Hacker News Reader

Comment by esoleyman | Hacker News Reader

I don’t like relative risk and relative risk reduction because it tends to overestimate the effectiveness of the intervention.

In this case, the absolute risk when measuring for death in the GIM pre-intervention and GIM post-intervention are 0.0215 (2.15%) and 0.0146 (1.46%) with an absolute risk reduction of 0.0069 (.69%).

While the relative risk is 26% across the pre- and post-intervention, the absolute risk reduction is only 0.69% with a NNT (number needed to treat) of 1/156. Which means that 1 patient in 156 was helped by this intervention.

In addition, they had 2 false alarms for each true alarm and could suggest that interventions were performed in patients who did not require it — more tests, medications and possibly increased risk from said interventions.

This shows that the CHARTwatch ML/AI is not helping at all that much clinically.

View on HN

I like this analysis, although I come to a different conclusion: if AI can give early warning to nursing staff, telling them 'look closer', and over 1/3 of the time, it was right, that seems great. Right now in a 30 bed unit, nurses have to keep track of 30 sets of data. With this, they could focus in on 3 sets when an alarm goes off. I believe these systems will get better over time as well. But, as a patient, I'd 100% take a ward that early AI warning with 66% chance of false positives over one with no such tech. Wouldn't you?

I would not. High false alarm rates are a problem in all sorts of industry when it comes to warnings and alerts. Too many alerts, or too many false positive alerts cause operators (or nurses in this example) to start ignoring such warnings.

This is the real problem. In a perfect world, everyone pays attention to alarms with the same attentiveness all the time. But it just isn't reality. Before going into building software, I was in the Navy and after that did work as a chemical system tech. In the Navy, I worked in JP-5 pumprooms. In both environments we had alarms and in both environments we learned what were nuisance alarms and what weren't, or just took alarms with a grain of salt and there for never paid proper attention to them.

That is always the issue with alarms. You have a fine line to walk. Too many alarms and people become complacent and learn to ignore alarms. Too few alarms and you don't draw the attention that is needed.

More data with appropriate confidence intervals can always be leveraged for good. I hear this application often in medical systems, and recognize the practical impact. The problem is incorrect use of this knowledge (eg to overtreat); not having the knowledge.

No, the problem is information overload. Even without these errors nurses are often overburdened with work and paperwork. Adding another alarm, with a >50% false positive rate is going to make that situation worse. And the nurses will start ignoring the unreliable warning.

I suspect we are on the same page. My point is in regards to using information as described in the article to improve the system. I do not think an on/off "alarm" is the way to do this. The key is to use information from signal processing theory (eg how a Kalman filter updates) to provide input into what medical action to take. The reactions against more diagnostics etc is due to how they are applied, like a brute force alarm, leading to worse outcomes through, for example, unnecessary surgeries etc.

The reduction I am arguing against is: "Historically, extra information and diagnostics that have an error margin results in worse outcomes because we misapply it; therefore don't build these systems."

Yeah, we agree. That reduction is also what I originally commented against.

Yeah, but GP gives the example of a 33% chance for true positive. That's more than enough to keep you on your toes.

At work, we had an appliance which went into failsafe on average 8 times per day. The failsafe is meant to remove power from a device-under-test in case of something like fire in the DUT. The few actual critical failures were not detected by the appliance.

Instead, the failsafe has the effect of merely invalidating the current test, and making the appliance unable to run a test correctly until either power cycled or the appliance's developer executes a secret series of commands that are not shared with us.

So of course an operator of the appliance found a way to feed in a false "I'm here!" with a loop, to trick the appliance into never going into failsafe…

That's for ~6.8% of all tests being false-positive, ~93.2% being true-negative, and ~3 tests that should have triggered failsafe did not.

Sooooo... You're saying that the chance of a true positive given an alert was much less than 33%?

I don't if you meant it as a counterpoint for what I said, but it really isn't.

