Hi everyone,
\n\nI have this concern about the raw data we got for this competition. Since the recorded Grasp-and-Lift tasks aren't imagery (require the subject to actually perform the movement) the data has a lot of information that comes from muscular activity. In fact this EMG information can be used to help with classification of each phase.
\n\nIf that was an EEG-only research such EMG information would be treated as noise an highpass filtered. Since that is not the case here, someone can use this EMG information to improve his/hers results. Although the accuracy will increase, this will no longer be the EEG classification problem.
\n\nI would like to know what's the case here? Is using this muscular movement related information fair and the "EEG" in the competition's name is just to gain attention, or is it in some way forbidden?
", "messages": [ { "id": "89471", "postDate": "08/15/2015 22:20:16", "content": "Hi everyone,
\n\nI have this concern about the raw data we got for this competition. Since the recorded Grasp-and-Lift tasks aren't imagery (require the subject to actually perform the movement) the data has a lot of information that comes from muscular activity. In fact this EMG information can be used to help with classification of each phase.
\n\nIf that was an EEG-only research such EMG information would be treated as noise an highpass filtered. Since that is not the case here, someone can use this EMG information to improve his/hers results. Although the accuracy will increase, this will no longer be the EEG classification problem.
\n\nI would like to know what's the case here? Is using this muscular movement related information fair and the "EEG" in the competition's name is just to gain attention, or is it in some way forbidden?
", "rawMarkdown": "Hi everyone,\r\n\r\nI have this concern about the raw data we got for this competition. Since the recorded Grasp-and-Lift tasks aren't imagery (require the subject to actually perform the movement) the data has a lot of information that comes from muscular activity. In fact this EMG information can be used to help with classification of each phase. \r\n\r\nIf that was an EEG-only research such EMG information would be treated as noise an highpass filtered. Since that is not the case here, someone can use this EMG information to improve his/hers results. Although the accuracy will increase, this will no longer be the EEG classification problem. \r\n\r\nI would like to know what's the case here? Is using this muscular movement related information fair and the \"EEG\" in the competition's name is just to gain attention, or is it in some way forbidden?", "votes": null }, { "id": "89473", "postDate": "08/15/2015 22:43:29", "content": "BinaryB, how would you separate the EMG from the EEG signal, what does it have special? I've read it is known which muscles influence activity on which electrodes, but how would one make the difference between neuronal and muscular activity?
", "rawMarkdown": "BinaryB, how would you separate the EMG from the EEG signal, what does it have special? I've read it is known which muscles influence activity on which electrodes, but how would one make the difference between neuronal and muscular activity?", "votes": null }, { "id": "89480", "postDate": "08/16/2015 00:44:24", "content": "The rules for this competition state that "future" data are not permitted in any computations (e.g., EEG signals when muscular activity is already taking place cannot be used for predicting onset of muscle movement). I would be more concerned about the usage of evoked activity from the visual cue as the predictor of movement, since it precedes the intent of movement.
", "rawMarkdown": "The rules for this competition state that \"future\" data are not permitted in any computations (e.g., EEG signals when muscular activity is already taking place cannot be used for predicting onset of muscle movement). I would be more concerned about the usage of evoked activity from the visual cue as the predictor of movement, since it precedes the intent of movement.", "votes": null }, { "id": "89504", "postDate": "08/16/2015 12:49:26", "content": "@Ana_4min
\n\nThe best way would be to have a reference EMG signal for adaptive filtering. Unfortunately that is not the case here. You could also use some blind source separation methods, but that would validate causality requirement.
\n\nHere, at least some parts of such artifacts would lay in low frequency range. These can be observed as significant changes in signal's trend and can be removed with highpass filtering. That would also help with eye movements.
", "rawMarkdown": "Ana_4min\r\n\r\nThe best way would be to have a reference EMG signal for adaptive filtering. Unfortunately that is not the case here. You could also use some blind source separation methods, but that would validate causality requirement.\r\n\r\nHere, at least some parts of such artifacts would lay in low frequency range. These can be observed as significant changes in signal's trend and can be removed with highpass filtering. That would also help with eye movements.", "votes": null }, { "id": "89505", "postDate": "08/16/2015 12:50:07", "content": "@JemmyChen
\n\nYou can still use this muscular information with some delay. That way your algorithm's output would stabilize faster on a correct value. lateralized readiness potential precedes the movement for only a few milliseconds. If you use muscle movement information here you would missclassify those only first few samples, but then it would be good.
