1

I am very new to Machine Learning. I've read about it and I successfully did a tutorial where we looked at the dimensions of Iris petals and predicted what other Iris petal sizes may be. This all made sense because we were just looking at numbers.

I'm having trouble taking the next step. Ultimately the project I want to accomplish is this:

  1. Train with a set of diagnosis codes from emergency room visits.

  2. Determine a trend for example, let's say people with asthma then develop COPD.

  3. Then test with data of other people. The idea being to say, these people have asthma, they may develop COPD.

The issue I'm encountering making the leap is that the Diagnosis codes are not numerical. If you can suggest how to approach this or give me hints on what to study, I would appreciate it.

Thank you

Funkavenger
  • 11
  • 1
    I don't see this example as a Machine Learning problem. If you want to say something like "You were diagnosed with asthma so you have a 4% chance of developing COPD within 5 years" then what you need is statistics. It can test such a hypnosis just fine. Machine Learning might come in handy if you had no hypnosis at all and wanted to look at the data and try to mine one out of it. – candied_orange Nov 15 '18 at 01:00

1 Answers1

2

ML problems can broadly be divided into regression problems where we want to predict numerical values, and classification problems where we want to predict a categorical value. However, these two are related and you can often turn a classification problem into a regression problem, e.g. by encoding each class as a variable that can be zero or one. So instead of a single variable Diagnosis ∈ { Asthma, COPD, … } you might have multiple variables { Asthma ∈ [0,1], COPD ∈ [0,1], … }.

Note that there is no clear delineation between ML algorithms and computational statistics, i.e. ordinary statistics that a computer makes feasible. Sometimes it may be appropriate to simply calculate the correlations between the diagnosis in one time slice with the diagnosis in the previous time slice, and use that correlation to predict probabilities for a given diagnosis in the next time slice. If appropriate care is taken, an encoding of variables on the interval [0, 1] can be directly interpreted as a probability. Even if no closed form for such calculations is available, Monte Carlo techniques can often produce good results.

amon
  • 134,135
  • I figured there is a way to encode like you're describing. Thank you so much- this glossary will help me move forward. Also, to clarify, ultimately I want to take all diagnosis codes and aggregate them to look for trends across wide sets of data. I just used the most basic of examples to get me past my rut. – Funkavenger Nov 15 '18 at 14:55
  • (Remark) In this question, one or more columns of *inputs* are categorical, in addition to one or more outputs that are categorical. – rwong Nov 15 '18 at 17:59