{
  "id": 678327,
  "title": "Welcome to AI-EMT | Competition Guide, Resources & Q&A",
  "url": "/competitions/ai-emergency-medicine-turkiye/discussion/678327",
  "author_name": "Olaf Yunus Laitinen Imanov",
  "post_date": "2026-02-24T05:55:40.132000",
  "votes": 5,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Welcome to the AI in Emergency Medicine - Turkiye (AI-EMT) Competition!</p>\n<p>We are excited to have you here. This thread is the official hub for \nannouncements, questions, and community discussion throughout the competition.\nPlease read this post carefully before getting started.</p>\n<hr>\n<h2>What You Are Predicting</h2>\n<p>This competition asks you to predict two binary clinical outcomes for \nemergency department encounters across Turkiye:</p>\n<ul>\n<li>adverse_outcome: ICU admission OR in-hospital death (prevalence ~41%)</li>\n<li>readmission_30d: unplanned ED return within 30 days (prevalence ~14%)</li>\n</ul>\n<p>Evaluation metric: Macro-Averaged AUC-ROC across both targets.\nScore = (AUC_adverse_outcome + AUC_readmission_30d) / 2</p>\n<hr>\n<h2>Getting Started</h2>\n<ol>\n<li>Download the data from the Data tab.</li>\n<li>Open the official starter notebook: setup_solution.ipynb</li>\n<li>The notebook walks you through EDA, feature engineering, a Logistic \nRegression baseline, and a LightGBM model with 5-fold cross-validation.</li>\n<li>Format your submission using sample_submission.csv and submit.</li>\n</ol>\n<p>Benchmark score to beat: 0.6955 macro AUC (Logistic Regression, 45 features, \nno tuning). A well-engineered LightGBM baseline achieves approximately 0.8507.</p>\n<hr>\n<h2>Key Things to Know</h2>\n<p>Missing values in troponin_i_ngml, bnp_pgl, d_dimer_mgfl, lactate_mmoll, \nlipase_ul, procalcitonin_ngl, and albumin_gdl are Missing Not At Random (MNAR). \nThese tests are selectively ordered for sicker patients. The missingness \nindicator itself is predictive. We strongly recommend creating binary \n{column}_missing features before imputation.</p>\n<p>The disposition column in train.csv is a direct proxy for both targets \n(it encodes ICU admission and death explicitly). Using it as a predictor \nconstitutes data leakage and will result in disqualification.</p>\n<p>supplemental_data.csv contains 100,000 additional labeled encounters. \nYou may use this file for any purpose including additional training, \npseudo-labeling, or calibration.</p>\n<hr>\n<h2>File Summary</h2>\n<table>\n<thead>\n<tr>\n<th>File</th>\n<th>Rows</th>\n<th>Purpose</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>train.csv</td>\n<td>700,000</td>\n<td>Training data with targets</td>\n</tr>\n<tr>\n<td>test.csv</td>\n<td>200,000</td>\n<td>Test data, predict these</td>\n</tr>\n<tr>\n<td>sample_submission.csv</td>\n<td>200,000</td>\n<td>Submission format template</td>\n</tr>\n<tr>\n<td>supplemental_data.csv</td>\n<td>100,000</td>\n<td>Extra labeled data, free to use</td>\n</tr>\n<tr>\n<td>hospital_reference.csv</td>\n<td>25</td>\n<td>City-level hospital metadata</td>\n</tr>\n<tr>\n<td>icd10_reference.csv</td>\n<td>21</td>\n<td>ICD-10 code reference table</td>\n</tr>\n</tbody>\n</table>\n<hr>\n<h2>Submission Limits</h2>\n<ul>\n<li>5 submissions per day</li>\n<li>Select 2 final submissions for private leaderboard scoring</li>\n<li>Maximum team size: 4 participants</li>\n<li>Team merger deadline: 5 days before competition close</li>\n</ul>\n<hr>\n<h2>Competition Organizers</h2>\n<p>This competition is organized by the Informatiqs community.</p>\n<p>Kutluk Atalay - Community President\nOzan Mohurcu - Community Vice President\nOlaf Yunus Laitinen Imanov - Lead Competition Organizer</p>\n<p>For questions, tag <a href=\"https://www.kaggle.com/OlafYunusLaitinenImanov\" target=\"_blank\">@OlafYunusLaitinenImanov</a> in the discussion thread.\nWe aim to respond within 48 hours.</p>\n<hr>\n<h2>Discussion Guidelines</h2>\n<ul>\n<li>Be respectful and constructive.</li>\n<li>Share domain knowledge, research references, and general approaches freely.</li>\n<li>Do not share complete working code pipelines with competing teams before \nthe competition closes.</li>\n<li>For bug reports or data questions, describe the issue clearly and include \na minimal reproducible example where possible.</li>\n</ul>\n<hr>\n<h2>Good luck to everyone.</h2>\n<p>We look forward to seeing what the Informatiqs community builds.\nThe organizers will post updates and clarifications in this thread \nas the competition progresses.</p>",