\n\nOn the other hand I wouldn't worry that much about visually evoked potentials (VEP). You can discard channels that cover visual cortex. Even better solution would be to select only EEG channels from motor cortex.
", "rawMarkdown": "JemmyChen \r\n\r\nYou can still use this muscular information with some delay. That way your algorithm's output would stabilize faster on a correct value. lateralized readiness potential precedes the movement for only a few milliseconds. If you use muscle movement information here you would missclassify those only first few samples, but then it would be good.\r\n\r\nOn the other hand I wouldn't worry that much about visually evoked potentials (VEP). You can discard channels that cover visual cortex. Even better solution would be to select only EEG channels from motor cortex.", "votes": null }, { "id": "89523", "postDate": "08/16/2015 19:02:39", "content": "BinaryB, thank you. What sort of frequencies bands are muscle activity and eye movements expected to be in?
", "rawMarkdown": "BinaryB, thank you. What sort of frequencies bands are muscle activity and eye movements expected to be in?", "votes": null }, { "id": "89526", "postDate": "08/16/2015 20:36:42", "content": "@Ana_4min
\n\nI'd say that it is best to cut out frequencies below 1Hz to get rid of this unwanted trending.
\n\nWith eyeblinks (also very problematic next to eye movements) it is a bit more complicated. As I said the best practice would be to use additionally provided EOG signal for adaptive filter. Unfortunately, we don't have that. Most papers on ocular artifacts in EEG say that eyeblink related artifacts are strongest below 10Hz. That means they are mixed with brain's alpha activity.
\n\nMy suggestion would be to perform time-frequency analysis of short signal fragment with eye artifact, so that you have the idea how this works. Probably highpass filter with cut-off frequency set to 10-12Hz would do the trick. You must be aware of the tradeoff here: the higher cut-off the better ocular artifacts are filter, but also more EEG information is lost.
", "rawMarkdown": "Ana_4min\r\n\r\nI'd say that it is best to cut out frequencies below 1Hz to get rid of this unwanted trending. \r\n\r\nWith eyeblinks (also very problematic next to eye movements) it is a bit more complicated. As I said the best practice would be to use additionally provided EOG signal for adaptive filter. Unfortunately, we don't have that. Most papers on ocular artifacts in EEG say that eyeblink related artifacts are strongest below 10Hz. That means they are mixed with brain's alpha activity.\r\n\r\nMy suggestion would be to perform time-frequency analysis of short signal fragment with eye artifact, so that you have the idea how this works. Probably highpass filter with cut-off frequency set to 10-12Hz would do the trick. You must be aware of the tradeoff here: the higher cut-off the better ocular artifacts are filter, but also more EEG information is lost.", "votes": null }, { "id": "89585", "postDate": "08/17/2015 18:48:40", "content": "BinaryB, thank you.
", "rawMarkdown": "BinaryB, thank you.", "votes": null } ], "comments": [ { "id": 89473, "author_name": "ana4min", "author_url": "", "post_date": "08/15/2015 22:43:29", "content": "BinaryB, how would you separate the EMG from the EEG signal, what does it have special? I've read it is known which muscles influence activity on which electrodes, but how would one make the difference between neuronal and muscular activity?
", "votes": null, "replies": [] }, { "id": 89480, "author_name": "jemmyc", "author_url": "", "post_date": "08/16/2015 00:44:24", "content": "The rules for this competition state that "future" data are not permitted in any computations (e.g., EEG signals when muscular activity is already taking place cannot be used for predicting onset of muscle movement). I would be more concerned about the usage of evoked activity from the visual cue as the predictor of movement, since it precedes the intent of movement.
", "votes": null, "replies": [] }, { "id": 89504, "author_name": "binaryb", "author_url": "", "post_date": "08/16/2015 12:49:26", "content": "@Ana_4min
\n\nThe best way would be to have a reference EMG signal for adaptive filtering. Unfortunately that is not the case here. You could also use some blind source separation methods, but that would validate causality requirement.
\n\nHere, at least some parts of such artifacts would lay in low frequency range. These can be observed as significant changes in signal's trend and can be removed with highpass filtering. That would also help with eye movements.
", "votes": null, "replies": [] }, { "id": 89505, "author_name": "binaryb", "author_url": "", "post_date": "08/16/2015 12:50:07", "content": "@JemmyChen
\n\nYou can still use this muscular information with some delay. That way your algorithm's output would stabilize faster on a correct value. lateralized readiness potential precedes the movement for only a few milliseconds. If you use muscle movement information here you would missclassify those only first few samples, but then it would be good.
