I would like to analyse head rotation data in space. For this I measured at 15HZ the rotation around the X, Y and Z angles for a little more than ten minutes. I would like to use these movements to analyse a second variable, also continuous, but with a single measurement. I have 35 participants, and if necessary I can use only one of the axes that is more informative than the others (Z axis). I have already done the classic analyses (velocity, coefficient of variation etc), but I would like to explore the possibility of training a model to predict the second variable, but all I can find is to predict a data item following the time series. Any ideas? Thanks in advance!
$\begingroup$ Please clarify your specific problem or provide additional details to highlight exactly what you need. As it's currently written, it's hard to tell exactly what you're asking. $\endgroup$– Community BotMar 11 at 21:25
If the second variable is a binary variable indicating whether the participant got dizzy or not, you need to consider other factors beyond the raw head rotation data to accurately predict whether the participant experienced dizziness.
One factor that can contribute to dizziness is the movement of fluid in the inner ear. To capture this, you can compute the velocity and acceleration of head movements along the X, Y, and Z axes, and analyze the frequency components of these signals. Specifically, you can look for sudden changes in acceleration or velocity, and check if they occur at frequencies that are known to be associated with inner ear disturbances.
Another factor that can contribute to dizziness is the orientation of the head relative to gravity (I don´t know if relative the the space station floor or Earth). To capture this, you can compute the orientation of the head in space, and look for changes in orientation that could result in a mismatch between the visual and vestibular systems.
In addition to these factors, you should also consider other potential contributors to dizziness, such as changes in blood pressure or heart rate, and use them to inform the features you extract and the model you build.
IF It is that what you are after... it seems to be a complex problem. It requires feature selection and to do the feature engineering you need a deep understanding of the underlying physiological mechanisms.
The type of task you are describing is sometimes called time series extrinsic regression. There's not a lot of literature about, this but a paper that provides a good introduction to and evaluation of some ML algorithms is Tan et al.'s Time series extrinsic regression. [Disclaimer: I'm part of the same research group as some of the authors].
I noticed you tagged this "neural-network". In principle, any NN for time series classification can be readily adapted for extrinsic regression by simply replacing the softmax output layer with a linear layer. However, 35 instances is not a lot of data to train a neural network.