Hacker News Reader

It's interesting that you mention the criminology scenario. Courts are starting to use software algorithms to predict recidivism rates for purposes of setting bond amounts. The reports are even given to judges to inform sentencing. However, there is an enormous backlash against them, because e.g. they predict higher rates for racial minorities: https://www.propublica.org/article/machine-bias-risk-assessm....

The propublica study has been shown to be a lie, see https://www.chrisstucchio.com/blog/2016/propublica_is_lying....

Anyway all I'm claiming is that statistical methods are much more accurate than humans. Nothing there disputes that claim. If you want the most accurate predictions possible, you should use an algorithm.

That article implies that humans are somehow fair or unbiased. That is a completely ridiculous claim that has been proven false many times. Human judges give ugly people twice the sentences of attractive people. Judges have been shown to give significantly harsher sentences just before lunch, when they are hungry. Not to mention all the classic biases against gender/race/political affiliation/etc. Studies have shown interviews are worse than useless at assessing how good someone will be as an employee. Instead employers are biased by how much they like the candidate. We should hardly expect traditional parole interviews to be any different.

But almost no one cares about these results. Yet when an algorithm is shown to have a (not statistically significant) bias, people freak out. This, if anything, proves my point that algorithm aversion is a serious problem.

Chris Stucchio's point is different from that of the article. The problem is that the word "bias" is overloaded, and used in two different ways. When ProPublica uses the word bias, they mean that one race is affected in a systematically different way than another race. When Chris Stucchio uses the word bias, he means a statistical method which has a built in bias to achieve a particular result. It is quite possible for both to be true: the statistical methods have no built in bias, but their prediction ends up making different predictions for different races. I find the most reasonable explanation to be that society itself has a racial bias, and society itself is the data generator.

Yes, this is a broad criticism of ML in public policy: that it's an amplifier for the status quo.

> amplifier for the status quo

While there are respectable ML folks making those criticisms, the commentary I've read seems more click-bait than science.

ML can be used just as traditional statistics to make causal inferences and predict the effect of intervention. There's nothing about ML that reinforces status quo more than traditional statistics, let alone case studies (aka anecdotes) or "common sense."

While this is definitely a clickbaity topic, and algorithms are usually better than people there are still real issues that come from seeking accuracy, particularly since real world data is so often biased to begin with.

Google had an interesting take on how you could control for some dimensions of bias here https://research.google.com/bigpicture/attacking-discriminat...

> There's nothing about ML that reinforces status quo more than traditional statistics

Yes but that is not how ML is marketed. It's marketed as way better than traditional statistics, and soon even better than humans.

But reality is that a trained system is only as good as the data used to train it.

> only as good

But ML can improve upon other methods of interpreting that (biased) data. Thus, in some ways better than traditional statistics and non-mathematical human intuition.

> Human judges give ugly people twice the sentences of attractive people. Judges have been shown to give significantly harsher sentences just before lunch, when they are hungry.

It would be nice if you could provide sources for these claims.

It took only 10 seconds of googling to come up with this:

https://www.scientificamerican.com/article/lunchtime-lenienc...

http://www.villagevoice.com/news/study-ugly-people-more-like...

Here's the paper for the just before lunch claim: http://www.pnas.org/content/108/17/6889

The lunch statistic probably came from this study: http://www.economist.com/node/18557594 http://www.pnas.org/content/108/17/6889

(To be fair, the study size was small).

There's a study on attractiveness and juror bias here (it's more complicated than just "ugly people get worse sentences" but some bias does show up for certain juror personality types): http://onlinelibrary.wiley.com/doi/10.1002/bsl.939/abstract

If there is ever a dress for success time it is when you are in court.

The ProPublica you cite as example for bias in criminal justice is plainly wrong. A follow up study found the the system they examined was not biased at all.

Reference: Flores, Bechtel, Lowencamp; "False Positives, False Negatives, and False Analyses: A Rejoinder to “Machine Bias: There’s Software Used Across the Country to Predict Future Criminals. And it’s Biased Against Blacks.”", Federal Probation Journal, September 2016, You can find the article here: http://www.uscourts.gov/statistics-reports/publications/fede...

In fact the ProPublica analysis (written by journalists,not scientists) was so wrong that the authors of the study wrote:

"It is noteworthy that the ProPublica code of ethics advises investigative journalists that "when in doubt, ask" numerous times. We feel that Larson et al.'s (2016) omissions and mistakes could have been avoided had they just asked. Perhaps they might have even asked...a criminologist? We certainly respect the mission of ProPublica, which is to "practice and promote investigative journalism in the public interest." However, we also feel that the journalists at ProPublica strayed from their own code of ethics in that they did not present the facts accurately, their presentation of the existing literature was incomplete, and they failed to "ask." While we aren’t inferring that they had an agenda in writing their story, we believe that they are better equipped to report the research news, rather than attempt to make the research news."

Ouch...

Is there evidence that there are high-performing algorithms that are available to doctors that they aren't using?

The basic algorithms like RCRI and CHADS2VASC are almost universally applied in the appropriate context (at my quaternary care teaching hospital, granted).

If there are numerous algorithms that we should be using but aren't, I'd argue that the bigger barrier is making them accessible and easy enough to use that they can be applied during a clinical encounter (among all the other things that need to be done).

