1
$\begingroup$

I am thinking of using machine learning type classification models to compare them with the traditional approach of logistic regression (dichotomous outcome) in a sample of patients with diabetic foot where the dependent variable (y) would be the level of severity of diabetic foot: low (reference) , high(1). Then, it would have many dependent variables such as age, sex, blood glucose, cholesterol, glycosylated hemoglobin, etc.

My question is, apart from obtaining precision metrics in Machine Learning classification (ROC, sensitivity, accuracy etc), is it possible to obtain some odds ratio or relative risk with the top machine learning models in benchmarking?

$\endgroup$
1
  • $\begingroup$ Note that when you have some output from different models for the same patients (or even using the same model - or multiple models from the same model class - repeatedly on the same patient), you should also account for this in your model (i.e. the outcomes for these records are correlated, so e.g. could use a fixed or random effect for case). The (logistic) regression model you describe would not do that, but can obviously adapted to do it. $\endgroup$
    – Björn
    Nov 23, 2023 at 11:38

1 Answer 1

1
$\begingroup$

This is a problematic way to start. This is not a setting for forced-choice classification but rather for estimating tendencies, i.e., probabilities (risk). And the accuracy measures you mentioned are problematic. Use a direct probability model, which may be built upon relative log odds (logistic model) or relative risk (other models, usually fit less well). Quantities such as odds ratios are building blocks to allow one to estimate overall absolute risk. Though logistic regression is often the best solution (especially if you use modern versions of it that don’t make linearity assumptions) but there are machine learning algorithms for estimating probabilities (but not through odds/risk ratios).

$\endgroup$
2
  • $\begingroup$ Thank you Frank for putting greater emphasis on the differences between both approaches. Your help has been useful for me. $\endgroup$
    – ronald
    Nov 27, 2023 at 12:08
  • 1
    $\begingroup$ I misstated one thing. You can build a neural network around a logistic regression model (which uses odds ratios) just as you can with the Cox proportional hazards regression model. $\endgroup$ Nov 27, 2023 at 16:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.