基于人工智能的高血压性脑出血医疗文本信息自动识别系统

AI-based automatic recognition system of medical text for hypertensive intracerebral hemorrhage

摘要:

目的 构建基于人工智能的高血压性脑出血医疗文本信息自动识别系统，快速识别和分析患者临床信息，高效地输出正确的诊疗方案。方法 基于国内外最新高血压性脑出血诊疗指南，经多位高年资神经外科医生和专业人工智能团队共同讨论，构建基于语言表征模型和专家模块的高血压性脑出血医疗文本信息自动识别及决策系统（即H系统）。随后将收集到的高血压性脑出血病例分为训练集、测试集和验证集，以数据库中病例的真实治疗方案为金标准，先总体评价H系统的准确性，再将其与神经外科医生进行对比，分析H系统的判读效率。结果 在测试集中，H系统所输出的治疗方案的准确率为94.0%（91.5%~96.5%），特异度为91.8%（86.3%~97.3%），灵敏度为95.5%（89.3%~98.2%），曲线下面积（area under the curve，AUC）值为0.936（0.922~0.950）（P=0.000）；在验证集中，H系统所输出的治疗方案的准确率为93.3%（89.5%~97.1%），特异度为 89.9%（83.4%~96.4%），灵敏度为95.8%（92.3%~99.3%），AUC值为0.928（0.891~0.966）（P=0.000）。在处理同样的70例病例时，H系统用时（334.60±4.46）s，而神经外科医生用时（12 550.28±95.45）s；在50 min内，H系统处理的病例数为（383±3）例，而神经外科医生处理的病例数为（11±4）例。结论 本研究所构建的H系统能够对高血压性脑出血患者的急诊病例进行自动识别和分析，并快速输出准确的诊疗方案，可协助医生对高血压脑出血进行急诊诊疗。

Abstract:

Objective To design an automatic medical text recognition system based on artificial intelligence（AI）， which can quickly spot and analyze patients’ clinical information then output accurate treatment plans.Methods After discussions by senior neurosurgeons and professional AI teams，we designed the automatic recognition and decision-making system of medical text information for hypertensive cerebral hemorrhage（namely H-system） based on the language representation model and the expert mode. Taking the real treatment plans in the database as the gold standard， the accuracy of the H-system was evaluated as a whole，and then it was compared with the senior neurosurgery to analyze the efficiency of the H-system.Results In the testing sets，the accuracy of the output by H-system was 94.0%（91.5%-96.5%），the specificity was 91.8%（86.3%-97.3%），the sensitivity was 95.5%（89.3%-98.2%），and the AUC was 0.936（0.922-0.950）（P=0.000）. Meanwhile，in the validation sets，the accuracy of the output by H-system was 93.3%（89.5%-97.1%），the specificity was 89.9%（83.4%-96.4%），the sensitivity was 95.8%（92.3%-99.3%），and the AUC was 0.928（0.891-0.966）（P=0.000）. In processing the same 70 cases，the H-system took（334.60±4.46） s compared to （12 550.28±95.45） s for the neurosurgeon；in 50 minutes， the H-system processed（383±3） cases compared to （11±4） cases for the neurosurgeon.Conclusion The H-system constructed in this study can automatically recognize and analyze the medical text data of patients， and quickly output an accurate treatment plan. In clinical practice， it may assist doctors to provide an efficient and reliable treatment for patients with hypertensive intracerebral hemorrhage.