\n\nOn the other hand I wouldn't worry that much about visually evoked potentials (VEP). You can discard channels that cover visual cortex. Even better solution would be to select only EEG channels from motor cortex.
", "votes": null, "replies": [] }, { "id": 89523, "author_name": "ana4min", "author_url": "", "post_date": "08/16/2015 19:02:39", "content": "BinaryB, thank you. What sort of frequencies bands are muscle activity and eye movements expected to be in?
", "votes": null, "replies": [] }, { "id": 89526, "author_name": "binaryb", "author_url": "", "post_date": "08/16/2015 20:36:42", "content": "@Ana_4min
\n\nI'd say that it is best to cut out frequencies below 1Hz to get rid of this unwanted trending.
\n\nWith eyeblinks (also very problematic next to eye movements) it is a bit more complicated. As I said the best practice would be to use additionally provided EOG signal for adaptive filter. Unfortunately, we don't have that. Most papers on ocular artifacts in EEG say that eyeblink related artifacts are strongest below 10Hz. That means they are mixed with brain's alpha activity.
\n\nMy suggestion would be to perform time-frequency analysis of short signal fragment with eye artifact, so that you have the idea how this works. Probably highpass filter with cut-off frequency set to 10-12Hz would do the trick. You must be aware of the tradeoff here: the higher cut-off the better ocular artifacts are filter, but also more EEG information is lost.
", "votes": null, "replies": [] }, { "id": 89585, "author_name": "ana4min", "author_url": "", "post_date": "08/17/2015 18:48:40", "content": "BinaryB, thank you.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "89471": "Hi everyone,\r\n\r\nI have this concern about the raw data we got for this competition. Since the recorded Grasp-and-Lift tasks aren't imagery (require the subject to actually perform the movement) the data has a lot of information that comes from muscular activity. In fact this EMG information can be used to help with classification of each phase. \r\n\r\nIf that was an EEG-only research such EMG information would be treated as noise an highpass filtered. Since that is not the case here, someone can use this EMG information to improve his/hers results. Although the accuracy will increase, this will no longer be the EEG classification problem. \r\n\r\nI would like to know what's the case here? Is using this muscular movement related information fair and the \"EEG\" in the competition's name is just to gain attention, or is it in some way forbidden?", "89473": "BinaryB, how would you separate the EMG from the EEG signal, what does it have special? I've read it is known which muscles influence activity on which electrodes, but how would one make the difference between neuronal and muscular activity?", "89480": "The rules for this competition state that \"future\" data are not permitted in any computations (e.g., EEG signals when muscular activity is already taking place cannot be used for predicting onset of muscle movement). I would be more concerned about the usage of evoked activity from the visual cue as the predictor of movement, since it precedes the intent of movement.", "89504": "Ana_4min\r\n\r\nThe best way would be to have a reference EMG signal for adaptive filtering. Unfortunately that is not the case here. You could also use some blind source separation methods, but that would validate causality requirement.\r\n\r\nHere, at least some parts of such artifacts would lay in low frequency range. These can be observed as significant changes in signal's trend and can be removed with highpass filtering. That would also help with eye movements.", "89505": "JemmyChen \r\n\r\nYou can still use this muscular information with some delay. That way your algorithm's output would stabilize faster on a correct value. lateralized readiness potential precedes the movement for only a few milliseconds. If you use muscle movement information here you would missclassify those only first few samples, but then it would be good.\r\n\r\nOn the other hand I wouldn't worry that much about visually evoked potentials (VEP). You can discard channels that cover visual cortex. Even better solution would be to select only EEG channels from motor cortex.", "89523": "BinaryB, thank you. What sort of frequencies bands are muscle activity and eye movements expected to be in?", "89526": "Ana_4min\r\n\r\nI'd say that it is best to cut out frequencies below 1Hz to get rid of this unwanted trending. \r\n\r\nWith eyeblinks (also very problematic next to eye movements) it is a bit more complicated. As I said the best practice would be to use additionally provided EOG signal for adaptive filter. Unfortunately, we don't have that. Most papers on ocular artifacts in EEG say that eyeblink related artifacts are strongest below 10Hz. That means they are mixed with brain's alpha activity.\r\n\r\nMy suggestion would be to perform time-frequency analysis of short signal fragment with eye artifact, so that you have the idea how this works. Probably highpass filter with cut-off frequency set to 10-12Hz would do the trick. You must be aware of the tradeoff here: the higher cut-off the better ocular artifacts are filter, but also more EEG information is lost.", "89585": "BinaryB, thank you." }, "source": "meta" }