Time for a new era for medicine prescription?

[kk2796]

After taking a course in datamining, it seems that we have an tool that has promise to greatly benefit our society in many areas. No - no, I'm not talking about domestic spying. But I am wondering about how datamining could be applied in the realm of medicine. In my mind, medical prescription is in something of a stone-age right now. I walk into my doctor's office, she takes in my symptoms, and thinks I may benefit from medication X, Y, or Z. Because I currently take medicine A, she rules out Z. Based on an article she'd read last month about high incidences of side effects from Y, she opts for X. Because I weigh about 180 pounds, she gives me the 200 mg dosage. If things don't improve significantly in a week or so, she'll up it to the 250 mg, or perhaps switch me over to Y.

Though I've cut out a few things, this is effectively the best that today's clinical medical practices will ever be able to offer - sure, tremendously experienced and knowledgable doctors might be able to make meager improvements upon the above; but the entire system is limited: all doctors are humans prone to error, prone to bias, and capable of taking only about 4 or 5 factors into consideration at a time, at best. Furthermore, for any given drug, a sheet with not much more than the following will be available to medical professionals:

Drug X is safe for treatment of symptoms S1, S2 and S3, with side-effects A, B, and C. Not safe during pregnancy. Not to be used with medicine Z. Dosages are 100 mg, 150 mg, or 200 mg. Take on an empty stomach.

Modern clinical trials are designed to give results based on single-variable variance, with all other factors controlled. Consider medication "Y". Who knows what, if any, impact taking Vitamin A and omega-3 intake will have on the risk of heart problems from taking Y? Do a trial. What about high blood pressure? Do a trial. How about a Ibuprofin table every few days thrown into the mix? Do a trial. What about routine exercise? Trial. Obesity? Trial. But will these trials be misleading: what if high blood pressure only increases the risk of heart disease if accompanied with obesity? And given all of the above trials, consider a person who takes Vitamin A, omega-3, doesn't take Ibuprofin, doesn't exercise, is not overweight, and has no family history of heart problems (despite coming from a huge family)... can the clinical trials above answer whether or not Y is the safest medicine for this person? And what about other risks aside from heart disease, such as kidney stones, migraines, etc.? Now is Y the safest?

I long to live in a world where I go into a doctor's office, undergo a physical examination, take some tests, and the data is fed into a computer: I have symptoms S1, S2, S3, and S4 with varying severities (and am fine on several other fronts), am 6'3", weigh 181 pounds, already take 500 mg of medicin A per day, am 28 years old, have a body fat % of 14, blood pressure of 140 / 80, cholesterol of 300 (??), take a multi-vitamin daily (Brand Foo, by the way), am a vegetarian, excercise twice a week, had only one aunt - who died of with heart disease, [insert about 20 other factors]....

After a little humming, the computer prints up the following: I am prescribed 187.5 mg of X, my dose of Z is adjusted to 578 mg, and I'm given a multi-vitamin recipe with the exact levels of 30 vitamins and minerals that would best help my body absorb X and Z, while mitigating side-effects to the greatest possible extent. I'm given a short list of foods to avoid and a list of foods to try to work into my diet.

In my mind, datamining is the exact tool that would make this feasible. Modern clinical medicine is built upon the single variance model: set up an experiment, control all factors but one variable, and determine the efficacy of modifications of that variable. Datamining throws this right out the window - it says, "give me as many symptoms as you'd like, as many variables as you'd like (diet, age, weight... the more the better!); as many medicines + dosages as you'd like, and give me as many results in as many different areas over whatever time periods you'd like"... and then, like magic, it will build a data model making unbiased connections between the above. And with each new patient and each new set of data added, the model will get just a little better.

Back to my example:
1 week later, I take another physical - the results are fed back into the computer, which has 2 effects:
1) The computer makes any minute adjustments to my prescription as needed.
2) My results are linked with my attributes/symptoms/prescription, and uploaded (anonymously - no name attached) into the national data repositiory and incorporated into the national data model along with millions of other results taken daily from other patients around the country. And with my results taken into consideration, every patient nationwide from that day forward will be prescribed a slightly different amount of whatever medicine they need... the prescription will be a *hair* more accurate, because the underlying data model of the human body will be slightly better based on my reaction to my prescriptions.

Am I a dreamer?

[ChristineR]

For most drugs, I don't think you'll ever get enough data points to find fits to your 30 odd factors. And for drugs, that could be disastrous--for example, someone who weighs 300 lbs but has liver malfunction may be killed by a dose appropriate for someone who weighs 100 lbs and has normal liver function.

[El Greco]

Am I a dreamer?

No, such programs exist and are being used already in hospitals. The point is that they are far from perfect because beyond a basic level things become subjective and ambiguous. There may be 200 studies with varying and contradicting results or no studies at all.

