Rodz Rants: Medical Marijuana and Machine Learning (caution: geeky)

 

 

backprop1

(above is the first time I got backprop to work in a ML NN exercise. I was so thrilled I got it tattooed on my back !)

Rodz Rant: the Medical Marijuana part

We think we know a lot about marijuana strains, their effects, and what problems they are good for.  Google anything and you will get page upon page of advice, opinion, and recommendations. So I googled anything and started reading, but soon found conflicting experiences and advice. What is the difference (chemically) between indicas and sativas? Why do some people get one effect and others get a totally different effect from the same dose of the same strain? How can any practitioner make a rational recommendation for any one patient?

Rodz Rant: segue

So how can we make recommendations to our patients? Well, like any complex system we can use rules of thumb, “heuristics,” to make a better-than-chance recommendation. And these rules of thumb work pretty well. It’s like when you look outside in the morning and decide you will take your raincoat to work today. So that is our level of understanding and recommendation so far. But I think we can do a lot better than this.

Rodz Rant: the Machine Learning part

In modeling terms, recommending the right medical marijuana preparation for any one patient resembles the kind of system that Amazon, Google, or Netflix uses to recommend a product to you.  In technical terms, it is a multivariate logistic regression problem with multiple output classes.  And it turns out that neural network modeling techniques are built do do this kind of stuff.  So I think that we will start making way better recommendations when we start plugging our outcomes data into a system like this. There’s a lot of math, but none of it is overly complicated. The goal is to improve on our rule of thumb recommendations to help the next patient choose a medical marijuana medicine that has a high probability of working from the start.

Rodz Rant: conclusion

This kind of ties into the “personalized medicine” thread that is beginning to gain traction in cancer chemotherapy and other medical fields.  The domain of medical marijuana is sufficiently new so that you don’t have to fight against entrenched pharmaceutical interests to make this kind of thing happen.  We are starting to collect data and build models now.

Watch This Space.