Digital Health

환자의 생각을 말로 바꾸는 AI (AI can turn a person's thought to speech)
  • 유시형 기자
  • 기사입력 2021.07.15
    • UCSF’s medical science research on decoding speech in a paralyzed person with anarthria was published in the New England Journal of Medicine (NEJM) on the 15th of July. The study examined a 36-year-old man who has had quadriparesis (muscle weakness is all four limbs) and anarthria (inability to articulate speech) for the past 16 years due to a brainstem stroke. Normally, paralyzed persons with anarthria are unable to use assistive communication devices because their movement is severely impaired. However, artificial intelligence and technology can help decode speech from a person’s brain activity.
      지난 16년 동안 말을 하지 못하던 뇌졸중 환자가 인공지능(AI)의 도움을 받아 자신의 생각을 문장으로 전환하는 데 성공했다. 미국 샌프란시스코 캘리포니아대(UC샌프란시스코, UCSF)의 연구진은 15일 국제 학술지 ‘뉴 잉글랜드 저널 오브 메디슨(NEJM)’에 “Neuroprosthesis for Decoding Speech in a Paralyzed Person with Anarthria(뇌졸중 환자가 말을 하려고 할 때 뇌에서 나오는 전기신호를 문장으로 전환했다)”라고 밝혔다. 뇌졸중 환자는 운동신경이 심각하게 손상되고, 실어증이 발생하기 때문에, 일반적으로는 보조 통신 장치를 사용할 수 없다. 하지만, AI는 사람의 뇌신호에서 직접 말을 해독하는 것을 도울 수 있다.
    • AI-based system for decoding speech in a paralyzed person with anarthria/Image credit to New England Journal of Medicine and UCSF
      ▲ AI-based system for decoding speech in a paralyzed person with anarthria/Image credit to New England Journal of Medicine and UCSF

      The research was conducted 50 times over 81 weeks and involved planting a high-density, multielectrode array in the participant’s sensorimotor cortex (part of the brain that controls speech). From there, the array can pick up neural signals (speech processing) from cortical regions as he attempted to speak. The recorded signals were then transmitted to a real-time processing system via a detachable digital link. The 50 sessions had the participant attempt to articulate words from a 50-word vocabulary and come up with sentences using those words.
      UCSF의 연구진은 뇌졸중 환자의 뇌에서 언어기능을 담당하는 감각운동피질에 다채널기록법(multielectrode array)을 부착했다. 이후 환자에게 컴퓨터 화면에 뜬 50개의 단어 중 하나를 제시하고 곧바로 해당 단어를 말해보라고 했다. 음성은 나오지 않았지만 언어 영역의 뇌에서 나오는 전기신호는 달라졌다. 인공지능은 단어마다 달라지는 환자의 뇌신호 형태를 스스로 터득했다. 해당 인공지능에는 문장에서 특정 단어 뒤에 어떤 단어가 나올 가능성이 큰지 예측하는 자연어 모델도(natural language modeling) 추가됐다. 실험 결과, 뇌졸중 환자의 뇌신호로 만든 문장은 약 25%의 착오율을 보였다. 30% 미만의 착오율을 가진 시스템은 사용 가능한 것으로 간주되기 때문에, 이는 큰 성과라고 보여진다.

      Machine learning, computational and natural language models were used to decode the words and sentences the participant was attempting to articulate. The results were that sentences were decoded with a median accuracy of 75% and the median number of words decoded per minute was 15.2. The median word error rate was 25.6%, which is more than sufficient for everyday communication. The median word error rate of 25.6% was a key figure in this research, since the target was getting below 30%, which is the threshold for usable speech-decoding approaches. The specific decoding AI used in the researched showed promising results even with large quantities of training data without daily recalibration, which suggests this decoding method can be applied for long-term direct-speech (neuroprosthetic) applications.
      Beyond speech decoding for paralyzed persons, companies like Elon Musk’s Neuralink are working on developing a fully-implanted, wireless brain-machine interface (BMI) to help people with paralysis operate computers and mobile devices directly from their neural activity. Artificial intelligence and technology not only possess the ability to restore the ability to communicate for paralyzed person, but also improve the autonomy and quality of life for all types of people.

    최신 뉴스