AI matches expert diagnosis rate

INSTANT ANSWER::The deep learning model is able to diagnose Parkinson’s disease with an 80% accuracy rate, even using smartphone video of motor function testing

By Liao Hsueh-ju and Jonathan Chin  /  Staff reporter, with staff writer

Fri, Aug 23, 2019 - Page 2

An artificial intelligence (AI) deep learning model developed by researchers from Taiwan, Japan and the UK is capable of diagnosing Parkinson’s disease at the same success rate as human movement disorder specialists.

Since the discovery of Parkinson’s two centuries ago, physicians have relied on their eyesight to detect the disease’s telltale signs: tremors, a shuffling gait and slow and stiff movement, China Medicine University Hsinchu Hospital neurologist Chen Jui-cheng (陳睿正) said on Monday.

As these symptom are similar to the impairments associated with other ailments, such as joint issues and strokes, movement disorder specialists are relied upon to diagnose the disease, he said.

However, humans are limited by geography and time, causing delays that patients and their families cannot afford, he said.

Researchers from China Medicine University, National Tsing Hua University, the University of London and the University of Tokyo worked together to solve this problem, Chen said.

Using deep learning, the teams built an image recognition AI that taught itself to identify the symptoms of Parkinson’s with 30,000 archived video files and specialists’ diagnoses, he said.

Since the beginning of this year, the machine correctly identified 80 percent of the 109 cases it was presented, becoming the first automated program to approximate the performance of human experts, he said.

The program is capable of instantaneously generating a diagnostic reported based on videos recorded on smartphones of people performing motor function tests, Chen said, adding that it is a valuable resource for people in medically underserved areas.

The prevalence rate of the disease has increased in Taiwan over the past 30 years to about 40,000 people, likely driven by unknown environmental factors, he said.

Treatment options include medication, surgery and focused ultrasounds that have recently been accepted by regulators in the US and Japan, Chen added.

The team presented its findings at the annual conference of the Institute of Electrical and Electronic Engineers in Medicine and Biology and is to be published next year by the institute’s official publication, the Open Journal of Engineering in Medicine and Biology.