Hi everyone,
\nI am wondering which approach is best between hand-made interpretable feature engineering on raw EEG epochs with a classification machine learning model on these features and deep learning approach as the deep model can learn informative features directly from the raw EEG signals.
\nFor classical feature engineering ideas, I am thinking of some features :
\nFor Deep Learning techniques, I am thinking of CNN model for the best feature learning approach but have not much experience about it.
\nAlso, in the medical field, interpretability of the solution is key to understand what the model is doing, so maybe manual feature engineering is a lot more interpretable than a deep learning approach.
\nFeel free to correct any misunderstanding and share your thoughts below ! 😃
\nEnjoy the competition !
", "messages": [ { "id": 2594224, "postDate": "2024-01-09T18:14:08.083Z", "content": "Hi everyone,
\nI am wondering which approach is best between hand-made interpretable feature engineering on raw EEG epochs with a classification machine learning model on these features and deep learning approach as the deep model can learn informative features directly from the raw EEG signals.
\nFor classical feature engineering ideas, I am thinking of some features :
\nFor Deep Learning techniques, I am thinking of CNN model for the best feature learning approach but have not much experience about it.
\nAlso, in the medical field, interpretability of the solution is key to understand what the model is doing, so maybe manual feature engineering is a lot more interpretable than a deep learning approach.
\nFeel free to correct any misunderstanding and share your thoughts below ! 😃
\nEnjoy the competition !
", "rawMarkdown": "Hi everyone,\n\nI am wondering which approach is best between hand-made interpretable feature engineering on raw EEG epochs with a classification machine learning model on these features and deep learning approach as the deep model can learn informative features directly from the raw EEG signals.\n\nFor classical feature engineering ideas, I am thinking of some features :\n- Statistics on raw epoch of EEG signal (time features)\n- Frequency features\n- Time-frequency features based on coefficients of the wavelet transform\n- Hjorth parameters : [Wikipedia link](https://en.wikipedia.org/wiki/Hjorth_parameters)\n- Power Spectral density, Entropy features\n- Features from EEG waves : delta: below 3.5 Hz (0.1–3.5 Hz), theta: 4–7.5 Hz, alpha: 8–13 Hz, beta: 14–40 Hz, and gamma: above 40 Hz. \n\nFor Deep Learning techniques, I am thinking of CNN model for the best feature learning approach but have not much experience about it.\n\nAlso, in the medical field, interpretability of the solution is key to understand what the model is doing, so maybe manual feature engineering is a lot more interpretable than a deep learning approach.\n\n\nFeel free to correct any misunderstanding and share your thoughts below ! 😃\n\nEnjoy the competition !", "votes": 13 }, { "id": 2619652, "postDate": "2024-01-25T15:42:34.227Z", "content": "It is an interesting question! I think a model might capture these features implicitly before overfitting if there could be enough data. However, I see that some manual feature engineers helped models learn general patterns before overfitting in my limited experience with Kaggle.
", "rawMarkdown": "It is an interesting question! I think a model might capture these features implicitly before overfitting if there could be enough data. However, I see that some manual feature engineers helped models learn general patterns before overfitting in my limited experience with Kaggle.", "votes": 1 }, { "id": 2629492, "postDate": "2024-01-31T20:14:10.840Z", "content": "I have done a lot of feature engineering in my catboost notebook, and indeed the score improves but not by a very large margin. The most important features stay simple features such as the mean eeg signal value for the middle 10 seconds. But they're very represented in the other top 25 features.
", "rawMarkdown": "I have done a lot of feature engineering in my catboost notebook, and indeed the score improves but not by a very large margin. The most important features stay simple features such as the mean eeg signal value for the middle 10 seconds. But they're very represented in the other top 25 features.", "votes": 2 }, { "id": 2629094, "postDate": "2024-01-31T16:09:46.597Z", "content": "The Hjorth parameters look really useful. The PDs and RDAs require a semi-constant frequency, so I imagine the Hjorth complexity would be useful here. I wonder if some of these parameters might help us differentiate between lateralized and generalized versions too. Thanks for this!
