Hacker News Reader

Comment by vonklaus | Hacker News Reader

Theranos is working on a technology that will exist in the near future, confirmed by many many other companies working on it, so it seems quite likely: someone will succeed in using less blood, to run more tests.

That doesn't sound like a totally alien idea, and since Theranos will probably not deliver on this promise fully, someone else likely will. However, the uBeam sounds super dubious. I think wireless energy is probably possible, but without a prototype a consumer could feasibly use in their home, and the large amount if scientists I coming out against this a few years ago doesn't seem to have shrunk.

So, yes it is a really hard interesting problem, but all the fun ones are. I wouldn't have invested with the expectation of success, but that the technology developed could be used for something else/is valuable.

View on HN

I don't think they will. Or rather they won't deliver anything as good as what they promise. Imagine you're testing for something that requires 10 modified white blood cells per million. A pinprick drop of blood has about 7,000,000 white blood cells in it. So you might expect to see 70 modified cells... but only if those cells are evenly distributed throughout the patients blood. They're not. They're randomly distributed. So the statistical likelihood of getting a false negative is high.

The answer to that is to take more blood, from a vein. You can't make a more accurate test because you simply don't have the necessary information. If you do take more blood then the test gets as expensive as a normal test and you no longer have any cost advantage.

> Theranos is working on a technology that will exist in the near future, confirmed by many many other companies working on it, so it seems quite likely: someone will succeed in using less blood, to run more tests.

Except we don't need more blood tests[0]. With uBeam, there's no doubt that if the technology works, it would be a game-changer. Theranos was facing both technical and market challenges, but with uBeam, it's a purely technical challenge.

0: http://fivethirtyeight.com/features/theranos-is-wrong-we-don...

I can't agree more our startup is working on time travel there's no doubt that if the technology works, it would be a game-changer. Since it's a purely technical challenge we expect to raise a trillion dollars in our seed round.

And if not, pivot to perpetual motion, like this chap:

https://vimeo.com/148032441

I should point out that time travel would fix the phone charging problem quite handily. Just send your phone back in time a few hours. Not sure if there are other applications.

You don't even need to do this - if you need a trillion dollars then just get in your time machine once it is built, head back 25 before today and invest $1 in all the hot stocks and winning lotto number so that by today you have the trillion you need to build the time machine. Simple.

I disagree with that article as it makes no sense. Their assertion that tests === diagnostics that map 1:1 to a disease was baseless. Tests don't answer the question "doe patient have diabetes?", they answer the question "what is the patients glucose level?" 115 mg/dl.

Maybe the patient forgot to fast, maybe the test was a fluke? Either way a consistent average of 85 would put that out of range. SO I found the five thirty eight article stupid so I guess I misunderstood why more health data would be bad.

The 538 article actually makes a lot of sense and is a pretty decent introduction to a common public health issue.

> Their assertion that tests === diagnostics that map 1:1 to a disease was baseless. Tests don't answer the question "doe patient have diabetes?", they answer the question "what is the patients glucose level?" 115 mg/dl.

Medical tests that aren't actionable aren't useful. So while you're correct in a very narrow sense, you have to use the results to make a decision about treatment. If you're going to make treatment decisions based on tests that have some amount of inherent error, you may end up over or under-treating patients.

> Maybe the patient forgot to fast, maybe the test was a fluke? Either way a consistent average of 85 would put that out of range.

Medical tests have both random and systematic error. What you're talking about is random error (i.e. a single test result being bad). However, many tests have systematic error (i.e. the result being wrong every time the test is taken) just due to personal variation in what's "normal".

The article makes the point that if more tests lead to worse outcomes[1] (due to overtreatment) then adding more tests is useless or worse than useless. This effect is particularly pronounced when testing across large populations, even if the test is highly accurate.

[1] And in many cases they do. For example, early mammogram screenings result in worse outcomes for patients when done across a large population. See http://annals.org/article.aspx?articleid=2480757

Medical tests that aren't actionable aren't useful.

True, but screening tests are very useful. If you can separate the population into "definitely diabetic", "definitely not diabetic", and "needs more testing" quickly and cheaply, that's a win. The fact that it doesn't always give you an actionable result doesn't make it useless.

> If you can separate the population into "definitely diabetic", "definitely not diabetic", and "needs more testing" quickly and cheaply, that's a win.

That's the fundamental problem. It turns out with a reasonably accurate test, a low incidence rate, and a large population, the "definitely X" class will be filled with people who aren't actually X. Going back and trying to separate the two ends up costing more and causing more harm to the non-Xs then just waiting until there's a problem. I'd encourage you to take a look at the link I posted since it takes a very clear look at the problem in the context of mammograms and breast cancer.

One way to look at it is, as you point out, our estimates of a measure are more accurate when you take more measures.

But the major issue with more data is Bayes' Law. For decades, running tests and collecting data was expensive, so we only ran them on people who were sick or suspected of being so. For prostate cancer, this data would tell us the probability that one would have a high PSA measurement given that they have prostate cancer (p(high PSA|prostate cancer). E.g., we know that 70% of prostate cancer patients have high PSA.

But, as Bayes' Law points out, this does NOT tell us the probability that you have prostate cancer given that you come in for a normal checkup, and we find a high PSA (p(prostate cancer|high PSA)). E.g., we do NOT know that 70% of people with high PSA have prostate cancer (absent other symptoms).

In order to determine that, you need to know the probability of having a high PSA in general. And the point is, NOBODY KNOWS these population meaures yet. Collecting lots of data is a good way to establish these baselines, but it's not clear it will help with diagnoses at all.

Another real example: we've gotten better and better at detecting tiny tumors early. But once we started looking at the data, we noticed that a large fraction of tiny tumors simply disappear over time. So, when the tumor is small enough, it's not clear whether we should treat it with surgery/radiation/chemo (which have their own risks/downsides) or just wait-and-see.

Reply on news.ycombinator.com