Reliably transcribing medical words and phrases: FosterHealth AI
FosterHealth AI’s HIPAA compliant AI-powered scribe generates fact-checked clinical notes based on conversations between patients and physicians.
Challenges in transcribing medical terms
Our current customers include leading research institutions, individual practices and small clinics spanning across the United States and India. Additionally, our customers come from various specialties, including but not limited to internal medicine, oncology, and gastroenterology. In order to ensure that physicians trust the product, our technology must reliably transcribe medical words and phrases that are used across diverse specialties and geographies.
Currently, state-of-the-art transformer architecture based speech processing models are not sufficiently reliable to perform speech processing related tasks in medical context. For example: “Amyatinib” instead of “Imatinib” — while the AI transcribed word is phonetically similar, it constitutes an error. We need technological and interface innovations to address these reliability related issues.
Our design philosophy
Every healthcare operator must find our application trustworthy. They should feel comfortable and confident about our AI model’s outputs every time they use our application. To achieve this design goal, we incorporate three main design principles into our product development process:
- Design an independent fallback system that enhances the reliability by actively monitoring the primary AI system’s output
- If needed, the AI model should ask the user for help before generating the final output
- The user should be able to review the AI outputs in a seamless manner
In this blog, we describe how our application leverages these principles and reliably transcribes medical words and phrases.
Medi-Spellcheck system
First, we enforce multiple engineering controls to improve the reliability of primary AI (complex, transformer architecture based model). We then extract medically relevant terms from the transcript using simpler, classical natural language processing (NLP) algorithms (fallback system). Due to our partnerships with leading research institutions, we have trusted, physician reviewed text corpus data in our stack. We compare the medical terms in the transcript with the terms in the trusted text corpus and extract terms which do not have a match (errors self-identified by the system). We then use NLP algorithms to suggest what may be the right spellings for these terms based on the trusted text corpus.
Our Medi-Spellcheck interface enables the user to quickly review the medical terms with errors, as identified by the system. The interface enables the user to accept or reject the suggestions presented by the algorithm. Additionally, the interface allows the user to quickly search the internet in case they want to use suggestions from the internet.
It gets better as you use more: it learns from your instructions
Our Medi-Spellcheck system continuously learns every time the user accepts or rejects the suggestion. It adds the correct word specified by the user to user-specific text corpus. Our Medi-Spellcheck interface uses both general medical text corpus, developed in collaboration with our research institute partners and user-specific text corpus to check for errors. Consequently, as the user uses the system more, their user-specific text corpus becomes larger, enhancing the spelling handling system and ultimately resulting in a superior user experience.
Our goal is to deliver the state of the art technology in a reliable and trustworthy manner. We are constantly talking to our users, collaborating with leading research institutes and healthcare experts and continually improving our service. If you have any additional questions or if you want to partner with us, please contact us here.
FosterHealth AI team