Estimating joint angle using machine learning methods

I am currently working on a project which uses IMU sensor data in order to predict the Knee Angle at each time point. My input data consists of 9 sensors, I plotted an example section below.

The goal of the algorithm now is to find the point where the knee angle reversal takes place (indexed in red) or predict the knee angle over the whole time section. This would depend either I choose a classification model (1 for a swing phase reversal, 0 for no reversal) or a regression model predicting the knee angle.

I already tried creating an input matrix taking sensor values from 10 time steps (1 actual + 9 from the past) and the knee angle as an expected value and feed it into a CNN. (Matrix 9x10)

I also thought about an LSTM but I am stuck in the modelling.

• How to choose the input length
• Overlapping of data, currently my step size is 1. So the data is fully overlapped.

My new approach should look like this, however I am not able to implement such model in MATLAB as I cant use a 2D Conv on sequence data. https://www.researchgate.net/publication/313643430_Motor_Imagery_Classification_Based_on_Deep_Convolutional_Neural_Network_and_Its_Application_in_Exoskeleton_Controlled_by_EEG

Is there any other solution I am not thinking of which can be used within MATLAB?