I think it will be an interesting approach if you try to use recurrent NN like LSTM to model all participants data as time series data. LSTM has the ability to remember important information from the past as well as forget information which is not important. So when the predictor variables go to the next subject, the model can forget those information that are not relevant to the new subject. In this way, you will get a more generalized model. Now, you can augment your model strength to perform better on a personal level by concatenating additional learning models/algorithms based on some personal traits or context as features.

In one of my pervious work, where my objective was to come with a more generalized model, I used this technique, and it worked really well. You can refer to this paper.

Article Machine Learning Based Stress Monitoring in Older Adults Usi...

Try ARIMA modelling.

A NN behaves as statistical filter. If you there is very little data, it must not have too much noise in the signal if it is used for forecasting. The easiest way to handle this problem is to check your monitoring data collection methodology to make sure it collects only data that you intend to observe. If you are collecting unrelated events (which is very common), then you will find that a NN will not stabilize to offer statistically significant predictions. The data collection process is the key to stabilizing a NN, not the modeling technique. Filter out data not related to the specific question you seek to answer. That is the best anyone can do.

Thank you Rajdeep Kumar Nath . I had a look at your interesting paper.

So what you are saying is that it is possible to feed LSTM with all subjects data at once (after being splitting into training and testing) and create a generic model. This would still be considered as time series data? Or time series data should have temporal sequence order from the first to the last data instance?