{
  "id": 19523,
  "title": "Accurate to within ~10% is impressive",
  "url": "/competitions/second-annual-data-science-bowl/discussion/19523",
  "author_name": "",
  "post_date": "2016-03-15T01:50:13.690Z",
  "votes": 2,
  "comment_count": 3,
  "views": 963,
  "content": "<p>Does anyone else think this is pretty impressive? When I considered that the manual labeling would be necessarily idiosyncratic, with a lot of judgement calls on the edge cases, I didn't think it possible on the test dataset. Congratulations to everyone near the top.</p>",
  "messages": [
    {
      "id": "111510",
      "postDate": "03/15/2016 01:50:13",
      "content": "<p>Does anyone else think this is pretty impressive? When I considered that the manual labeling would be necessarily idiosyncratic, with a lot of judgement calls on the edge cases, I didn't think it possible on the test dataset. Congratulations to everyone near the top.</p>",
      "rawMarkdown": "Does anyone else think this is pretty impressive? When I considered that the manual labeling would be necessarily idiosyncratic, with a lot of judgement calls on the edge cases, I didn't think it possible on the test dataset. Congratulations to everyone near the top.",
      "votes": null
    },
    {
      "id": "111530",
      "postDate": "03/15/2016 07:32:53",
      "content": "<p>Here are my results.\nIt's even better than 10%\n<img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/3909/results.png\" alt=\"enter image description here\" title></p>\n\n<p>Validate mi is biggest error too low and validate mx is biggest error to high.\nFold 4 contains 2 corrupt patients (595 and 599) with 3 slices. (hence the big mi)\nApart from the 2 corrupt patients there were some big errors from wrongly provided volumes (esp patient 429)</p>\n\n<p>I used only 1 model for everything.\nThe main thing that limited my performance was not knowing how to label and wrongly provided volumes. It tried to make up for this by doing some calibration.</p>\n\n<p>So I basically consider this problem solved.</p>",
      "rawMarkdown": "Here are my results.\r\nIt's even better than 10%\r\n![enter image description here][1]\r\n\r\nValidate mi is biggest error too low and validate mx is biggest error to high.\r\nFold 4 contains 2 corrupt patients (595 and 599) with 3 slices. (hence the big mi)\r\nApart from the 2 corrupt patients there were some big errors from wrongly provided volumes (esp patient 429)\r\n\r\n  [1]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/3909/results.png\r\n\r\nI used only 1 model for everything.\r\nThe main thing that limited my performance was not knowing how to label and wrongly provided volumes. It tried to make up for this by doing some calibration.\r\n\r\nSo I basically consider this problem solved.",
      "votes": null
    },
    {
      "id": "111904",
      "postDate": "03/17/2016 07:54:22",
      "content": "<blockquote>\n  <p>So I basically consider this problem solved.</p>\n</blockquote>\n\n<p>I agree, I doubt people doing medical research realize what is possible in terms of automation. I sure didn't. There are so many problems similar to this that consume 2-3 years of someones life. </p>",
      "rawMarkdown": "> So I basically consider this problem solved.\r\n\r\nI agree, I doubt people doing medical research realize what is possible in terms of automation. I sure didn't. There are so many problems similar to this that consume 2-3 years of someones life.",
      "votes": null
    },
    {
      "id": "111995",
      "postDate": "03/17/2016 18:43:09",
      "content": "<p>Problem solved? Probably not for clinical applications. It would be interesting to run an ejection fraction classifier using the same techniques you used in the competition. I'm guessing, given your numbers, that there would way too many false positives and, more critically, false negatives for this technique to be used in practice. But it would be an interesting exercise.</p>",
      "rawMarkdown": "Problem solved? Probably not for clinical applications. It would be interesting to run an ejection fraction classifier using the same techniques you used in the competition. I'm guessing, given your numbers, that there would way too many false positives and, more critically, false negatives for this technique to be used in practice. But it would be an interesting exercise.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 111530,
      "author_name": "juliandewit",
      "author_url": "",
      "post_date": "03/15/2016 07:32:53",
      "content": "<p>Here are my results.\nIt's even better than 10%\n<img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/3909/results.png\" alt=\"enter image description here\" title></p>\n\n<p>Validate mi is biggest error too low and validate mx is biggest error to high.\nFold 4 contains 2 corrupt patients (595 and 599) with 3 slices. (hence the big mi)\nApart from the 2 corrupt patients there were some big errors from wrongly provided volumes (esp patient 429)</p>\n\n<p>I used only 1 model for everything.\nThe main thing that limited my performance was not knowing how to label and wrongly provided volumes. It tried to make up for this by doing some calibration.</p>\n\n<p>So I basically consider this problem solved.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 111904,
      "author_name": "halfwit",
      "author_url": "",
      "post_date": "03/17/2016 07:54:22",
      "content": "<blockquote>\n  <p>So I basically consider this problem solved.</p>\n</blockquote>\n\n<p>I agree, I doubt people doing medical research realize what is possible in terms of automation. I sure didn't. There are so many problems similar to this that consume 2-3 years of someones life. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 111995,
      "author_name": "spammy",
      "author_url": "",
      "post_date": "03/17/2016 18:43:09",
      "content": "<p>Problem solved? Probably not for clinical applications. It would be interesting to run an ejection fraction classifier using the same techniques you used in the competition. I'm guessing, given your numbers, that there would way too many false positives and, more critically, false negatives for this technique to be used in practice. But it would be an interesting exercise.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "111510": "Does anyone else think this is pretty impressive? When I considered that the manual labeling would be necessarily idiosyncratic, with a lot of judgement calls on the edge cases, I didn't think it possible on the test dataset. Congratulations to everyone near the top.",
    "111530": "Here are my results.\r\nIt's even better than 10%\r\n![enter image description here][1]\r\n\r\nValidate mi is biggest error too low and validate mx is biggest error to high.\r\nFold 4 contains 2 corrupt patients (595 and 599) with 3 slices. (hence the big mi)\r\nApart from the 2 corrupt patients there were some big errors from wrongly provided volumes (esp patient 429)\r\n\r\n  [1]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/3909/results.png\r\n\r\nI used only 1 model for everything.\r\nThe main thing that limited my performance was not knowing how to label and wrongly provided volumes. It tried to make up for this by doing some calibration.\r\n\r\nSo I basically consider this problem solved.",
    "111904": "> So I basically consider this problem solved.\r\n\r\nI agree, I doubt people doing medical research realize what is possible in terms of automation. I sure didn't. There are so many problems similar to this that consume 2-3 years of someones life.",
    "111995": "Problem solved? Probably not for clinical applications. It would be interesting to run an ejection fraction classifier using the same techniques you used in the competition. I'm guessing, given your numbers, that there would way too many false positives and, more critically, false negatives for this technique to be used in practice. But it would be an interesting exercise."
  },
  "source": "meta"
}