Dear researchers

Objective:

We've applied machine learning methods such as artificial neural networks, random forest, and support vector machines to predict stroke patient's recovery.

Materials and methods:

We have stroke patients' clinical data from EMRs(electronic medical records) and their kinematic data obtained by the exoskeleton robot's sensor system(from gait training).

The clinical data are ordinal and categorical, and the kinematic data are time-series data.

Clinical data and kinematic data have been integrated into tabular data by applying moving windows to time-series data (obtained mean, std, median, max, and min).

Limitations:

In our experience, it was not easy to use all the data for training at once because the types and characteristics of clinical data and kinematic data were different.

Thus, we are applying the ensembling method to various neural network models.

(We've tried conventional bagging or stacking algorithms to the outputs of the neural networks.)

Question:

At this point, we would like to know some reasonable, preferred, recommended methods for ensembling the neural network models with different data learned separately. (i.e., how to combine a neural network model trained by clinical data and another model trained by kinematic data)