Confusion about the length of the input signal

[CuiTingXu]

Hello, I am learning about your library recently. I want to know if there is any limit on the input length of the four models trained? If I want to test my input with the first three models 'dFC_2s.pkl', rFC_2s.pkl, and 'rfc_2s_norm.pkl', do I need to fix my signal to the length of 2s? Similarly, if I need to use the 'xGB_9s. joblib' classifier, do I have to fix my test signal to a length of 9s? I really hope you can reply to me, thank you!

[alexisgcomte]

Hello JaeCao, thank you for your question, this point has to be clearer in documentation to come.

Models have been trained with different parameters of preprocessing, ie input seconds and normalization before SQI computing. Nothing prevents you from inputing a longer segment: SQIs will be computed the same way, but their distribution will change slightly as they witness more signal. However, the thresholds computed by Decision Trees models (with or without ensemble learning) should stay consistant and it could even improve predictions. Please note that the sampling frequency parameter is very important in the computing, so it must match your sample.

Most importantly, present models are dependent of the characteristics of the signal we have used for training:

• ECGs can have different distribution of signals
• We have re-annoted signal to better suit "our" notation of quality (this part will be documented later) and the prediction depend on our exigency

Therefore, the safest way to use this library is to test it on a sample and analyze results. Currently, models do not generalize on all kind of signals. When the time allows, we plan on publishing the automated training pipeline we used for community to train on new signal.

Feel free to ask for more questions!