Imagine sifting through mountains of survey data to uncover hidden health trends. It's a daunting task, but what if AI could lend a hand? Researchers are exploring innovative ways to automate disease phenotyping – essentially identifying and classifying diseases mentioned in survey responses – using a powerful combination of domain-specific models and large language models (LLMs). This approach tackles the challenge of standardizing the vast and varied language people use when discussing their health. The study centers around BERN2, a model specifically designed to recognize and categorize disease mentions in biomedical text. Researchers put BERN2 to the test using data from the ORIGINS birth cohort, a large-scale study focusing on early life determinants of non-communicable diseases. They then supercharged BERN2 by integrating it with LLMs like Mistral and Llama, experimenting with different prompting techniques to refine the accuracy of disease identification. The results? While simply fine-tuning LLMs on survey data yielded limited improvements, a clever combination of few-shot inference (giving the model a few examples to learn from) and retrieval-augmented generation (allowing the model to access external knowledge bases) significantly boosted performance. This suggests that providing LLMs with both structure and context is key to unlocking their full potential. This research opens exciting possibilities for faster, more accurate analysis of survey data, ultimately leading to better understanding of disease prevalence and potential risk factors. While challenges remain in terms of computational resources and the need for larger datasets, the future of automated disease phenotyping looks bright, promising a more efficient way to glean valuable insights from the wealth of information hidden within survey responses.