Abstract: Extracting structured information from narrative clinical notes is a critical task in medical informatics. This paper proposes a sentence-level sequence labeling approach for detecting medical events in Chinese Electronic Health Records (EHRs). By utilizing a hierarchical architecture combining Convolutional Neural Networks (CNNs) and Conditional Random Fields (CRFs), the authors achieved a peak accuracy of 92.6%. Additionally, a novel "smoothed Viterbi decoder" is introduced, providing a computationally efficient, parameter-free alternative for clinical environments with limited resources.
Introduction
Narrative clinical notes are notoriously difficult to process due to their "noisy" nature, filled with abbreviations, irregular grammar, and non-standard terminology. In Chinese clinical writing, a single medical event—such as a "History of Treatment"—frequently spans several short sentences separated by commas.
Traditional Named Entity Recognition (NER) systems operating at the word level often fail to capture this broader structural context. This study argues that treating clinical notes as a sequence of short sentences is a more natural and effective method for event detection. By shifting the granularity from words to sentences, models can better preserve the integrity of complex medical descriptions.
Methods
The researchers developed an end-to-end deep neural network framework with a hierarchical structure designed to capture both local sentence features and global note context:
- Feature Extractor: This component converts word sequences within short sentences into vector embeddings. The study compared Convolutional Neural Networks (CNNs) and bidirectional Long Short-Term Memory (bi-LSTM) networks for this task.
- Sequence Labeller: This layer predicts labels for the entire note based on the sequence of sentence vectors. The study evaluated three decoders: a standard bi-LSTM, the industry-standard CRF, and a proposed Smoothed Viterbi Decoder.
The Smoothed Viterbi Decoder uses an exponential smoothing factor ($c$) to balance the model's current prediction with global statistical trends derived from the training data. This approach requires no additional parameter training, making it highly efficient.
Dataset
The model was trained and tested on a custom dataset of 2,000 clinical notes from the "History of Present Illness" section of a third-level grade-A hospital in China. Data was annotated at the "short sentence" level across 11 categories, including 10 specific medical events (e.g., Imaging Examination, Pathological Examination) and one "Others" category.
Results and Discussion
The CNN-CRF architecture emerged as the top performer, reaching a peak accuracy of 92.6%. Interestingly, CNNs proved superior to bi-LSTMs as feature extractors for short sentences, likely due to their ability to capture local keyword patterns within brief text segments.
The proposed Smoothed Viterbi Decoder achieved a competitive 90.7% accuracy. Analysis of the smoothing factor $c$ showed that a value of 0.3 optimized the performance, effectively balancing specific sentence features with corpus-wide transition logic.
Detailed evaluation via a confusion matrix reveals high precision in categories like "Description of Symptoms" (0.94) and "General Status" (0.94). However, the matrix also highlights challenges in distinguishing between closely related examination types, such as "Electrocardiogram Examination" occasionally being confused with "Imaging Examination."
Conclusion
This research demonstrates that sentence-level granularity is highly effective for medical event detection in Chinese EHRs. The chronological flow of clinical notes provides vital context that independent classification models miss. While newer Transformer-based models have since become dominant in NLP, the hierarchical approach and the efficient Smoothed Viterbi Decoder remain relevant contributions for specialized clinical applications where computational resources may be limited.
Summary updated