", "rawMarkdown": "The Hjorth parameters look really useful. The PDs and RDAs require a semi-constant frequency, so I imagine the Hjorth complexity would be useful here. I wonder if some of these parameters might help us differentiate between lateralized and generalized versions too. Thanks for this!", "votes": 2 }, { "id": 2594754, "postDate": "2024-01-10T03:57:37.140Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 2595028, "postDate": "2024-01-10T07:36:05.393Z", "rawMarkdown": "", "votes": 3, "isDeleted": true } ] } ], "comments": [ { "id": 2619652, "author_name": "makio323", "author_url": "", "post_date": "2024-01-25T15:42:34.227000", "content": "It is an interesting question! I think a model might capture these features implicitly before overfitting if there could be enough data. However, I see that some manual feature engineers helped models learn general patterns before overfitting in my limited experience with Kaggle.
", "votes": 1, "replies": [] }, { "id": 2629492, "author_name": "stefanoclss", "author_url": "", "post_date": "2024-01-31T20:14:10.840000", "content": "I have done a lot of feature engineering in my catboost notebook, and indeed the score improves but not by a very large margin. The most important features stay simple features such as the mean eeg signal value for the middle 10 seconds. But they're very represented in the other top 25 features.
", "votes": 2, "replies": [] }, { "id": 2629094, "author_name": "Tyler Feemster", "author_url": "", "post_date": "2024-01-31T16:09:46.597000", "content": "The Hjorth parameters look really useful. The PDs and RDAs require a semi-constant frequency, so I imagine the Hjorth complexity would be useful here. I wonder if some of these parameters might help us differentiate between lateralized and generalized versions too. Thanks for this!
", "votes": 2, "replies": [] }, { "id": 2594754, "author_name": "", "author_url": "", "post_date": "2024-01-10T03:57:37.140000", "content": "", "votes": 0, "replies": [ { "id": 2595028, "author_name": "", "author_url": "", "post_date": "2024-01-10T07:36:05.393000", "content": "", "votes": 3, "replies": [] } ] } ], "raw_markdown_by_id": { "2594224": "Hi everyone,\n\nI am wondering which approach is best between hand-made interpretable feature engineering on raw EEG epochs with a classification machine learning model on these features and deep learning approach as the deep model can learn informative features directly from the raw EEG signals.\n\nFor classical feature engineering ideas, I am thinking of some features :\n- Statistics on raw epoch of EEG signal (time features)\n- Frequency features\n- Time-frequency features based on coefficients of the wavelet transform\n- Hjorth parameters : [Wikipedia link](https://en.wikipedia.org/wiki/Hjorth_parameters)\n- Power Spectral density, Entropy features\n- Features from EEG waves : delta: below 3.5 Hz (0.1–3.5 Hz), theta: 4–7.5 Hz, alpha: 8–13 Hz, beta: 14–40 Hz, and gamma: above 40 Hz. \n\nFor Deep Learning techniques, I am thinking of CNN model for the best feature learning approach but have not much experience about it.\n\nAlso, in the medical field, interpretability of the solution is key to understand what the model is doing, so maybe manual feature engineering is a lot more interpretable than a deep learning approach.\n\n\nFeel free to correct any misunderstanding and share your thoughts below ! 😃\n\nEnjoy the competition !", "2619652": "It is an interesting question! I think a model might capture these features implicitly before overfitting if there could be enough data. However, I see that some manual feature engineers helped models learn general patterns before overfitting in my limited experience with Kaggle.", "2629492": "I have done a lot of feature engineering in my catboost notebook, and indeed the score improves but not by a very large margin. The most important features stay simple features such as the mean eeg signal value for the middle 10 seconds. But they're very represented in the other top 25 features.", "2629094": "The Hjorth parameters look really useful. The PDs and RDAs require a semi-constant frequency, so I imagine the Hjorth complexity would be useful here. I wonder if some of these parameters might help us differentiate between lateralized and generalized versions too. Thanks for this!", "2594754": "" } }