Trainable Prompt Steering LLM for Safe Medical Reports
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies for generating medical reports using large language models (LLMs) face challenges in ensuring safety, accuracy, and individualization, as they can produce incorrect and unsafe content, and lack the capability for doctors to provide feedback or for the LLM to learn from mistakes.
Innovation Solution
A computer-implemented machine learning method that generates safe text by creating a trainable prompt using negative and positive influential features of a predicted condition, which is then used to steer a pre-trained large language model (PLLM) to produce safe medical reports. This method includes obtaining patient data, predicting diseases using explainable AI (XAI), and training the prompt to optimize the generation of safe reports.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a pre-trained large language model is used to generate medical reports, then report generation speed and efficiency are improved, but safety and accuracy of the generated content deteriorate due to potential incorrect and unsafe output
Solution Approach 1:
The patent implements a feedback mechanism where doctors can provide corrections to generated medical reports. The system stores these corrections and uses them to retrain the prompt, enabling the model to learn from mistakes and continuously improve safety and accuracy while maintaining generation efficiency
Solution Approach 2:
The patent applies preliminary action by training a specialized prompt on medical knowledge and safety guidelines before the actual report generation process. This pre-trained prompt steers the pre-trained large language model to generate safer and more accurate medical reports from the outset
2Device complexity
If a fixed prompt structure is used for report generation, then system complexity is reduced, but adaptability to individual doctors' needs and patient cases deteriorates
Solution Approach 1:
The patent implements dynamics by making the prompt trainable and adaptable. Instead of a fixed prompt structure, the system allows the prompt to be dynamically adjusted through training on medical knowledge and individual doctor feedback, enabling personalization while maintaining manageable system complexity through automated training processes
3Measurement precision
If extensive training data and computational resources are used to improve model accuracy, then report accuracy improves, but computational resource consumption increases
Solution Approach 1:
The patent applies the taking out principle by extracting and isolating the trainable prompt component from the entire large language model. This allows training to focus on a small, targeted subset of parameters rather than the full model, achieving improved accuracy through specialized training while conserving substantial computational resources by not retraining the entire model
Data Source
AI summary
The present invention provides a computer-implemented, machine learning method for generating safe text. A first portion of a trainable prompt is generated using negative influential features and positive influential features of a predicted condition. A second portion of the trainable prompt is trained to steer a pre-trained large language model (PLLM) to generate the safe text using at least the first portion of the trainable prompt. The method has applications including, but not limited to, use cases in medicine (e.g., digital medicine, personalized healthcare, AI-assisted drug or vaccine development, diagnosis or treatment, disease prediction, etc.), and cyber security.