including Liu2024 Dataset

[tahatt13]

The dataset includes data from 50 acute stroke patients, and it involves two motor imagery tasks: right-handed and left-handed movements.
Here are more details about the dataset :

Name	#Subj	#Chan	#Classes	#Trials / class	Trials len	Sampling rate	#Sessions
Liu2024	50	29 EEG	2	40	4 sec	500 Hz	1

[bruAristimunha]

The question here is: is this dataset good enough to be merged?

[bruAristimunha]

Check the events

[bruAristimunha]

Hey @sebVelut!

Can you please review this code by next Friday? A shallow review is good!

[sebVelut]

Hi !
I have reviewed and tested it by getting data with paradigm.get_data() and by testing WithinSessionEvaluation with a Xdawn + TS + LDA and everything worked without problems. The code seems good in my point of view.

But I think you should be careful about the events id described as said by Bruno, it should be the same as in the mapping.
Moreover, the use of annotations seems to work, another idea is to use a stim channel (or transform the existing one) and put the label of the events as value of the channel. Then the moabb will search for the stim channel first (this method seems to be taken by several other dataset implementation, ex: Physionet_mi or mpi_mi). Up to you if you decide to change the method.

A last point, in whats_new.rst, it is missing the line to say that you have added the Liu dataset

[bruAristimunha]

Updating the line

[PierreGtch]

Thanks @tahatt13 ! Here are my comments

[PierreGtch]

@bruAristimunha is it ok to change the pre-commit file here?

[PierreGtch]

I’m not sure if we should include the instruction and break events...
Because then by default, when using this dataset with the MotorImagery paradigm, algorithms would also have to classify these events.
Plus, will the instructions and break epochs really correspond to these phases? i.e. 4s interval, starting 2s after the event.

I suggest we don’t include them by default. And eventually we can leave the possibility to get them with dataset arguments:

Suggested change

	def __init__(self):
	super().__init__(
	subjects=list(range(1, 50 + 1)),
	sessions_per_subject=1,
	events={"left_hand": 1, "right_hand": 2, "instr": 3, "break": 4},
	code="Liu2024",
	interval=(2, 6),
	paradigm="imagery",
	doi="10.1038/s41597-023-02787-8",
	)
	def __init__(self, break_events=False, instrumental_events=False):
	events = {"left_hand": 1, "right_hand": 2}
	if break_events:
	events[“instr”] = 3
	if insert_events:
	events[“break”] = 4
	super().__init__(
	subjects=list(range(1, 50 + 1)),
	sessions_per_subject=1,
	events=events,
	code="Liu2024",
	interval=(2, 6),
	paradigm="imagery",
	doi="10.1038/s41597-023-02787-8",
	)

[tahatt13]

I added that possibility. If I understood correctly what you meant, I think the epochs will be respected since the duration and onset are respected in the events file as you can see the break and instructions last for exactly 4seconds after each stimulu:

[PierreGtch]

The parameters break_events and inst_events are not working:

data = Liu2024(break_events=False, instr_events=False).get_data([1])
data[1]['0']['0'].annotations
#  <Annotations | 120 segments: break (40), instr (40), left_hand (20), ...>

( you still get annotations with all possible events even if you set these parameters to False)

The rest is great :)

[PierreGtch]

Thanks @tahatt13 :D