I'm currently utilizing the Temporal Fusion Transformer (TFT) to predict nursing workload for each patient over the next 3 shifts in a hospital, using past workload data and additional features. My dataset comprises 30,000 patient time series of varying lengths.
The model should predict the workload for the subsequent shifts regardless of the patient's length of stay in the hospital. Thus, it should be able to predict each timestep within a patient's time series based on the corresponding past true values.
My goal is to evaluate how well the model performs in relation to the number of available past values. For this, the TFT needs to return the full output sequence, similar to the return_sequences=True parameter in LSTMs. Could someone provide insights on how to enable this functionality in the TFT? Thank you very much!
I've already tried to divide each patient time series into "sub-timeseries" with increasing number of timesteps and then predicting the last three shifts. However, this results in a dataset that is far too large.