  "messages": [
    {
      "id": 3413042,
      "postDate": "2026-02-24T05:55:40.133Z",
      "content": "<p>Welcome to the AI in Emergency Medicine - Turkiye (AI-EMT) Competition!</p>\n<p>We are excited to have you here. This thread is the official hub for \nannouncements, questions, and community discussion throughout the competition.\nPlease read this post carefully before getting started.</p>\n<hr>\n<h2>What You Are Predicting</h2>\n<p>This competition asks you to predict two binary clinical outcomes for \nemergency department encounters across Turkiye:</p>\n<ul>\n<li>adverse_outcome: ICU admission OR in-hospital death (prevalence ~41%)</li>\n<li>readmission_30d: unplanned ED return within 30 days (prevalence ~14%)</li>\n</ul>\n<p>Evaluation metric: Macro-Averaged AUC-ROC across both targets.\nScore = (AUC_adverse_outcome + AUC_readmission_30d) / 2</p>\n<hr>\n<h2>Getting Started</h2>\n<ol>\n<li>Download the data from the Data tab.</li>\n<li>Open the official starter notebook: setup_solution.ipynb</li>\n<li>The notebook walks you through EDA, feature engineering, a Logistic \nRegression baseline, and a LightGBM model with 5-fold cross-validation.</li>\n<li>Format your submission using sample_submission.csv and submit.</li>\n</ol>\n<p>Benchmark score to beat: 0.6955 macro AUC (Logistic Regression, 45 features, \nno tuning). A well-engineered LightGBM baseline achieves approximately 0.8507.</p>\n<hr>\n<h2>Key Things to Know</h2>\n<p>Missing values in troponin_i_ngml, bnp_pgl, d_dimer_mgfl, lactate_mmoll, \nlipase_ul, procalcitonin_ngl, and albumin_gdl are Missing Not At Random (MNAR). \nThese tests are selectively ordered for sicker patients. The missingness \nindicator itself is predictive. We strongly recommend creating binary \n{column}_missing features before imputation.</p>\n<p>The disposition column in train.csv is a direct proxy for both targets \n(it encodes ICU admission and death explicitly). Using it as a predictor \nconstitutes data leakage and will result in disqualification.</p>\n<p>supplemental_data.csv contains 100,000 additional labeled encounters. \nYou may use this file for any purpose including additional training, \npseudo-labeling, or calibration.</p>\n<hr>\n<h2>File Summary</h2>\n<table>\n<thead>\n<tr>\n<th>File</th>\n<th>Rows</th>\n<th>Purpose</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>train.csv</td>\n<td>700,000</td>\n<td>Training data with targets</td>\n</tr>\n<tr>\n<td>test.csv</td>\n<td>200,000</td>\n<td>Test data, predict these</td>\n</tr>\n<tr>\n<td>sample_submission.csv</td>\n<td>200,000</td>\n<td>Submission format template</td>\n</tr>\n<tr>\n<td>supplemental_data.csv</td>\n<td>100,000</td>\n<td>Extra labeled data, free to use</td>\n</tr>\n<tr>\n<td>hospital_reference.csv</td>\n<td>25</td>\n<td>City-level hospital metadata</td>\n</tr>\n<tr>\n<td>icd10_reference.csv</td>\n<td>21</td>\n<td>ICD-10 code reference table</td>\n</tr>\n</tbody>\n</table>\n<hr>\n<h2>Submission Limits</h2>\n<ul>\n<li>5 submissions per day</li>\n<li>Select 2 final submissions for private leaderboard scoring</li>\n<li>Maximum team size: 4 participants</li>\n<li>Team merger deadline: 5 days before competition close</li>\n</ul>\n<hr>\n<h2>Competition Organizers</h2>\n<p>This competition is organized by the Informatiqs community.</p>\n<p>Kutluk Atalay - Community President\nOzan Mohurcu - Community Vice President\nOlaf Yunus Laitinen Imanov - Lead Competition Organizer</p>\n<p>For questions, tag <a href=\"https://www.kaggle.com/OlafYunusLaitinenImanov\" target=\"_blank\">@OlafYunusLaitinenImanov</a> in the discussion thread.\nWe aim to respond within 48 hours.</p>\n<hr>\n<h2>Discussion Guidelines</h2>\n<ul>\n<li>Be respectful and constructive.</li>\n<li>Share domain knowledge, research references, and general approaches freely.</li>\n<li>Do not share complete working code pipelines with competing teams before \nthe competition closes.</li>\n<li>For bug reports or data questions, describe the issue clearly and include \na minimal reproducible example where possible.</li>\n</ul>\n<hr>\n<h2>Good luck to everyone.</h2>\n<p>We look forward to seeing what the Informatiqs community builds.\nThe organizers will post updates and clarifications in this thread \nas the competition progresses.</p>",
