Archive for January, 2016
Superbe résumé d’une course particulièrement éprouvante, débutée dans le gros temps et terminée dans la pétole.
According to Eisner, there are currently three cultures of machine learning. Different people or projects will fall in different places on this "ML simplex" depending on what they care about most. They start with something in green and attempt to get blue as a way of achieving red.
For every complex problem there is an answer that is clear, simple, and wrong. H. L. Mencken
Once again, Wiley Miller gets it perfectly right.
A meta-analysis of blood pressuring lowering for the prevention of cardiovascular disease and death appeared just before Christmas. It attracted a day’s worth of comment before we all went off to do seasonal things and then recover from them. I guess the debate will start to build up again now, and when the article appears in print. It is certainly worth a careful read: it’s a model for this type of systematic review and it is very clearly written, covering a vast range of trials using blood pressure lowering drugs both for high blood pressure and for other indications. It confirms that BP lowering, like the use of statins, should be governed by total risk and not by a specific level of systolic BP. Remember the Fifth Commandment: Thou shalt treat according to level of risk and not to level of risk factor. The study also identifies clear class differences for different drugs in relation to different outcomes, e.g. thiazide-like agents are better at reducing the incidence of heart failure, whereas the logic of using RAAS inhibiting drugs to prevent renal failure looks very shaky. The authors conclude: "Our results provide strong support for lowering blood pressure to systolic blood pressures less than 130 mm Hg and providing blood pressure lowering treatment to individuals with a history of cardiovascular disease, coronary heart disease, stroke, diabetes, heart failure, and chronic kidney disease."
I think this may mark a watershed moment in the use of BP lowering agents. But contrary to most commentators, I think it will lead to a large and welcome decline in their use. A clue comes in the sentence, “Rather than a decision based on an arbitrary threshold for a single risk factor, this approach needs individualised assessment of the balance of absolute risks and benefits when physicians decide on the blood pressure level at which to start blood pressure lowering and the target blood pressure.” So this is a decision for the physician, is it? Why isn’t it a decision for the person who is expected to take the drugs for the rest of their life? And it’s at this point that the whole stack of cards begins to fall apart. For a start, our cardiovascular risk prediction instruments fail to predict most of the absolute risk and have poor overlap with each other. There is no way that we can produce more than a vague ball-park guess about the likely contribution of various treatment possibilities—non-pharmacological as well as pharmacological—to the outcomes of particular individuals. And most of these outcomes are binary—you either have a stroke or you don’t. They cannot be expressed as days of life gained, but only in terms that are borrowed from gambling and so have no objective meaning for individuals. The trials lumped together here were on subsets of people for relatively short periods of time, so even if you make them into a smiley face chart, you will be giving out a false message about their predictive value. And when you do make them into such a chart, many sensible people will look at it and say, "You mean to say that I’d have to be one of 231 people to take these pills for ten years just so that one of us wouldn’t have a heart attack? Sod that." Public health physicians will hold up their hands in horror. Oxford professors will rage at GPs (and The BMJ) for not imposing the supposed good of the herd on individuals. But why? It is for each of us to play the odds of our lives as we choose. Many will choose to take the pills—myself included. Many will not. The only right choice is informed patient choice.