Institute for Medical Informatics, Statistics and Documentation, Medical University of Graz, Austria.
Department of Information and Process Management, Steiermärkische Krankenanstaltengesellschaft m.b.H. (KAGes), Graz, Austria.
Stud Health Technol Inform. 2023 May 18;302:788-792. doi: 10.3233/SHTI230267.
Clinical information systems have become large repositories for semi-structured and partly annotated electronic health record data, which have reached a critical mass that makes them interesting for supervised data-driven neural network approaches. We explored automated coding of 50 character long clinical problem list entries using the International Classification of Diseases (ICD-10) and evaluated three different types of network architectures on the top 100 ICD-10 three-digit codes. A fastText baseline reached a macro-averaged F1-score of 0.83, followed by a character-level LSTM with a macro-averaged F1-score of 0.84. The top performing approach used a downstreamed RoBERTa model with a custom language model, yielding a macro-averaged F1-score of 0.88. A neural network activation analysis together with an investigation of the false positives and false negatives unveiled inconsistent manual coding as a main limiting factor.
临床信息系统已经成为半结构化和部分注释的电子健康记录数据的大型存储库，这些数据已经达到了一个关键数量，使得它们对于有监督的数据驱动的神经网络方法很有意义。我们探索了使用国际疾病分类（ICD-10）对 50 个字符长的临床问题列表条目的自动编码，并在 100 个 ICD-10 三位数代码上评估了三种不同类型的网络架构。FastText 基线达到了 0.83 的宏平均 F1 分数，其次是字符级别的 LSTM，其宏平均 F1 分数为 0.84。表现最好的方法是使用带有自定义语言模型的下游 RoBERTa 模型，其宏平均 F1 分数为 0.88。神经网络激活分析以及对假阳性和假阴性的研究揭示了不一致的手动编码是一个主要的限制因素。
