AI-powered stethoscope can spot pneumonia by listening to how a patient breathes | Hacker News Reader

AI-powered stethoscope can spot pneumonia by listening to how a patient breathes

dailymail.co.uk

9 pointsknown8 comments

Hello, a little self promotion here : We have just released Wisteria (https://news.ycombinator.com/item?id=19128287), which is a platform which does real-time audio processing in the browser using deep-learning, powered by Tensorflow.js

Currently targeted at music education, it can be probably be extended to abnormal noise monitoring, like this AI-powered stethoscope given the right dataset.

The article says the accuracy of the stethoscope’s predictions is 87%. I wonder what the precision and recall are. Precision saves money, recall saves lives.

87% accurate means a lot of people will get misdiagnosed. It would be good to give a baseline of humans to compare to so we can better understand the problem. Of course the daily mail isn’t exactly known for rigorous reporting...

Most people do not have pneumonia and accuracy has no meaning without data on what it was measured on.

It is important to know if it was tested on suspected pneumonia cases, any breathing problems, any illnesses or sample of general population.

The number is less than impressive if it was just suspected pneumonia cases. (Because then, the treatment is wide spectrum antibiotics blind anyway plus supportive. And culture to see what bug it is.)

Wow, they have noise cancelling in a digital stethoscope - and that is being touted as something amazing! :) This article is fun.

It says it was trained on data from Africa and Asia, yet it seems to be targeted globally. Shouldn't it be trained on many more healthy people from other parts of the world to be more trustworthy?

I assure you Asians and Africans breathe just the way white people do. No gills involved.

I also don’t quite understand why you are calling for more tests on „healthy“ people? The bad interpretation is that you consider all Asians and Africans to be „unhealthy“. The good one is... seriously, I can’t think of a good one. One would certainly want further tests to also include the sick people one hopes to identify?

They could have a bit different telltales of pneumonia, because of differences in their environment, different viral strains, etc. If you want to build trust in ai-powered medical devices, you should be very thorough in testing them.

Reply on news.ycombinator.com