利用多模态医学数据和机器学习构建脑出血预后预测模型的研究

Research on constructing prognosis prediction model of intracerebral hemorrhage using multimodal medical data and machine learning

目的利用多模态医学数据和机器学习构建脑出血预后预测的模型,并探讨其预测价值。方法回顾性分析华中科技大学协和深圳医院神经外科2020年1~12月收治的98例脑出血患者的临床资料及北京协和医院建立的全国多中心颅内出血数据库2020年1~12月纳入的302例脑出血患者的临床资料。构建脑出血影像学数据库,提取影像组学、临床相关因素标签,构建预测患者预后的模型。另选取华中科技大学协和深圳医院神经外科2021年1~12月收治的100例脑出血患者进行模型前瞻性验证。结果400例患者预后不良的发生率为19.00%,单因素及多因素分析结果显示GCS评分、收缩压、舒张压、血糖、血肿体积、周围水肿体积、纤维蛋白原均是预后的影响因素(P<0.05)。自编码影像特征-临床数据模型预测脑出血患者预后的灵敏度、特异度、准确度、曲线下面积(AUC)[95%置信区间(95%CI)]分别为100.00%、99.38%、99.50%、0.994(0.935~0.998),均高于自编码影像特征模型及传统模型,且自编码影像特征模型均高于传统模型(P<0.05)。经验证,自编码影像特征-临床数据模型预测脑出血患者预后的灵敏度、特异度、准确度、AUC(95%CI)分别为100.00%、97.47%、98.00%、0.974(0.922~0.996)。结论利用多模态医学数据和机器学习构建的自编码影像特征-临床数据模型预测脑出血预后的效能高。

Objective To construct a predictive model of cerebral hemorrhage prognosis prognosis using multimodal medical data and machine learning,and to explore its predictive value.Methods A total of 400 cerebral hemorrhage data was retrospective analyzed,included 98 cerebral hemorrhage patients in the Neurology Department of Peking Union Medical College Shenzhen Hospital,Huazhong University of Science and Technology from January to December 2020 and 302 cerebral hemorrhage patients from the National Multicenter Intracranial Hemorrhage Database established by Peking Union Medical College Hospital from January to December 2020 to build an imaging database of cerebral hemorrhage,extract the labels of imaging omics and clinical related factors,and build a model to predict patients'prognosis.Another 100 patients with intracerebral hemorrhage in the Neurology Department of Peking Union Medical College Shenzhen Hospital,Huazhong University of Science and Technology from January to December 2021 were selected for prospective verification of the model.Results The incidence of poor prognosis in 400 patients was 19.00%.The results of univariate analysis and multivariate logistic regression analysis showed that GCS score,systolic blood pressure,diastolic blood pressure,blood glucose,hematoma volume,peripheral edema volume,fibrinogen were all the influencing factors of poor prognosis(P<0.05).The sensitivity,specificity,accuracy and area under curve(AUC)[95% confidence interval(95%CI)]of self coding image feature clinical data model for predicting prognosis in patients with cerebral hemorrhage were 100.00%,99.38%,99.50% and 0.994(0.935-0.998)respectively,which were higher than those of self coding image feature model and traditional model(P<0.05),and those of the self coding image feature model were higher than those of traditional model(P<0.05).After verification,the sensitivity,specificity,accuracy and AUC(95%CI)of the self coded image feature clinical data model for predicting prognosis in patients with cerebral hemorrhage were 100.00%,97.47%,98.00%and 0.974(0.922-0.996)respectively.Conclusion The self coding image feature clinical data model based on multi-modal medical data and machine learning has high efficiency in predicting cerebral hemorrhage prognosis.