Here's a simple algorithm that would help millions, yet it is not accessible to most patients:

Ordering A1C, remembering to actually check the results, and telling patient to stop eating sugar - well before they are symptomatic for diabetes.

> Ordering A1C, remembering to actually check the results, and telling patient to stop eating sugar - well before they are symptomatic for diabetes.

Yeah, it's that last part ("telling patient to stop eating sugar" [and getting them to actually change their behavior]) that's the hard part.

This is the holy grail of medicine and has been for decades. We've done billions of dollars of research over the last century, and the one clear thing we've found is that we do not know of any cost-effective way to get patients to change their behavior at large[0].

Seriously, if you really have the answer to that question, you could be a multi-billionaire.

[0] That is, we can change behavior of some patients, but they're generally the patients who would have changed their behavior anyway. And we can change the behavior of some of the rest, but they're not cost-effective.

If you knew how to do that, you could also solve the climate change / global warming problem, as well as host of other big social issues at the same time. As it is, people simply follow the incentive gradient, and most of the time there's damn little you can do about it.

And so it turns out once again, in every field, people really are the problem ...

There is, of course, the (unintended?) side-effect that if there was a cost-effective way for controlling peoples' behavior then we would be effectively under the tyranny of whomever possessed those means.

Sometimes we just have to let other people live with the consequences of their actions; which is hard, because often we also have to live with those consequences.

Whatever we say about "other people", we should also remember that we are "other people" to somebody.

Taxes?

Like, what might be the huge negative cost of tobacco taxes?

(There's an argument to be made for freedom, but you are making an economic argument about costs, not a moral one)

Unfortunately, taxes don't always make significant reductions of usage for addictive substances and may shift commerce towards illegal markets.

Right. Patient compliance is a hard problem. But doctor compliance with basic protocols is also very bad, and seems fixable with computers.

That is already done at a population level, and when I was in primary care I would get notifications about patients whose A1c was overdue.

Some hospitals do it, some don't. My wild guess - at best 50% of patients get it. And maybe 25% of those get it interpreted as well as you would for your family.

I've had exactly one of these tests, and that was only because the insurance I had required all employees to get a battery of blood tests drawn (or pay more each month for insurance). They planned on requiring this yearly, with plans to force folks to start meeting the metrics or pay more for insurance. I'm not sure how this panned out for them.

Anyhow, if it wasn't for this, I would have never had such a test. I'm in my late 30's. My father was diabetic at my age, actually, and it runs in the family. I'm a prime candidate for preventive measures, but still nothing. Now, I understand that I go to the doctor rarely but at any one of these I could have been asked about a physical and stuff like that.

This would be great for diabetes patients. It might be a good early intervention measure for folks that are prone to be diabetic.

But does such a test actually provide benefit to the general public? Would an otherwise heatlhy 30-year old with a bad diet fail this test - or would it only show that despite his bad diet, that his blood glucose levels have been pretty stable the last 3 months? I'm thinking that in an ordinary person with a working pancreas, it wouldn't show the spikes of diabetes.

That said, I do think these sorts of things work, especially when paired with labor protection policies that allow folks to miss work to go to the doctor (or be sick). For example, once I was in the Norwegian health system, I got a letter encouraging me to see the doctor for my cancer screening (I'm female). Turns out, they want folks to get one every 3 years and keep a database to keep track of who has gotten it and who hasn't. I don't know how well this sort of thing would work in a fragmented system like the US has, though.

I've seen reviews comparing clinical-decision-support-systems, which advise doctors on diagnosis and treatment and the machines did better.

There's also some research, in general , about the resistance of doctors to such systems. And yes while some of the reasons are UI, some are motivated by doctors preferences.

And even if the UI is a bit slow, we should at least seen such systems common in the more critical settings, or just with vulnerable or complex patients.

Yes, I've tried to use clinical decision tools for complex cases. They aren't integrated into our EHR (which is the most common EHR), so inputting the data for a case takes between 10-15 minutes. That's as much time as I get to see a patient at an outpatient visit.

I'm a doctor. I'd be delighted to have decision support for every problem. As it stands I have decision support for medication dosing and for appropriate use of imaging, both of which are very helpful.

Don't blame the customer for not buying if the product is not good.

There's isabel healthcare's CDSS, the interface seems rapid and one of the reviewers("sprecher") says that he can use it in a patient visit:

http://www.isabelhealthcare.com/customer-satisfaction/testim...

Could be just marketing, but maybe.

There's a free test. if you do test, please share your opinion of it, it's interesting.

> The problem is there has been enormous resistance to the use of algorithms in medicine

That's not 100% true for all disciplines. E.g., in radiology, there are some assisting technologies around for identifying tumors etc. While many radiologists use them daily, the evidence isn't always that clear. Often they just lead to a higher false positive rate.

I've spent a lot of time doing manual tracings of brain structures and I tell you now that while automated methods are improving rapidly, and just now approaching the average ability of a tracer, they do not rival experts yet. The most time efficient methods now is semi-automatic. Begin with the automated methods, then make manual corrections.

Given all that, brain radiology is a lot easier that pathological radiology (e.g. tracing the extent of a lung tumor). There is a lot of research into automating this, but dice overlaps are still poor.