TL;DR
This paper introduces a novel hybrid model (HyM) for automatically classifying patients' medical specialties based on their symptom descriptions. By integrating features from TF-IDF, BERT, LSTM, and TEXT-CNN, HyM achieves over 93.5% accuracy, significantly outperforming individual models and addressing a critical need for efficient patient triage in China.
What Problem Does This Paper Address?
In China, the healthcare system often lacks sufficient general practitioners, leading patients to directly register with specialists. This requires patients to possess significant medical knowledge to accurately identify the correct specialty for their symptoms. This often results in misdiagnosis, wasted time, and inefficient use of medical resources. The paper aims to solve this by developing an automated system that can accurately direct patients to the appropriate medical specialty based solely on their self-described symptoms. This not only improves patient experience but also enhances the efficiency of medical staff.
Methods
The core of this research is the proposed Hybrid Model (HyM), which combines multiple feature extraction techniques to capture both shallow and deep semantic information from patient symptom descriptions.
- Two-Channel Feature Extraction:
- Channel 1 (Shallow Features): Utilizes TF-IDF (Term Frequency-Inverse Document Frequency) to capture the distinctiveness and importance of words within a symptom description. This method is good for identifying keywords.
- Channel 2 (Deep Features): This channel processes symptom descriptions after they are converted into numerical vectors (using Word2Vec). It then employs three powerful deep learning models:
- BERT (Bidirectional Encoder Representations from Transformers): Leverages pre-trained language understanding to capture contextual and semantic information.
- LSTM (Long Short-Term Memory): A type of recurrent neural network adept at processing sequential data and capturing long-term dependencies in text.
- TEXT-CNN (Text Convolutional Neural Network): A specialized CNN for text that extracts local features and patterns.
- Feature Concatenation and Classification: The features extracted from both channels are concatenated to form a comprehensive representation of the symptom text. This combined feature vector is then fed into a fully connected neural network (FC) with a dropout mechanism for classification into one of the medical specialties.
- Text Preprocessing: Before feature extraction, symptom descriptions undergo preprocessing including Chinese word segmentation using the Jieba library, removal of punctuation, stop words, and special characters. Text length is standardized by padding or truncation.
Datasets
The study utilizes a custom dataset compiled from two sources:
- Online Database: Data from www.39.net, a Chinese online health service platform, comprising over 30,000 symptom descriptions classified into eight common medical departments (e.g., Cardiovascular Diseases, Pediatrics).
- Offline Database: Data from offline hospitals, sourced from www.xywy.com, including approximately 20,000 symptom descriptions also classified into eight departments. These instances were double-checked by doctors.
In total, 48,128 patient description texts were used for analysis, covering eight common medical specialties. This dataset appears to be a newly curated benchmark for this specific problem in the Chinese medical context.
Key Insights and Ideas
- Hybridization is Key: The most significant insight is that combining shallow (TF-IDF) and deep (BERT, LSTM, TEXT-CNN) features leads to superior performance compared to using any single model. This suggests that different models capture complementary aspects of symptom descriptions.
- Context Matters: BERT's ability to understand context and semantics is crucial for accurately interpreting complex symptom descriptions.
- Sequential Information is Valuable: LSTM's strength in handling sequential data helps in understanding the flow and progression of symptoms.
- Local Patterns are Important: TEXT-CNN's ability to identify local patterns and n-grams contributes to feature extraction.
- Addressing a Real-World Problem: The paper tackles a practical and pressing issue in healthcare systems, particularly in regions with a shortage of general practitioners, offering a tangible solution.
Key Results
The proposed HyM model achieved an impressive accuracy of over 93.5%. This significantly outperforms:
- TEXT-CNN: 88.7% accuracy
- LSTM: 81.3% accuracy
- BERT: 90.5% accuracy
The paper also highlights that HyM demonstrates high generalization capacity and stability across various test areas, with some achieving up to 97.3% accuracy. The F-scores for all eight labels were consistently above 90%, indicating strong performance in both precision and recall.
Limitations and Future Work
- Limitations:
- The study focuses on Chinese language symptom descriptions.
- The dataset, while substantial, is limited to eight common medical specialties.
- The paper acknowledges that while the model is accurate, it's still a step towards a more comprehensive system.
- Future Work:
- Develop a more colloquial dataset to better match patient language.
- Investigate the role of attention mechanisms and hyperparameter tuning further.
- Incorporate additional data types such as disease symptom texts, medical numerical data, and medical dictionaries to build a more comprehensive categorization model.
- Expand the model's application to network interrogation, triage, intelligent hospital guidance, and medical text data mining.
Summary updated