      "rawMarkdown": "Welcome to the AI in Emergency Medicine - Turkiye (AI-EMT) Competition!\n\nWe are excited to have you here. This thread is the official hub for \nannouncements, questions, and community discussion throughout the competition.\nPlease read this post carefully before getting started.\n\n---\n\n## What You Are Predicting\n\nThis competition asks you to predict two binary clinical outcomes for \nemergency department encounters across Turkiye:\n\n- adverse_outcome: ICU admission OR in-hospital death (prevalence ~41%)\n- readmission_30d: unplanned ED return within 30 days (prevalence ~14%)\n\nEvaluation metric: Macro-Averaged AUC-ROC across both targets.\nScore = (AUC_adverse_outcome + AUC_readmission_30d) / 2\n\n---\n\n## Getting Started\n\n1. Download the data from the Data tab.\n2. Open the official starter notebook: setup_solution.ipynb\n3. The notebook walks you through EDA, feature engineering, a Logistic \n   Regression baseline, and a LightGBM model with 5-fold cross-validation.\n4. Format your submission using sample_submission.csv and submit.\n\nBenchmark score to beat: 0.6955 macro AUC (Logistic Regression, 45 features, \nno tuning). A well-engineered LightGBM baseline achieves approximately 0.8507.\n\n---\n\n## Key Things to Know\n\nMissing values in troponin_i_ngml, bnp_pgl, d_dimer_mgfl, lactate_mmoll, \nlipase_ul, procalcitonin_ngl, and albumin_gdl are Missing Not At Random (MNAR). \nThese tests are selectively ordered for sicker patients. The missingness \nindicator itself is predictive. We strongly recommend creating binary \n{column}_missing features before imputation.\n\nThe disposition column in train.csv is a direct proxy for both targets \n(it encodes ICU admission and death explicitly). Using it as a predictor \nconstitutes data leakage and will result in disqualification.\n\nsupplemental_data.csv contains 100,000 additional labeled encounters. \nYou may use this file for any purpose including additional training, \npseudo-labeling, or calibration.\n\n---\n\n## File Summary\n\n| File                  | Rows    | Purpose                          |\n|-----------------------|---------|----------------------------------|\n| train.csv             | 700,000 | Training data with targets       |\n| test.csv              | 200,000 | Test data, predict these         |\n| sample_submission.csv | 200,000 | Submission format template       |\n| supplemental_data.csv | 100,000 | Extra labeled data, free to use  |\n| hospital_reference.csv| 25      | City-level hospital metadata     |\n| icd10_reference.csv   | 21      | ICD-10 code reference table      |\n\n---\n\n## Submission Limits\n\n- 5 submissions per day\n- Select 2 final submissions for private leaderboard scoring\n- Maximum team size: 4 participants\n- Team merger deadline: 5 days before competition close\n\n---\n\n## Competition Organizers\n\nThis competition is organized by the Informatiqs community.\n\nKutluk Atalay - Community President\nOzan Mohurcu - Community Vice President\nOlaf Yunus Laitinen Imanov - Lead Competition Organizer\n\nFor questions, tag @OlafYunusLaitinenImanov in the discussion thread.\nWe aim to respond within 48 hours.\n\n---\n\n## Discussion Guidelines\n\n- Be respectful and constructive.\n- Share domain knowledge, research references, and general approaches freely.\n- Do not share complete working code pipelines with competing teams before \n  the competition closes.\n- For bug reports or data questions, describe the issue clearly and include \n  a minimal reproducible example where possible.\n\n---\n\n## Good luck to everyone.\n\nWe look forward to seeing what the Informatiqs community builds.\nThe organizers will post updates and clarifications in this thread \nas the competition progresses.",
