Medical Note Transcription Using AI Chunking and LLM Summarization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems in medical settings face challenges due to immature technology and high recognition error rates, limiting their widespread adoption across all clinical domains, despite advancements in consumer applications.
Innovation Solution
A system and method using speech-to-text artificial intelligence to transcribe doctor-patient interactions into text format, employing techniques to split transcriptions into chunks, summarize relevant portions, and extract medically relevant information to overcome the token length limitations of large language models, allowing seamless integration with existing hardware and enabling efficient medical note generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If speech recognition systems are used for medical documentation, then transcription capability is provided, but recognition error rates are high and technology is immature
Solution Approach 1:
The patent segments the medical conversation transcript into multiple chunks or segments, processes each segment separately through the LLM, and then aggregates the results. This approach allows the system to handle long transcripts that exceed token limits while maintaining processing accuracy and reducing error rates through manageable processing units.
Solution Approach 2:
The patent introduces an intermediary processing layer that includes techniques such as summarization, key information extraction, and structured formatting before feeding data to the LLM. This intermediary layer acts as a mediator between the raw transcript and the AI model, improving recognition accuracy and reducing errors by pre-processing the input data.
2Measurement precision
If LLMs are used to determine medical note content, then accurate medical documentation is generated, but token length limitations prevent processing of full medical interactions
Solution Approach 1:
The patent divides the complete medical interaction transcript into multiple smaller segments or chunks that fit within the LLM's token limits. Each segment is processed independently to generate accurate medical note content, and the results are then integrated to form the complete documentation, thereby overcoming the length limitation while maintaining precision.
Solution Approach 2:
The patent transforms the problem from a single-dimension (processing the entire transcript at once) to a multi-dimensional approach by processing segments in parallel or sequence, then combining results. This dimensional change allows the system to handle arbitrarily long transcripts by distributing the processing across multiple LLM calls while maintaining overall accuracy.
3Measurement precision
If new AI technologies are deployed in medical settings, then improved transcription accuracy is achieved, but hardware installation and system setup become complex
Solution Approach 1:
The patent designs the system to work with existing, universally available hardware (standard microphones, cameras, and processors already present in medical devices). By making the AI solution compatible with existing infrastructure rather than requiring specialized hardware, the system achieves improved transcription accuracy without increasing device complexity or installation burden.
Solution Approach 2:
The patent enables the system to utilize existing hardware resources and computational capabilities already present in medical devices. The AI processing leverages built-in processors and memory, eliminating the need for additional hardware installations while maintaining high transcription accuracy through efficient use of available resources.
Data Source
AI summary
A system and method using speech-to-text artificial intelligence to transcribe a doctor-patient interaction into a text format. A website or application on a computer, phone, or device records an interaction. A speech-to-text artificial intelligence that will transcribe the doctor-patient interaction into a text format. After the system of the present invention has received the transcription, it will ask the doctor what sections he would like in his medical note. After a selection of the pieces of the note desired, the transcription of the recording between the doctor and patient is sent to the application server. The application server will use a large-language model AI to determine the content of any of the medical note sections. This input length is almost universally significantly shorter than the length of a standard medical interaction. The application uses three techniques to create a single note from the input.