TL;DR
This paper investigates how well current AI models can summarize doctor-patient conversations when trained on one type of data but tested on another (out-of-domain). They found that models struggle with "objective" medical information when not specifically trained for it, and while SOAP-formatted summaries help, they can lead to "hallucinations" (generating made-up medical facts).
What Problem Does This Paper Address?
Summarizing doctor-patient conversations into clinical notes is crucial for healthcare efficiency and patient care. However, creating accurate summaries is challenging due to specialized medical jargon and the need to structure information into specific sections like Subjective (patient's report), Objective (doctor's observations), Assessment (diagnosis), and Plan (treatment). While many AI models exist for this, they often perform poorly when applied to data different from what they were trained on. This "out-of-domain" problem is particularly relevant because collecting large, in-domain datasets is difficult due to privacy concerns. This paper aims to understand why these models fail on out-of-domain data and how to improve their robustness.
Methods
The researchers evaluated two main configurations of summarization models:
- General Summarization Model: These models aim to produce a summary without explicitly trying to create distinct SOAP sections.
- SOAP-Oriented Model: These models are designed to generate summaries with explicit Subjective, Objective, Assessment, and Plan sections.
They analyzed state-of-the-art fine-tuned language models (like FLAN-T5 and LED) and large language models (GPT-3.5 and GPT-4) on these configurations. To understand the root causes of performance drops, they conducted analyses including Linguistic Inquiry and Word Count (LIWC) to examine word distribution and emotional content in notes, and a hallucination analysis to count generated medical terms not present in the original dialogue.
Datasets
The study used two publicly available doctor-patient conversation datasets:
- MTS-Dialog: This dataset features shorter conversations and notes that are typically focused on a single category (e.g., just subjective information).
- ACI-BENCH: This dataset contains longer conversations and notes that include all SOAP sections.
The researchers evaluated models in cross-dataset settings, meaning models trained on one dataset were tested on the other.
Key Insights and Ideas
- Out-of-Domain Performance Drop: Fine-tuned language models, while performing well on their training data, suffer a significant performance drop when applied to out-of-domain datasets. This suggests they tend to "overfit" to their training data.
- The Challenge of "Objective" Information: General summarization models struggle to retain "objective" medical information (like numerical data from exams or labs). This information is often crucial but can be easily omitted by models not specifically guided to capture it.
- SOAP Structure Helps, But Isn't Perfect: Models designed to output SOAP sections generally perform better than general models, especially in retaining information across different categories. However, this structured approach can lead to "hallucinations," where models invent medical terms or facts, particularly in the Assessment and Plan (AP) sections.
- Format Mismatch Isn't the Main Culprit: Analysis using LIWC suggests that differences in word distribution or format between datasets are unlikely to be the primary reason for performance decline. The models' inability to generalize core medical concepts seems to be the main issue.
- GPT Models Show Lower Hallucination Rates: While fine-tuned models like LED can hallucinate significantly, GPT-based models generally exhibit lower hallucination rates, though they might still invent information, especially in the AP section.
Key Results
- ACI-BENCH Performance: Table 4 shows that models evaluated on ACI-BENCH (which has full SOAP notes) experienced a performance decrease when tested out-of-domain. General models performed particularly poorly on objective notes, with performance drops of up to 71% for FLAN-T5. SOAP-oriented GPT models showed better performance, with GPT-4 SOAP achieving only a 2% drop in aggregate score compared to the in-domain baseline.
- MTS-Dialog Performance: Table 5, evaluating on MTS-Dialog (shorter, single-category notes), also revealed performance declines for out-of-domain models. The general model again struggled with objective notes.
- Hallucination Ratios: Figure 6 indicates that the LED model had a notable hallucination problem, especially in objective sections. SOAP-oriented GPT models had higher hallucination ratios in the AP section compared to other sections.
Limitations and Future Work
- Data Imbalance: The MTS-Dialog dataset is unbalanced, with most references focusing on subjective information, which could affect model training.
- Dataset Representation: The study acknowledges that two datasets cannot fully represent the complexity and diversity of real-world doctor-patient conversations across various medical specialties.
- Model Variability: The performance of GPT models can vary between runs and is sensitive to prompt engineering.
- Future Work: The authors suggest that adaptations are essential for general models to preserve objective information, potentially by treating objective note generation as a distinct task. They also highlight the need to further examine the assessment and plan sections for hallucination issues and explore ways to improve robustness for real-world applications.
Summary updated