AI Framework Learns Tailored Prompts to Deliver Personalized Healthcare Advice

Large language models are increasingly being deployed in healthcare settings, yet most systems rely on generic medical knowledge without accounting for each patient's specific circumstances. A new research framework aims to bridge that gap by learning individualized prompts that steer language models toward more tailored clinical guidance.

The Problem with Generic Guidance#

Healthcare providers and patients want LLM systems that can analyze personal medical records and offer recommendations suited to their unique situations. However, existing approaches typically treat all patients the same, drawing from broad medical knowledge bases and failing to adapt to individual differences. This one-size-fits-all approach limits the practical value of AI assistance in clinical decision-making.

How Personalized Prompt Learning Works#

The proposed approach, called personalized prompt learning (PPL), constructs custom prompts in two steps. First, it builds initial prompts by combining two types of signals: information directly from each patient's own medical history and "peer-informed" signals drawn from cases of clinically similar patients. These seed prompts are then refined using reinforcement learning, iteratively adjusting them to better match the recommendations that physicians have written for that specific patient.

A key design feature is that PPL operates using standard text-based hard prompts, allowing it to work with proprietary language models without requiring modifications to the underlying systems.

Testing on Real Patient Data#

The team evaluated PPL on obstetrics and gynecology cases from real-world clinical settings. Expert physicians compared the personalized recommendations generated by the system against standard clinical guidance, with PPL prevailing in 97 out of 100 head-to-head assessments. The results suggest the approach successfully produces more individualized, clinically appropriate advice.

Broader Implications#

The framework points toward a pathway for deploying LLMs in healthcare contexts where tailored guidance matters. By learning patient-specific prompts rather than modifying the models themselves, the approach maintains compatibility with commercial systems while improving clinical relevance. The researchers have made code and data publicly available to support broader adoption and further research.