Few-shot Prompting Enables Large Language Models to Translate Professional Medical Reports into Layman's Terms

The research team from Ulm University open-sourced the GISelA project. Through carefully designed few-shot prompting strategies, they proved that large language models can convert complex professional medical reports into patient-friendly lay language, with translation quality comparable to professional human experts. This project provides a breakthrough solution to the information asymmetry problem in medical reports and has significant clinical application value.

Modern medical reports are filled with professional terms, abbreviations, and complex expressions. Ordinary patients find it hard to understand the real meaning and clinical significance of content like "mildly decreased hemoglobin level" or "sinus arrhythmia". This information asymmetry not only affects patients' participation in medical decision-making but may also lead to anxiety, misunderstandings, and reduced treatment adherence. Traditional solutions relying on professional interpretation are costly and difficult to scale, so researchers explored the possibility of automated solutions using large language models.

The core contribution of the GISelA (German to Simple Language for Patients) project is to verify the effectiveness of few-shot prompting strategies in medical text simplification tasks. Compared to zero-shot prompting, it guides the model to understand task requirements and style preferences by presenting high-quality input-output examples. The project open-sourced experimental code (including prompt templates, example selection strategies, and blind evaluation processes).

Successful few-shot prompts need to follow three key principles:

1. Example quality first: Selecting 3-5 high-quality examples works better than random multiple examples;
2. Domain-specific style: Maintain accuracy while reducing cognitive load (e.g., "Imaging examination shows pulmonary infiltrates" is converted to "X-ray shows inflamed areas in your lungs");
3. Consistency constraints: Ensure key information such as values, time, and drug doses are accurate and error-free.

The study used strict blind comparative evaluation: evaluators scored translations from dimensions like accuracy, readability, and completeness without knowing the source (human expert vs. large language model), eliminating evaluation bias. Results showed that model translations based on few-shot prompting reached the level of professional human translations in multiple indicators. This is of great significance for medical environments with limited resources—providing high-quality report interpretation services without increasing labor costs.

GISelA technology is suitable for standardized medical texts such as outpatient reports, discharge summaries, and test reports. When integrated into hospital information systems, patients can get lay versions synchronously, improving the efficiency of doctor-patient communication.

However, deployment faces challenges:

1. Regulatory compliance: Need to meet strict medical device regulations;
2. Responsibility definition: The responsibility for model translation errors needs to be clearly defined;
3. Specialty customization: There are large differences in terminology systems across different specialties, so prompt examples need to be customized accordingly.

This study proves that under the guidance of high-quality examples, general-purpose large language models can adapt to highly specialized domain tasks without expensive domain fine-tuning, lowering the threshold for professional applications. This methodology also has reference significance for scenarios such as legal document simplification and technical document popularization. The core lies in building a high-quality example library and designing prompt templates that effectively activate the model's domain knowledge.

The GISelA project is a successful exploration of AI in the field of medical accessibility, showing that carefully designed prompt strategies can make existing large language models reach a practical level in specific vertical domains. With the maturity of technology and the improvement of regulatory frameworks, such intelligent translation tools are expected to become standard configurations for doctor-patient communication, making medical information truly benefit every patient.