AI promis­ing as a physi­cian as­sis­tant

The Buffalo News - - WEATHER -

per­for­mance of ex­pe­ri­enced physi­cians. It was more than 90 per­cent ac­cu­rate at di­ag­nos­ing asthma; the ac­cu­racy of physi­cians in the study ranged from 80 to 94 per­cent.

In di­ag­nos­ing gas­troin­testi­nal dis­ease, the sys­tem was 87 per­cent ac­cu­rate, com­pared with the physi­cians’ ac­cu­racy of 82 to 90 per­cent.

Un­leash­ing un­knowns

Able to rec­og­nize pat­terns in data that humans could never iden­tify on their own, neu­ral net­works can be enor­mously pow­er­ful in the right sit­u­a­tion. But even ex­perts have dif­fi­culty un­der­stand­ing why such net­works make par­tic­u­lar de­ci­sions and how they teach them­selves.

As a result, ex­ten­sive test­ing is needed to re­as­sure both doc­tors and pa­tients that these sys­tems are re­li­able.

Ex­perts said ex­ten­sive clin­i­cal tri­als are now needed for Zhang’s sys­tem, given the dif­fi­culty of in­ter­pret­ing de­ci­sions made by neu­ral net­works.

“Medicine is a slow-mov­ing field,” said Ben Shickel, a re­searcher at the Univer­sity of Florida who spe­cial­izes in the use of deep learning for health care. “No one is just going to de­ploy one of these tech­niques with­out rig­or­ous test­ing that shows ex­actly what is going on.”

It could be years be­fore deeplearn­ing sys­tems are de­ployed in emer­gency rooms and clin­ics. But some are closer to real-world use: Google is now run­ning clin­i­cal tri­als of its eye-scan sys­tem at two hos­pi­tals in south­ern In­dia.

Deep-learning di­ag­nos­tic tools are more likely to flour­ish in coun­tries out­side the United States, Zhang said. Au­to­mated screen­ing sys­tems may be par­tic­u­larly use­ful in places where doc­tors are scarce, in­clud­ing in In­dia and China.

The sys­tem built by Zhang and his col­leagues ben­e­fited from the large scale of the data set gath­ered from the hospi­tal in Guangzhou. Sim­i­lar data sets from U.S. hos­pi­tals are typ­i­cally smaller, both because the av­er­age hospi­tal is smaller and because reg­u­la­tions make it dif­fi­cult to pool data from mul­ti­ple fa­cil­i­ties.

Zhang said he and his col­leagues were care­ful to pro­tect pa­tients’ pri­vacy in the new study. But he ac­knowl­edged that re­searchers in China may have an ad­van­tage when it comes to col­lect­ing and an­a­lyz­ing this kind of data.

“The sheer size of the pop­u­la­tion – the sheer size of the data – is a big dif­fer­ence,” he said.

Or­ga­ni­za­tions are build­ing sys­tems that an­a­lyze health records to flag con­di­tions like os­teo­poro­sis, di­a­betes, hy­per­ten­sion and heart fail­ure.

Newspapers in English

Newspapers from USA

© PressReader. All rights reserved.