Project: Patient-Friendly Clinical Text Summarization
Exploring transformer-based summarization models to improve patient comprehension of complex clinical notes.
Overview
This project investigates transformer-based summarization approaches—specifically T5, BART, and PEGASUS—to generate patient-friendly clinical text from complex medical notes. The focus is on serving low-health-literacy populations, ensuring that summaries retain both informativeness and readability.
Key Features
- Evaluation of state-of-the-art transformer summarization models in the clinical domain.
- Dual evaluation objective:
- Automatic metrics: ROUGE, BLEU.
- Readability metrics: Flesch-Kincaid Grade Level, SMOG Index.
- Experiments conducted on real-world and semi-synthetic clinical datasets.
- Special emphasis on balancing semantic fidelity with patient comprehension.
Approach
- Model Selection
- Fine-tuned T5, BART, and PEGASUS for abstractive summarization.
- Leveraged domain-specific preprocessing to handle medical terminology.
- Dual Evaluation Objective
- Informativeness assessed using ROUGE and BLEU scores.
- Readability assessed using Flesch-Kincaid and SMOG indexes.
- Datasets
- Real-world clinical notes.
- Semi-synthetic patient education datasets.
- Analysis
- Compared readability and informativeness across models.
- Identified trade-offs between lexical simplification and information preservation.
Results & Evaluation
- BART achieved the best balance between informativeness and readability.
- PEGASUS produced highly faithful summaries but often retained technical jargon.
- T5 excelled at simplifying text but occasionally omitted key clinical details.
| Model | ROUGE-L | BLEU | Flesch-Kincaid ↓ | SMOG ↓ |
|---|---|---|---|---|
| T5 | 0.xx | 0.xx | x.xx | x.xx |
| BART | 0.xx | 0.xx | x.xx | x.xx |
| PEGASUS | 0.xx | 0.xx | x.xx | x.xx |
(Lower readability scores = simpler, more patient-friendly text)
Key Insights
- Readability scores often inversely correlated with informativeness metrics.
- Controlled generation and lexical simplification techniques are promising next steps.
- Balancing semantic fidelity with ease of understanding remains the main challenge.
Codebase
Conclusion
This work highlights the potential of transformer-based summarization in improving patient comprehension of complex clinical documents. Future work will focus on controlled text generation, vocabulary simplification, and integrating domain adaptation for specialized medical contexts.