摘要目的:基于人工智能算法构建预测患者脑卒中严重程度的院前无创筛查预测模型，为脑卒中患者及家属提供筛查指导和预警作用，为临床决策提供数据支持。方法:采用回顾性研究方式，从大连医科大学附属第二医院医渡云大数据服务器系统中提取就诊时间为2001年1月1日至2023年7月31日的脑卒中患者（ n＝53?793）的临床信息。结合单因素筛选结果以及神经内科高级职称专家意见确定输入变量，输出变量为入院时反映病情严重程度的美国国立卫生研究院卒中量表（NIHSS）评分。采用Python 3.7构建DeepFM算法模型，并同时构建Logistic回归、CART决策树、C5.0决策树、贝叶斯神经网络、深度神经网络（DNN）等5种数据挖掘模型。将原始数据随机分为80%训练集和20%验证集，分别用于训练和测试模型，调整各模型参数，分别计算6种模型的准确度、敏感度、F指数，进行模型的综合评价；绘制受试者工作特征曲线（ROC曲线）和校准曲线，对比DeepFM模型与其他5种数据发掘模型的预测性能。此外，提取大连市中心医院脑卒中患者（ n＝1?028）的数据进行模型的外部验证。 结果:共筛选出14?015例信息完整的脑卒中患者，其中训练集11?212例，验证集2?803例。经单因素筛选后纳入14个指标用于构建模型，即性别、年龄、复发、肢体障碍、面部问题、言语障碍、头部反应、意识障碍、视觉障碍、呛咳吞咽异常、危险因素、家族史、是否吸烟、是否饮酒。DeepFM模型采用两阶交叉特征，DNN层隐藏层层数为3层，使用Dropout丢弃神经网络中的神经元，Rule作为激活函数，各层采用Dense全连接，目标函数为随机梯度下降，迭代次数为15次，训练参数共133?922个。比较6个模型的预测价值，DeepFM模型的准确度为0.951、敏感度为0.992、特异度为0.814、F指数为0.950，曲线下面积（AUC）为0.916；其他5种数据挖掘模型的准确度在0.771～0.780，敏感度在0.978～0.987，F指数在0.690～0.707，AUC介于0.568～0.639。DeepFM模型的校准曲线较其他5种数据挖掘模型更贴合理想曲线。提示DeepFM模型的预测性能最好。对DeepFM模型进行外部验证，其准确度为0.891，说明模型的泛化性能良好。结论:基于DeepFM构建的院前无创筛查预测模型能够较为准确地预测脑卒中患者严重程度分级，在脑卒中快速筛查和早期临床决策中具有潜在的应用价值。

abstractsObjective:To construct a non-invasive pre-hospital screening model and early based on artificial intelligence algorithms to provide the severity of stroke in patients, provide screening, guidance and early warning for stroke patients and their families, and provide data support for clinical decision-making.Methods:A retrospective study was conducted. The clinical information of stroke patients ( n = 53?793) were extracted from the Yidu cloud big data server system of the Second Affiliated Hospital of Dalian Medical University from January 1, 2001 to July 31, 2023. Combined with the results of single factor screening and the opinions of experts with senior professional titles in neurology, the input variable was determined, and the output variable was the National Institutes of Health Stroke Scale (NIHSS) representing the severity of the disease at admission. Python 3.7 was used to build DeepFM algorithm model, and five data mining models including Logistic regression, CART decision tree, C5.0 decision tree, Bayesian network and deep neural network (DNN) were built at the same time. The original data were randomly divided into 80% training set and 20% test set, which were used to train and test the models, adjust the parameters of each model, respectively calculate the accuracy, sensitivity and F-index of the six models, carry out the comprehensive comparison and evaluation of the model. The receiver operator characteristic curve (ROC curve) and calibration curve were drawn, compared the prediction performance of DeepFM model and the other five algorithms. In addition, the data of stroke patients ( n = 1?028) were extracted from Dalian Central Hospital for external verification of the model. Results:A total of 14?015 stroke patients with complete information were selected, including 11?212 in the training set and 2?803 in the testing set. After univariate screening, 14 indicators were included to construct the model, including gender, age, recurrence, physical impairment, facial problems, speech disorders, head reactions, disturbance of consciousness, visual disorders, abnormal cough and swallowing, high risk factor, family history, smoking history and drinking history. DeepFM model adopted the two-order crossover feature. The number of hidden layers in DNN layer was 3. Dropout was used to discard the neurons in the neural network. Rule was used as the activation function. Each layer used Dense full connection. The objective function was random gradient descent. The number of iterations was 15. There were 133?922 training parameters in total. Comparing the predictive value of the six models showed that the accuracy of DeepFM model was 0.951, the sensitivity was 0.992, the specificity was 0.814, the F-index was 0.950, and the area under the curve (AUC) was 0.916. The accuracy of the other five data mining models were between 0.771-0.780, the sensitivity were between 0.978-0.987, the F-index were between 0.690-0.707, and the AUC were between 0.568-0.639. The calibration curve of the DeepFM model was more aligned with the ideal curve than the other five data mining models. Suggesting that the prediction performance of DeepFM model was the best. External validation was conducted on the DeepFM model, and its accuracy was 0.891, indicating good generalization performance of the model.Conclusion:The pre-hospital non-invasive screening prediction model based on DeepFM can accurately predict the severity grading of stroke patients, and has potential application value in rapid screening and early clinical decision-making of stroke.