Nvidia Presents Bio-Megatron Medical Speech Transcription System
Artificial intelligence doctors talking to patients digitally could play a crucial role in medical care, according to research presented by Nvidia at this year’s Conference for Machine Intelligence in Medical Imaging. The research looks into new ways to use the kind of medical transcription and analysis service that has grown increasingly popular during the ongoing COVID-19 health crisis and the subsequent strain on medical resources.
The researchers presented a system for making specialized, but flexible models that can transcribe and analyze conversations between doctos and patients. The scale and potential specificity are what make their creation, named Bio-Megatron, stand out. Trained on the 6.1 billion words in the PubMed database, Bio-Megatron has 345 million parameters. The system was tuned with natural language processing and automatic speech recognition software developed by the NIH and could be used to map what is said to health databases lower the time and energy doctors spend on research.
“We’re pushing the state of the art when it comes to natural language understanding,” Nvidia product manager for healthcare AI Raghav Mani told Voicebot in an interview. “We’ve had ongoing initiative focusing on it for the last few years, and the benefits of it are becoming more evident. [The new models] improve on capturing information, extracting data, and improving patient experiences.”
Bio-Megatron is 92.05% accuracy accurate after one millisecond of processing, according to the paper. If the AI is that reliable, a doctor won’t have to check the transcript line by line for mistakes in most cases. And it could serve as just part of a larger system that reduces the problems doctors face when they have too much to do and start to burnout.
“It’s a key enabler for telemedicine and other tools for documentation,” Nvidia global head of medical AI Dr. Mona Flores told Voicebot. “The doctor can concentrate on the patient, instead of taking notes.”
There are a growing number of such products available, with new capabilities and upgrades coming out at a regular clip. Amazon came out in December with Amazon Transcribe Medical, an automated transcription service for medical professionals, while Nuance and Microsoft have partnered to upgrade and merge Nuance’s Dragon Medical Virtual Assistant with Microsoft’s Azure platform. Other startups offering related services to medical professionals are also on the rise. The list includes medical voice assistant developer Saykara, which has raised $9 million from investors, and Suki, which has channeled $40 million in funding to integrate into broader products as well as providing a standalone product. Amid a surge of COVID-19-related use, Orbita raised $9 million for its healthcare AI. Speech recognition technology startup Deepgram meanwhile donated $1 million worth of its automatic speech transcription and analysis platform to assist medical providers during the current crisis. Nvidia’s creation could augment many of their platforms with its own approach to transcription.
“I can imagine clinical products enabling using this tech to improve their own offering,” Mani said. “We do a lot of work producing models and provide tools for developers to create even more focused conversational AI for their domain. It’s definitely complementary to the work that others are doing.”
There’s also interest in bringing the transcribed and analyzed medical conversations to research trials. For instance, Stefanini Group and CliniOps launched a platform called Trust that uses Stefanini’s Sophie virtual assistant to aid medical researchers by digitizing and automating the process of setting up trials. The Sophie AI handles the automation process faster and with less potential for human error than traditional tests. The trials are completed more quickly and at a lower cost than would typically be the case. Sophie also assists the researchers more directly, interacting with users to answer questions and fetch documents across the platform. The system built by Nvidia could also be applied to clinical research. The breadth and depth of the datasets would be of use for performing studies that rely on high-quality data for accuracy.
“We are creating a new corpus of data for clinical trials,” Flores said. “We have data in a format that can enable downstream analysis data useful for clinical trials. We’re providing ingredients of a higher quality for a specific dish.”