radiomics and its application values

中文关键词:  慢性非特异性腰痛  椎旁肌  影像组学  人工智能  多模态数据整合

英文关键词:Chronic nonspecific low back pain  Paraspinal muscles  Radiomics  Artificial intelligence  Multimodal data integration

中文摘要:

  【摘要】 目的：基于临床特征与影像组学，应用人工智能技术构建并验证慢性非特异性腰痛（chronic nonspecific low back pain，cNLBP）患者疼痛程度的预测模型。方法：回顾性分析2023年5月~2025年5月期间于大连医科大学附属第一医院就诊的cNLBP患者共148例，男62例、女86例；年龄15~78岁。以腰痛视觉模拟量表（visual analog scale，VAS）评分将患者分为轻度疼痛组（52例，VAS评分＜5分）和重度疼痛组（96例，VAS评分≥5分）。通过MRI获取腰椎前凸角，L4椎体横截面积，L4-5双侧小关节方向角，L4-5多裂肌（multifidus，MF）和竖脊肌（erector spinae，ES）的横截面积、脂肪浸润面积、脂肪浸润率、MRI平均灰度值作为临床特征，采用单因素和多因素Logistic回归分析筛选影响疼痛程度的临床特征，构建临床模型。在腰椎MRI上提取MF和ES的影像组学特征，包括：一阶统计特征、基于3D形状特征、灰度共生矩阵、灰度依赖矩阵、灰度级运行长度矩阵、灰度大小区域矩阵、邻域灰度差分矩阵及其小波变换特征、对数-西格玛特征、对数特征、平方特征、平方根特征、局部二值模式特征等13大类特征，通过最小冗余度最大相关性算法和最小绝对收缩和选择算子（least absolute shrinkage and selection operator，LASSO）回归分析筛选出与疼痛分级高度相关的特征。按8∶2的比例将患者数据随机分为训练集（n=119）和测试集（n=29），训练集用于训练模型，测试集用于评估模型的性能和泛化能力。使用多种人工智能算法分别构建针对MF和ES的多个影像组学模型，依据受试者工作特征曲线下面积（area under curve，AUC）、准确度、敏感度、特异度等评价指标对比各模型性能，从而选择各自最优的影像组学模型，将临床模型与最优的MF及ES影像组学模型分别结合，构建对应的联合模型，并通过计算预测模型的AUC、准确度、敏感度、特异度评估模型性能及稳定性，受试者工作特征曲线、校准曲线和决策曲线评估模型预测效能。结果：在上述收集的临床特征中，单因素Logistic和多因素Logistic回归分析显示右侧多裂肌脂肪浸润率（OR=0.758，P<0.05）和左侧多裂肌脂肪浸润率（OR=1.418，P<0.05）是cNLBP患者疼痛程度的独立预测因子，而其他临床特征未呈显著性意义，据此构建了临床模型。在影像组学方面，经LASSO算法筛选出MF的8个关键特征、ES的7个关键特征，分别构建了影像组学模型。选择各自最优的影像组学模型后，进一步构建联合模型，其中基于MF特征的联合模型的AUC为0.8081，基于ES特征的AUC为0.8586，优于单一模型的区分能力和临床实用性。决策曲线分析表明，相较于单一的临床模型，联合模型在多数阈值概率范围内具有更高的净收益；校准曲线分析显示联合模型的预测概率与实际观察结果一致性良好。结论：多裂肌的脂肪浸润等影像学指标与cNLBP患者的疼痛程度密切相关。

英文摘要:

  【Abstract】 Objectives: To construct and validate a predictive model for chronic nonspecific low back pain(cNLBP) based on clinical and radiomics features through artificial intelligence technology. Methods: A total of 148 patients with cNLBP admitted to the First Affiliated Hospital of Dalian Medical University from May 2023 to May 2025 included retrospectively analyzed in this study(62 males, 86 females; aged 15-78 years). According to the visual analog scale(VAS) score of low back pain, the patients were divided into a mild pain group(n=52, VAS score<5) and a severe pain group(n=96, VAS score≥5). Clinical features, including lumbar lordosis angle, L4 vertebral cross-sectional area(CSA), L4-5 bilateral facet joint orientation angle, as well as the CSA, fat infiltration area, fat infiltration rate, and mean gray value of the L4-5 multifidus(MF) and erector spinae(ES), were obtained via MRI examination. Univariate and multivariate logistic regression analyses were employed to screen clinical features affecting pain severity and to construct a clinical model. Radiomics features of the MF and ES were extracted from lumbar MRI, comprising 13 categories: first-order statistics, 3D shape-based features, gray level co-occurrence matrix(GLCM), gray level dependence matrix(GLDM), gray level run length matrix(GLRLM), gray level size zone matrix(GLSZM), neighborhood gray tone difference matrix(NGTDM), and their wavelet transform, log-sigma feature, logarithm, square, square root, and local binary pattern(LBP) features. Key features highly correlated with pain grading were screened using the minimum redundancy maximum relevance(mRMR) method and least absolute shrinkage and selection operator(LASSO) regression. The data were randomly divided into a training set(n=119) and a test set(n=29) at a ratio of 8∶2. The training set was used for model training, while the test set was used to evaluate model performance and generalization ability. Multiple radiomics models for MF and ES were separately constructed using various artificial intelligence algorithms. The optimal radiomics models for MF/ES were selected based on performance metrics such as area under the curve(AUC), accuracy, sensitivity, and specificity. The clinical model was then combined with the optimal MF and ES radiomics models respectively to construct corresponding combined models. Model performance and stability were evaluated by calculating AUC, accuracy, sensitivity, and specificity, while predictive efficacy was assessed by plotting ROC curves, calibration curves, and decision curves. Results: Among the collected clinical features, univariate and multivariate logistic regression analyses showed that the right MF fat infiltration rate(OR=0.758, P<0.05) and left MF fat infiltration rate(OR=1.418, P<0.05) were independent predictors of pain severity in cNLBP patients. A clinical model was constructed based on these findings. In terms of radiomics, 8 key features for MF and 7 key features for ES were selected via the LASSO algorithm to construct respective radiomics models. The further constructed combined model based on MF features achieved an AUC of 0.8081, and the combined model based on ES features achieved an AUC of 0.8586, demonstrating better discrimination ability and clinical utility than single models. Decision curve analysis(DCA) indicated that compared to the single clinical model, the combined models provided higher net benefits across most threshold probability ranges. Calibration curve analysis showed good consistency between the predicted probabilities of the combined models and actual observations. Conclusions: Imaging indicators such as fatty infiltration of the MF are closely related to pain severity in cNLBP patients.