基于机器学习预测创伤性颈脊髓损伤患者的神经功能状态

Prediction of neurological function status in patients with traumatic cervical spinal cord injury based on machine learning

中文关键词:  创伤性颈脊髓损伤  机器学习  神经功能预测  集成技术  ASIA评分

英文关键词:Traumatic cervical spinal cord injury  Machine learning  Neurological function prediction  Integrated technology  ASIA scores

中文摘要:

  【摘要】 目的：提出一种基于机器学习预测创伤性颈脊髓损伤（traumatic cervical spinal cord injury，TCSCI）患者神经功能状态的方法。方法：回顾性分析上海长征医院收治的180例TCSCI患者的临床资料，包括患者伤后24h内的颈椎MRI图像和美国脊髓损伤协会（American Spinal Injury Association，ASIA）评分及伤后1年随访时的ASIA评分。采用完全随机法将180例患者按8∶2的比例分为训练集144例和测试集36例。整体上，利用两阶段集成理念提出临床-影像预测新方法，即利用TCSCI伤后24h内的ASIA评分和MRI图像实现伤后1年时患者感觉和运动功能的全特征预测。第一阶段，采用GradentBoosting、GaussianNB、KNeighbors、 DecisionTree、RandomForest和support vector classifier模型分别独立预测132个皮节点和肌节点的感觉运动功能恢复情况。第二阶段，通过对性能指标的横向和纵向比较分别筛选出针对每个特征预测的最优模型，从而最后实现56个轻触觉皮节点、56个针刺觉皮节点和20个关键肌节点处的神经功能的最佳预后预测。构建的预测模型经训练和验证完成后，使用准确率、精确率、召回率、平均精度和F1值对测试集的预测进行评估。结果：从该预测模型在TCSCI患者伤后1年时感觉运动功能预测的整体表现来看，所有模型在测试集中预测的准确率均达到0.886以上，召回率达到0.845以上，精确率达到0.875以上，平均精度达到0.853以上，F1值达到0.859以上，说明各模型的正确预测能力和实际预测结果的质量和完整性都比较高。且两阶段预测模型能够实现各个模型对每个特征的预测效果的优选，预测性能更好。结论：感觉运动全特征预测方法可以有效预测TCSCI患者伤后1年随访时的神经功能恢复情况，其预测能力明显高于单一模型，有望为TCSCI患者的个性化诊疗和康复提供有益的借鉴。

英文摘要:

  【Abstract】 Objectives: To propose a method based on machine learning to predict the neurological functional status of patients with traumatic cervical spinal cord injury(TCSCI). Methods: The clinical data of 180 patients with TCSCI admitted to Shanghai Changzheng Hospital were retrospectively analyzed, including cervical spine MRI images and American Spinal Injury Association(ASIA) scores within 24 hours after injury and the ASIA scores at 1-year follow-up after injury. The 180 patients were randomly divided into a training set of 144 patients and a test set of 36 patients in a ratio of 8∶2. Overall, a new clinical-imaging prediction method was proposed using the two-stage integration concept, which used the ASIA scores and MRI images within 24 hours after TCSCI to achieve a full-feature prediction of the patient′s sensory and motor function one year after injury. In the first stage, models such as GradentBoosting, GaussianNB, KNeighbors, DecisionTree, RandomForest, and support vector classifier were used to independently predict the sensory-motor function recovery of 132 skin nodes and muscle nodes. In the second stage, the optimal model for each feature prediction was screened out through horizontal and vertical comparison of performance indicators, so as to finally achieve the best prognostic prediction of neurological function at 56 light touch and 56 acupuncture skin nodes, and 20 key muscle nodes. After the constructed prediction model is trained and verified, the prediction of the test set is evaluated using accuracy, precision, recall, average precision and F1 score. Results: In terms of the overall performance of this prediction model in predicting sensory-motor function in TCSCI patients 1 year after injury, all models in the test set achieved accuracy≥0.886, recall≥0.845, precision≥0.875, average precision≥0.853, and F1 score≥0.859, demonstrating that the correct prediction ability of each model and the quality and completeness of the actual prediction results were relatively high. Moreover, the two-stage prediction model can optimize the prediction effect of each model on each feature, and the prediction performance is better. Conclusions: This sensory-motor full-feature prediction method can effectively predict the neurological function recovery of TCSCI patients at 1-year follow-up after injury. Its predictive ability is significantly higher than that of a single model, and is expected to provide a useful reference for the personalized diagnosis, treatment and rehabilitation of TCSCI patients.