Boost ICU Stay Length Prediction
Budget: €8 – €30 EUR
I am competing in the Kaggle “ICU Length of Stay” challenge and need a fresh set of eyes to push my RMSE down from 4.16 to the 3.5–4.0 zone. The data come from MIMIC-III and include a mix of numeric vitals, and categorical chart events. Certain leakage-prone columns (DOD, DISCHTIME, DEATHTIME, HOSPITAL_EXPIRE_FLAG) must stay out of every fold and final model.
My first priority is model tuning. I have baseline LightGBM and XGBoost notebooks but they are plateauing, so I’m open to well-tuned tree ensembles, stacked regressors, KNN, SVM, or a lightweight neural network—whatever moves the needle. While you are free to add clever feature work, so everything plays nicely with the chosen model. Imputing the mixed missingness in a principled way is essential, but feel free to lean on tried-and-true methods unless something more sophisticated clearly pays off.
Deliverable
• A single, clean Jupyter Notebook (.ipynb) that:
– loads the public MIMIC-III competition files, performs the one-hot encoding and any additional transformations you deem useful,
– handles missing values properly
– walks through your tuning strategy step by step (cross-validation, early stopping, hyper-parameter search, ensembling logic, etc.),
– reports reproducible CV results plus a Kaggle test submission file in CSV format,
– includes concise markdown cells explaining the reasoning so I can justify the approach in a write-up.
I value transparency, so please comment source code clearly and cite any external libraries or snippets you borrow. Experience with regression tasks or clinical/health-care datasets will stand out—let me know what you’ve tackled before and how it went.
Looking forward to seeing how far you can push the score.
My first priority is model tuning. I have baseline LightGBM and XGBoost notebooks but they are plateauing, so I’m open to well-tuned tree ensembles, stacked regressors, KNN, SVM, or a lightweight neural network—whatever moves the needle. While you are free to add clever feature work, so everything plays nicely with the chosen model. Imputing the mixed missingness in a principled way is essential, but feel free to lean on tried-and-true methods unless something more sophisticated clearly pays off.
Deliverable
• A single, clean Jupyter Notebook (.ipynb) that:
– loads the public MIMIC-III competition files, performs the one-hot encoding and any additional transformations you deem useful,
– handles missing values properly
– walks through your tuning strategy step by step (cross-validation, early stopping, hyper-parameter search, ensembling logic, etc.),
– reports reproducible CV results plus a Kaggle test submission file in CSV format,
– includes concise markdown cells explaining the reasoning so I can justify the approach in a write-up.
I value transparency, so please comment source code clearly and cite any external libraries or snippets you borrow. Experience with regression tasks or clinical/health-care datasets will stand out—let me know what you’ve tackled before and how it went.
Looking forward to seeing how far you can push the score.