Temporal indexing of medical entity in Chinese clinical notes

Zengjian Liu, Xiaolong Wang, Qingcai Chen, Buzhou Tang, Hua XuBMC Medical Informatics Decis. Mak 2019 · DOI

Abstract

This paper presents a hybrid deep learning approach for temporal indexing of medical entities in Chinese clinical notes. By combining Bidirectional Long Short-Term Memory (Bi-LSTM) networks with Convolutional Neural Networks (CNNs) and a rule-based candidate selection process, the proposed model effectively maps symptoms, diagnoses, and treatments to a unified timeline. The system addresses common challenges in clinical NLP, such as data imbalance and linguistic complexity, significantly outperforming traditional rule-based and standard machine learning methods.

Introduction

Clinical notes are a vital source of longitudinal patient data, containing detailed records of symptoms, diagnoses, and treatments. However, this information is often buried in unstructured text. Temporal indexing—the process of linking medical entities to specific points in time or intervals—is a prerequisite for making this data useful for large-scale analysis. While extensively studied for English clinical notes, Chinese notes present unique linguistic hurdles. A major technical challenge is "data imbalance," where pairing every medical entity with every mentioned time expression creates an excessive number of irrelevant pairs, hindering model learning.

Methodology

The researchers proposed a three-step pipeline to automate temporal indexing:

1. Candidate Selection

To mitigate data imbalance, the system utilizes the inherent structure of clinical notes. By identifying specific sections (e.g., "Chief Complaint" or "Discharge Orders"), the system filters for time expressions likely to be relevant to a given medical entity, drastically reducing the search space.

Flowchart of the candidate selection process using document structure.

2. Temporal Relation Classification (RNN-CNN)

The core of the system is a hybrid deep learning architecture:

  • Bi-LSTM (RNN) Layer: Captures long-distance dependencies and contextual information for each word in a sentence.
  • CNN Layer: Extracts the most salient features from the LSTM’s output to create a robust representation of the relationship between the medical entity and the time expression.
  • Output Layer: Classifies the relationship into four categories: NONE, SIMULTANEOUS (time node), BEFORE, or AFTER (time intervals).

3. Final Indexing

The system selects the time expression with the highest confidence score from the candidate set as the final index for the medical entity.

The main flow for temporal indexing, showing the progression from candidate selection to classification and final indexing.

Data and Evaluation

The study utilized a manually annotated corpus of 563 de-identified Chinese clinical notes (12,611 medical entities and 4,006 temporal expressions). The dataset was split into 413 records for training and 150 for independent testing, with a high inter-rater agreement (Kappa of 0.7130).

Key Results

  • Classification Performance: The RNN-CNN model achieved a best F1-score of 75.97%, outperforming SVM (71.63%) and standard CNN/RNN models (~74.8%).
  • Indexing Accuracy: Under strict criteria, the model reached 71.96% accuracy. An ensemble of models (SVM, CNN, RNN, and RNN-CNN) further improved accuracy to 73.31%.

Discussion and Future Work

The hybrid model demonstrates that RNNs and CNNs are complementary: RNNs excel at understanding sentence flow, while CNNs are superior at identifying specific keyword features. Combining these with document-structure heuristics simplifies the task for the neural network.

However, limitations remain. The model currently struggles with "global" temporal relations—links between entities and time expressions far apart in a document. Additionally, performance on time intervals (BEFORE/AFTER) is lower than on time nodes (SIMULTANEOUS). Future work will focus on capturing long-range document-level links and improving the semantic "directionality" of time for interval indexing.

Summary updated

BibTeX
@article{liu2019temporal,
  author = {Zengjian Liu and Xiaolong Wang and Qingcai Chen and Buzhou Tang and Hua Xu},
  title = {{Temporal indexing of medical entity in Chinese clinical notes}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2019},
  doi = {10.1186/s12911-019-0735-x},
}
Figure