How to make Train-Test split on multivariate timeseries data

Question

I am building a model for the purpose of forcasting when someone is going into a stressful state.

I am using the WESAD dataset which has electrodermal activity (EDA) data on 11 subjects. I take this and use Neurokit2 to clean and extract features from the raw EDA data. The end result is that I have a list that stores each subject in the original dataset with 3 features and 1 label. The label is binary [0,1] and the features are normalized.

I only have experience running a timeseries model using a single factor and single subject. How would I correctly do the train-test split for multiple features on multiple subjects? Below is my code to create data generators for neural networks on one feature and one subject. Should I loop through each subject and do the same process as below? If I do as I suggest, how would I put this into a LSTM model?

from keras.preprocessing.sequence import TimeseriesGenerator

# Define the batch size
batch_size = 64
# Define the number of features and targets
num_features = 1
num_targets = 1
# Random State
random_state = 42
    
# Train Test Split
from sklearn.model_selection import train_test_split

# Validation split
X_dat, X_val, y_dat, y_val = train_test_split(subsampled_data, delayed_labels, 
                                                 test_size = 0.2,
                                                 random_state=random_state)
# Train test split
X_train, X_test, y_train, y_test = train_test_split(X_dat, y_dat,
                                                    test_size = 0.2,
                                                    random_state = random_state)


# Normalize the data
from sklearn.preprocessing import StandardScaler
# create the StandardScaler object
scaler = StandardScaler()
# fit the scaler on the training data
X_train_scaled = scaler.fit_transform(X_train.values.reshape(-1,1))
# transform the validation data
X_val_scaled = scaler.transform(X_val.values.reshape(-1,1))
# transform the test data
X_test_scaled = scaler.transform(X_test.values.reshape(-1,1))

# TimeSeriesGenerator parameters
shuffle = True

# Data Generator
train_data_gen = TimeseriesGenerator(X_train_scaled, y_train, 
                                     length=sequence_length, 
                                     batch_size=batch_size)
val_data_gen = TimeseriesGenerator(X_val_scaled, y_val, 
                                   length=sequence_length, 
                                   batch_size=batch_size)
test_data_gen = TimeseriesGenerator(X_test_scaled, y_test, 
                                    length=sequence_length, 
                                    batch_size=batch_size)

Answer 1

Scikit-learn's model_selection.TimeSeriesSplit is designed to appropriately split time series data. The result will include indices that can be used to reference the features, no matter how many features there are.