New to data science and am trying to be a self-starter and implement advanced data analytics in my subspecialty of surgery. Below is a description of my data set. I know that I will have to explore multiple methods, but wanted to get your take on which you think may be best. I will most likely be using R to achieve this analysis.

  • Have a data set with about 200 patients (rows)
  • Each patient has about 10-15 variables (preoperative and intraoperative)
  • Each patient has undergone either nonoperative or operative management
  • Success in nonoperative or operative management is determined by a questionnaire that patients fill out 1 year after they are seen. This questionnaire gives a binary outcome on whether they (1) Benefited or (2) Did not Benefit from the surgery.

My questions for the study are as follows:

  1. In the surgery group, I am trying to find out which variables lead to patients (1) Benefit vs (2) Do not benefit from surgery, and create a model which can better help predict which patients we can operate on (I have left out some details such as patient population, type of surgery, etc).
  2. In the second study, I would like to determine which patients we should operate on. In other words, I would like to find out which preoperative characteristics make some patients more likely to benefit from (1) Operative treatment vs (2) Nonoperative treatment and in this case the outcome will also be binary from the questionnaire.

I have tried linear and logistic regressions for this which have not been very good, hence why I am trying to learn more advanced models.

Models which are easier to comprehend by clinicians are more valuable which is why I haven't delved into neural nets. I appreciate any and all advice that can be provided. In addition, if I expand this data set to 600 people, would you use another model? I don't have access to large servers so most of this will be done on my laptop though I can use online resources if necessary (Azure etc).

Thank you all for your help and input.

  • $\begingroup$ You only have 200 data points with 15 features. Please note that deep learning is usually applied to 10,000s of training examples and 100s of features. $\endgroup$ Dec 17, 2018 at 6:38

1 Answer 1


Question 1: I would propose Decision Trees as a first thing you should look into, as they are easy to implement and the results are straightforward to interpret. You will end up with a tree-structure, where the nodes hold intervals/values of your variables. At each node, the tree tries to split your data most efficiently into your binary classifications $C_1$ and $C_2$. Thus, in the end, you will be able to extract most important features in your variables at the top of the tree. A typical measure of "efficiency" when selecting attributes is Entropy.

Question 2: I think it is also solved with the above method. For each node of the tree you know the amount of $C_1$ and $C_2$ classifications on the left and on the right. Assume your attribute in the node is "Has diabetes". Your data (and thus the tree) tells you, 100 out of 110 patients "failed" the operation if the person had diabetes. Thus, you can give an estimate about how certain attributed contribute to success of your operations, and conclude it would not make much sense to operate the given person.

Nonetheless, always be careful about how you interpret the results. You should not try to interpret attributes that split small subgroups, since this can easily lead to false assumptions about real-world behaviors.

  • $\begingroup$ Thanks for your response, Andre. Would testing this data on a blinded set from another institution make this more validated since if it performs well it would be more generalizable? Further, do you think Random Forest with or without bootstrapping make this model more robust or do you think the N (N from 200-500 with 12-15 variables for each N) I have in my data set is not large enough for such a method. $\endgroup$
    – AYP
    Sep 18, 2018 at 5:17
  • $\begingroup$ Validation through blinded set: Good idea! Random forests can improve your outcome, while making your results harder to interpret. Saying what's better from this point is very hard, since it highly depends on how well your data splits. Since Decision Trees are easy to implement, I would try them out (with pruning to combat overfitting), validate through your test set and see how it performs. Regarding $N$, I just can say the same: If the data splits nicely, 200 is OK; if that does not happen, even 500 can be too little. If this does not work, you can try Random forests. $\endgroup$
    – André
    Sep 18, 2018 at 8:17

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.