[kk2796]

For most drugs, I don't think you'll ever get enough data points to find fits to your 30 odd factors.

That is the incredible thing about data mining - you'd be shocked at how quickly meaningless data points (color of one's shirt) will be factored out, while important points (weight) are emphasized, so long as enough data points come in. The model will even tell you "I don't think shirt color is important, because only two people reported it", or "I know height isn't important because 1.2 million people reported it and it had no effect across all of them". As far as numbers go: don't think in terms of clinical trials studying 300 new mothers here or 1200 geriatric males there; think in terms of millions of patients treated daily across the nation - providing an unending *flood* of data points, each with an individual combination of factors *sure* to influence the data model in some way shape or form. In such an environment, no false pretense can survive for long; and no significant contraindication can go undetected.

And for drugs, that could be disastrous--for example, someone who weighs 300 lbs but has liver malfunction may be killed by a dose appropriate for someone who weighs 100 lbs and has normal liver function.

I should have said this up front - I think the basic safety/approval process shoul remain in place... Programs should only be able to select from drugs that have passed FDA approval (animal trials / clinical trials are well-fit for finding such contraindications as you mention above, and should be carried out well in advance of what I suggest). The data-mining would be more of an optimizer: once we know a range of dosages for a certain drug are safe enough for FDA approval; introduce it to the public at our "best guess" dosage, and let the program start it's work...

[Rolfe]

I recently heard a discussion of how useful this could be, but that it would require alteration of the law on medical confidentiality. The speaker maintained that it was important to have the identities of the patients so that multiple appearances in various medical environments could be collated, and that at the moment the requirements for strict anonymising of case reports were counter-productive.

His point was that there was sufficient information lurking unused in the cumulative health records of the population that significant benefits could be achieved if only researchers were allowed to investigate it without such strict confidentiality restrictions.

Rolfe.

[ChristineR]

That's just it: most drugs are not taken by millions of people. I'm out of my area here, but maybe someone else has info on how many people take drugs.

Here's a link to the top 300 of 2004: http://www.rxlist.com/top200.htm--from 92 million to 2 million. One database I found had 24,000 prescription drugs in it.

So I can see where it might work for the top 300, but not the other 23,000+.

[JamesM]

Do you have a particular type of data mining in mind, kk2796?

[kk2796]

Rolfe - I can certainly see the ethical tightrope involved. Sounds like another CS problem though... I can envision a system where most of a patient's medical info is kept private (no different than today), but other parts could be uploaded to a huge public database according to ANSI standards (with consensus on what data is and is not identifying and therefore should or should not be published). This might involve using a key from a one-way hash of an individual's name and a secret token (so individuals' records can be linked to each other and supplemented over time, but one couldn't use the public records to trace back to the individual or the individual's name to get their public records). As I said - it's a CS problem, but one unrelated to the datamining aspect...

James - interesting question. My initial inclination is to say a combination of techniques: cluster the database to identify people with "similar" body characteristics, then for each cluster, build an SVM across all symptoms, phyical attributes, and prescriptions to predicts the outcomes of said prescriptions. So when a patient arrives - take a physical, determine his/her cluster, punch in the desired results, and have a GA search the person's cluster's SVM input space for the drug/diet/behavior prescription most likely to give the desired results.

Then I thought to myself... WTF do I know? Here's a better idea - use capitalism. Make the public database easily available to everyone, then figure out a way to make building datamodels profitable (without crossing ethical lines of hiding the models from the poor). For instance, allow corporations, universities, etc., to make build data models and make them publically accessible... then have them compete for grant money by building superior models (superior models being determined by accuracy against new cases, evaluated on a regular basis). Other grants could be given for findings like "Conditions A + C + D are very unsafe".

[kk2796]

That's just it: most drugs are not taken by millions of people. I'm out of my area here, but maybe someone else has info on how many people take drugs.

Here's a link to the top 300 of 2004: http://www.rxlist.com/top200.htm--from 92 million to 2 million. One database I found had 24,000 prescription drugs in it.

So I can see where it might work for the top 300, but not the other 23,000+.

Right - again, the datamodel would be responsible for not only making suggestions, but expressing the confidence of it's suggestions. There should be some balancing acts going on: when it comes to new or rarely used drugs, data models should be seeded with FDA-recommended values... and work like a learning machines with each new case. The less experience it has with a drug, the less deviation it should have from FDA recommendations... additionally, for all drugs, the data models should be constrained by FDA-approved ranges and known contra-indicated combinations (which are informed by biochemistry, animal trials, clinical trials, and other sources of information that the model doesn't have at hand).

That said, if data models consitently recommend dosages at the extremes of these ranges, or consistently prefer certain combinations, the FDA ought to take note and followup with trials to loosen or tighten the system's constraints.