Hacker News Reader

Comment by eanzenberg | Hacker News Reader

Literally false positive and false negatives. In other words, if you test millions of people, there will be more people tested positive who DON'T have the disease vs. those who test positive and HAVE the disease.

View on HN

Exactly! For any medical tests, there are 2 measurements: sensitivity and specificity, one to judge false positive and another judge the false negatives.

https://en.wikipedia.org/wiki/Sensitivity_and_specificity

Spain sent testkits back to China because they were unreliable

In the NYT article on Italy today, frontline medics were saying that clinical symptoms were anecdotally more reliable than the tests because there are too many false negatives. Not sure whose tests they're using but I've heard similar things about US PCR tests.

https://www.nytimes.com/interactive/2020/03/27/world/europe/...

Yep! CT scans and clinician's experience is more reliable than any tests for COVID-19 diagnostics (which should only serving as a definitive confirmation). This also why I believe human doctors are still unbeatable by AI

I think that's missing the point, which is that the only way to get the necessary accuracy is to use multiple independent sources of information. It's not whether clinicians are better than tests or vice versa, but that no single datum is good enough.

No. The point is, COVID-19 testing kit is in extreme short supply everywhere in the world right now given most countries are locked down and the viral outbreak patients are growing at an exponential rate, the next reliable way to diagnose COVID-19 for clinicians without testing kits is to use CT (which gets results fast and inexpensive and the equipment is universally accessible in every hospital) scans and find some patterns, which is exactly what doctors in Wuhan did during the early phase of the outbreak[1].

[1]: https://www.ncbi.nlm.nih.gov/pubmed/32134681

Well, that's your point, but I think everyone knows the tests are scarce.

> This also why I believe human doctors are still unbeatable by AI.

Because you hypothesize that they can rapidly learn new, not very distinctive patterns without much data?

> hypothesize

Of course not! We've got half a million cases now, hundreds of thousands cases in hospitals, and hundreds of papers published in all the medical journals. So yeah, data is rich.

I was mainly referring to CT scan data. A symptomatic check should not require much experience, rather applying the knowledge in form of a simple decision tree.

Regarding CT scans .. machine learning will be so much better than humans in classifying those once enough data is available... it‘s completely out of question for me as an ML researcher.

I remember an article early in the outbreak that Chinese doctors found the most reliable diagnosis comes from CT scans. Test was merely the first filter. They tuned the settings to increase scan speed at the cost of resolution and setup a process to scan up to 200 patients a day per scanner. I wonder if this is still best practice that other countries aren't adopting. On the other hand I expect tests to have improved since the early days. There's also the consideration China simply had more CT scanners that could be mobilized to Wuhan which makes this less practical elsewhere.

My doctor friend who got the minutes of a call with a chinese doctor said they were also swabbing multiple areas of the body for PCR and doing an antibody screen as well (in addition to CT).

If you know the false positive and false negative rates, then you can predict how many in a sample size of millions will end up falsely identified in this way.

What I don't know and what the article doesn't seem to mention is if the false positive rates are known yet and if we have more or less than the expected number of 'reinfected'.

Nothing about Covid19 testing is statistical right now. Why? Because tests are scarce and are used by medical professionals to save lives. There will be time for rigorous statistical analysis after this epidemic is under control.

Reply on news.ycombinator.com