Zhao Peng, Pan Wencai, Zou Xin, Yang Jiaqing, Zhang Shihui, Liu Yufei, Li Yang
Chongqing Hospital, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, No. 799, Liangjiang Avenue, Longxing Town, Yubei District, Chongqing, 401121, China.
Department of Medical Engineering, Xinqiao Hospital Army Medical University, Chongqing, 400037, China.
Biomed Eng Online. 2025 May 20;24(1):62. doi: 10.1186/s12938-025-01394-5.
Rapid prehospital assessment of hemorrhagic shock severity is critical for trauma triage and intervention. Current non-invasive single-parameter monitoring shows limited diagnostic reliability. We developed a multi-parameter predictive model integrating mean arterial pressure (MAP), buccal mucosal CO₂ (PCO₂), transcutaneous oxygen (PO₂), and pulse pressure variation (PPV). using K-nearest neighbors (KNN) algorithm.
Forty-five Wistar rats were randomly divided into five groups (n = 9) with different blood loss amounts. MAP, PCO, PO, and PPV measurements were continuously obtained. A multi-parameter shock severity prediction model was established based on the KNN algorithm. Leave-one-out cross-validation was used to determine the value of K. Meanwhile, a prediction model based on the support vector machine (SVM) algorithm was established. The performance of the two prediction models was compared using confusion matrices and receiver operating characteristic (ROC) curve.
When the training vs testing data set ratio is 7:3 or 6:4, and K = 3, the KNN-based model has the best prediction accuracy (94.82% and 93.51%). The confusion matrix and ROC evaluation showed that the overall performance of the KNN-based model is superior to that of the SVM-based model, at all levels of blood loss (F1 = 95.09% and 93.99%, AUC = 1 and 0.97 for the KNN-based model at 7:3 and 6:4 dataset ratio; F1 = 83.84% and 84.86%, AUC = 0.97 and 0.97 for the SVM-based model at 7:3 and 6:4 dataset ratio).
Using the detection indicators MAP, PCO, PO and PPV, the KNN-based rat hemorrhagic shock severity prediction model has high accuracy and stability, and demonstrates potential feasibility for severity stratification of hemorrhagic shock in standardized preclinical models, providing a foundation for future clinical validation in prehospital environments.
对出血性休克严重程度进行快速的院前评估对于创伤分诊和干预至关重要。目前的非侵入性单参数监测显示出有限的诊断可靠性。我们使用K近邻（KNN）算法开发了一种整合平均动脉压（MAP）、颊黏膜二氧化碳分压（PCO₂）、经皮氧分压（PO₂）和脉压变异度（PPV）的多参数预测模型。
将45只Wistar大鼠随机分为五组（每组n = 9），每组有不同的失血量。连续获取MAP、PCO₂、PO₂和PPV测量值。基于KNN算法建立多参数休克严重程度预测模型。采用留一法交叉验证来确定K值。同时，建立基于支持向量机（SVM）算法的预测模型。使用混淆矩阵和受试者工作特征（ROC）曲线比较两种预测模型的性能。
当训练集与测试集比例为7:3或6:4且K = 3时，基于KNN的模型具有最佳预测准确性（分别为94.82%和93.51%）。混淆矩阵和ROC评估表明，在所有失血量水平上，基于KNN的模型的整体性能优于基于SVM的模型（在7:3和6:4数据集比例下，基于KNN的模型的F1值分别为95.09%和93.99%，AUC值分别为1和0.97；基于SVM的模型在7:3和6:4数据集比例下的F1值分别为83.84%和84.86%，AUC值分别为0.97和0.97）。
使用检测指标MAP、PCO₂、PO₂和PPV，基于KNN的大鼠出血性休克严重程度预测模型具有较高的准确性和稳定性，并在标准化临床前模型中显示出对出血性休克严重程度分层的潜在可行性，为未来在院前环境中的临床验证奠定了基础。