      "votes": 5
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3413042": "Welcome to the AI in Emergency Medicine - Turkiye (AI-EMT) Competition!\n\nWe are excited to have you here. This thread is the official hub for \nannouncements, questions, and community discussion throughout the competition.\nPlease read this post carefully before getting started.\n\n---\n\n## What You Are Predicting\n\nThis competition asks you to predict two binary clinical outcomes for \nemergency department encounters across Turkiye:\n\n- adverse_outcome: ICU admission OR in-hospital death (prevalence ~41%)\n- readmission_30d: unplanned ED return within 30 days (prevalence ~14%)\n\nEvaluation metric: Macro-Averaged AUC-ROC across both targets.\nScore = (AUC_adverse_outcome + AUC_readmission_30d) / 2\n\n---\n\n## Getting Started\n\n1. Download the data from the Data tab.\n2. Open the official starter notebook: setup_solution.ipynb\n3. The notebook walks you through EDA, feature engineering, a Logistic \n   Regression baseline, and a LightGBM model with 5-fold cross-validation.\n4. Format your submission using sample_submission.csv and submit.\n\nBenchmark score to beat: 0.6955 macro AUC (Logistic Regression, 45 features, \nno tuning). A well-engineered LightGBM baseline achieves approximately 0.8507.\n\n---\n\n## Key Things to Know\n\nMissing values in troponin_i_ngml, bnp_pgl, d_dimer_mgfl, lactate_mmoll, \nlipase_ul, procalcitonin_ngl, and albumin_gdl are Missing Not At Random (MNAR). \nThese tests are selectively ordered for sicker patients. The missingness \nindicator itself is predictive. We strongly recommend creating binary \n{column}_missing features before imputation.\n\nThe disposition column in train.csv is a direct proxy for both targets \n(it encodes ICU admission and death explicitly). Using it as a predictor \nconstitutes data leakage and will result in disqualification.\n\nsupplemental_data.csv contains 100,000 additional labeled encounters. \nYou may use this file for any purpose including additional training, \npseudo-labeling, or calibration.\n\n---\n\n## File Summary\n\n| File                  | Rows    | Purpose                          |\n|-----------------------|---------|----------------------------------|\n| train.csv             | 700,000 | Training data with targets       |\n| test.csv              | 200,000 | Test data, predict these         |\n| sample_submission.csv | 200,000 | Submission format template       |\n| supplemental_data.csv | 100,000 | Extra labeled data, free to use  |\n| hospital_reference.csv| 25      | City-level hospital metadata     |\n| icd10_reference.csv   | 21      | ICD-10 code reference table      |\n\n---\n\n## Submission Limits\n\n- 5 submissions per day\n- Select 2 final submissions for private leaderboard scoring\n- Maximum team size: 4 participants\n- Team merger deadline: 5 days before competition close\n\n---\n\n## Competition Organizers\n\nThis competition is organized by the Informatiqs community.\n\nKutluk Atalay - Community President\nOzan Mohurcu - Community Vice President\nOlaf Yunus Laitinen Imanov - Lead Competition Organizer\n\nFor questions, tag @OlafYunusLaitinenImanov in the discussion thread.\nWe aim to respond within 48 hours.\n\n---\n\n## Discussion Guidelines\n\n- Be respectful and constructive.\n- Share domain knowledge, research references, and general approaches freely.\n- Do not share complete working code pipelines with competing teams before \n  the competition closes.\n- For bug reports or data questions, describe the issue clearly and include \n  a minimal reproducible example where possible.\n\n---\n\n## Good luck to everyone.\n\nWe look forward to seeing what the Informatiqs community builds.\nThe organizers will post updates and clarifications in this thread \nas the competition progresses."
  }
}