Sorry, I meant to say that with only 6.8% of all tests triggering a false alarm (and 0% true alarm), a test operator still found a way to prevent the alarm from occurring rather than being kept on their toes.

Yeah, but again, the problem isn't the high false positive rate.

The problem is that given any positive at all, the chance it points to a problem is still virtually zero.

If it was 6.8% of all tests as false positives and 2% true positives, probably people wouldn't have silenced the alarm.

If it goes off 8 times a day and 2 of them are true positives, then people have recent memories of having to fix problems pointed by the alarm.

I hope you are joking.

I'm not. If you have three alerts a day, a 33% chance of true positive per alert means you'll get an alert pointing to a problem at least once per day.

That's enough to anchor "alert == I might find a problem" in the user's mind.

No, many people working in clinical units wouldn't. Because of what might happen on false alarms. What GP said: more meds, more interventions. It's not clear at all whether such systems would help with current workflows and current technology. One of the most famous books about medicine says that good medicine is doing nothing as much as possible. It's still very true in 2024, and probably for a long time still.

I like this analysis, although I come to a different conclusion: if AI can allow nurses to manage 10x as many beds (30 vs 3), a hospital can now let go 90% of its nursing staff. Wouldn’t you?

Luckily most hospitals in the world seem to be short-staffed, and the population of sick is growing (because people are living longer).

Generally speaking, they aren’t short staffed because there aren’t enough nurses, but because they can’t/won’t pay them enough. Those same hospitals hire large numbers of travel nurses to supplement their “short staff” at pay rates double or triple a local nurse.

And the nurses who want decent pay and can do travel nurse, do travel nurse

Coming to the conclusion that cutting 90% of nursing staff is possible and desirable is an astonishingly disconnected take

False positives are not harmless. Textbook example is breast cancer screening https://theconversation.com/breast-cancer-screening-in-austr...

That’s true, but no alarm is deadly. It’s the matter of risk assessment. Not to mention AI can be refined and improved.

It's not just a false positive rate, but also the rate you train nurses to ignore alerts.

>1 patient in 156 was helped by this intervention

the headline says we're talking about death: does that mean 1 life was saved for every 156 patients?

>In addition, they had 2 false alarms for each true alarm and ... and possibly increased risk from said interventions

but wouldn't this study have captured any deaths from those interventions, so the 1 out of 156 life-savings was net?

Would you suffer serious nonlethal complications from false alarm to (maybe) save your room neighbour that you've never met before? This wouldn't be captured.

an individual would probably not make that choice, but the population could easily, the insurance company might, religious leaders might, etc.

this study was measuring deaths and what you are suggesting would be outside this study, but it could be measured also.

this was excellent and necessary context on all fluff pieces like the OP. how can we automate this kind of analysis?

You can't automate it. You have to look at the data and charts to figure out the specifics you want and then you plug and chug. I haven't looked deeply at this though but whenever researchers use relative risk and it shows a profound effect, I always calculate the absolute risk to make sure that the intervention is effective.

Many researchers go to relative risk because it shows better results!

I know everyone hates "I asked ChatGPT" comments but...I feel it's relevant here.

It came to roughly the same conclusion as the gp comment when provided with the study PDF.

https://chatgpt.com/share/66eb09e3-7a74-8008-afa8-3b60161d24...

(Though obviously this approach still requires you to go and look at the PDF yourself to make sure it isn't making anything up)

I think that ChatGPT result is a Rorschach test, it wrote things like "The percentage reduction could be exaggerated based..."

Could is doing a lot of work in letting you interpret what it's saying however you like.

>How can we automate this kind of analysis?

Happy to talk about it.

Are you in the Healthcare industry?

Update (3 days later):

It was bs, as usual.

That’s a good point, a similar conversation was had around the Covid jabs, with some research re vaccine mandates concluding “heads of governments, schools, healthcare facilities, and private businesses (were) misled by the vaccines’ reported 95% relative risk reduction”

https://www.sciencedirect.com/science/article/pii/S277265332...

Reply on news.ycombinator.com