1
$\begingroup$

I have not found any resource on how to use a neural network to model measurements from a repeated-measures design. Let's say we have a small dataset (for example, 14 independent variables, 15-repetitions, and 50 subjects) of measured continuous variables as independent variables, plus some constant-valued independent variables such as age and sex. We also have one dependent variable measured at each repetition. I have tried to use a feedforward neural network with 10 neurons in one hidden layer to model this, and one-subject-out cross validation, but I think it can be improved maybe by following techniques: 1. Discretization of continuous input variables, since the dataset is small. 2. Using a feedback in training the network since a dependent variable (output) for each subject could be affected by the output from preceding outputs. 3. Remove redundant input variables using feature selection approaches.

The main problem is that I do not know how to input the constant-valued variable (age) to this model. And how to put this feedback into the model. Finally if I got the model accurate enough in one-subject-out validation, is it fine to average the weights and test this new model on the dataset and if it was good enough use the model for unseen data?

$\endgroup$
  • $\begingroup$ What are you actually trying to achieve with the model? Is this a regression problem, a classification problem, a little bit of both? $\endgroup$ – jkm Jun 29 '18 at 12:37
  • $\begingroup$ This is a regression problem but can be turned also into a binary classification problem.@jkm $\endgroup$ – RZM Jul 2 '18 at 9:13
2
$\begingroup$

I've been struggling with the same problem and finally stumbled across an article entitled "Subject Specific Treatment to Neural Networks for Repeated Measures".

They recommend one-hot encoding the subject ID and using that as an additional input to the model (and they specifically suggest that it's better to provide the input to a deeper hidden layer, not the first). However, if you need to make predictions on new data, it's not exactly clear how best to do that. Their solution strikes me a pretty clever, and it is probably worth testing your architecture with and without the subject ID input.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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