Upload DPR dataset splits and detailed data card

.gitattributes

@@ -60,3 +60,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
 features.csv filter=lfs diff=lfs merge=lfs -text
 train.csv filter=lfs diff=lfs merge=lfs -text
+test.csv filter=lfs diff=lfs merge=lfs -text
+train_raw.csv filter=lfs diff=lfs merge=lfs -text
+validation.csv filter=lfs diff=lfs merge=lfs -text

README.md

@@ -2,48 +2,167 @@
 license: unknown
 task_categories:
 - tabular-classification
+task_ids:
+- binary-classification
 tags:
 - healthcare
-- tabular
+- diabetes
 - readmission
-pretty_name: Diabetic 30-Day Readmission Dataset
+- electronic-health-records
+- uncertainty-quantification
+- clinical-prediction
+pretty_name: Diabetic Patient 30-Day Readmission Dataset
+size_categories:
+- 100K<n<1M
 ---
 
-# Dataset Card for DPR
+# Diabetic Patient 30-Day Readmission Dataset
 
-This dataset is derived from the diabetic readmission dataset and is formatted for binary prediction of 30-day hospital readmission.
+## Dataset Summary
+
+This dataset is derived from the Kaggle Diabetic Patients Readmission Prediction dataset. The original dataset contains electronic health record data from diabetic patient hospital encounters and is commonly used for hospital readmission prediction.
+
+This released version formats the data for binary classification of 30-day hospital readmission and provides reproducible train, validation, and test splits for evaluating clinical prediction models and uncertainty quantification methods.
+
+## Source Data
+
+Original source:
+
+- Kaggle: https://www.kaggle.com/datasets/saurabhtayal/diabetic-patients-readmission-prediction
 
 ## Prediction Task
 
-The target label is `readmit_30d`.
+The prediction task is binary classification.
+
+Target variable:
+
+- `readmit_30d`
+
+Label definition:
+
+- `1`: patient was readmitted within 30 days, corresponding to `readmitted == "<30"`
+- `0`: patient was not readmitted within 30 days, corresponding to `readmitted == ">30"` or `readmitted == "NO"`
+
+The original multiclass `readmitted` variable was converted into this binary target.
+
+## Dataset Files
+
+This repository contains the following files:
+
+- `train.csv`: processed training split
+- `validation.csv`: processed validation split
+- `test.csv`: processed test split
+- `train_raw.csv`: raw training split before imputation and one-hot encoding
+- `validation_raw.csv`: raw validation split before imputation and one-hot encoding
+- `test_raw.csv`: raw test split before imputation and one-hot encoding
+- `cohort.csv`: encounter identifiers, patient identifiers, original readmission label, and binary label
+- `features.csv`: raw feature table excluding the original `readmitted` target column
+
+## Dataset Statistics
+
+| Split | N | Negatives | Positives | Positive Rate |
+|---|---:|---:|---:|---:|
+| Train | 71236 | 63286 | 7950 | 0.1116 |
+| Validation | 15265 | 13561 | 1704 | 0.1116 |
+| Test | 15265 | 13562 | 1703 | 0.1116 |
+| Total | 101766 | 90409 | 11357 | 0.1116 |
+
+Additional dataset information:
 
-- `1`: patient was readmitted within 30 days (`readmitted == "<30"`)
-- `0`: patient was not readmitted within 30 days
+- Number of total samples: 101766
+- Number of raw feature columns: 49
+- Number of processed feature columns after one-hot encoding: 2332
+- Feature types: mixed categorical and numerical clinical variables
+- Missing value marker in original dataset: `"?"`
 
-## Files
+## Feature Overview
 
-- `train.csv`: training split
-- `validation.csv`: validation split
-- `test.csv`: test split
-- `cohort.csv`: patient/encounter identifiers and labels
-- `features.csv`: feature columns excluding the target label
+The raw dataset includes clinical and administrative variables such as:
 
-## Splits
+- Demographic information
+- Admission and discharge information
+- Time in hospital
+- Laboratory test results
+- Medication indicators
+- Diagnosis code fields
+- Prior inpatient, outpatient, and emergency visit counts
 
-| Split | N | Positives | Positive Rate |
-|---|---:|---:|---:|
-| Train | 71236 | 7950 | 0.1116 |
-| Validation | 15265 | 1704 | 0.1116 |
-| Test | 15265 | 1703 | 0.1116 |
+Categorical features were one-hot encoded in the processed split files.
 
-## Preprocessing
+## Data Processing Pipeline
 
-Missing values originally encoded as `?` were replaced with missing values. The binary label `readmit_30d` was created from the original `readmitted` column. The data were split into 70/15/15 train/validation/test partitions using stratification on `readmit_30d`.
+The released dataset was generated using the following preprocessing procedure:
+
+1. Loaded `diabetic_data.csv`
+2. Replaced all `"?"` entries with missing values
+3. Created the binary label `readmit_30d`
+4. Created a cohort file containing `encounter_id`, `patient_nbr`, `readmitted`, and `readmit_30d`
+5. Created a raw feature file by removing the original `readmitted` target column
+6. Split the full dataset into train, validation, and test partitions using a 70/15/15 split
+7. Stratified both split steps by `readmit_30d`
+8. Removed identifier columns from modeling features:
+   - `encounter_id`
+   - `patient_nbr`
+9. Removed target columns from modeling features:
+   - `readmitted`
+   - `readmit_30d`
+10. Identified categorical and numerical columns using the training split only
+11. Imputed numerical features using the training-set median
+12. One-hot encoded categorical features using categories learned from the training split
+13. Aligned validation and test feature columns to the training feature columns
+14. Normalized column names to avoid special characters such as `<`, `>`, `[`, and `]`
+
+## Reproducibility
+
+The splits and preprocessing are reproducible using the following settings:
+
+- Train/temporary split seed: `42`
+- Validation/test split seed: `42`
+- Split ratio: 70% train, 15% validation, 15% test
+- Stratification variable: `readmit_30d`
+- Numerical imputation: median imputation fit on training split only
+- Categorical encoding: one-hot encoding fit on training split only
+- Unknown categorical levels in validation/test: ignored during encoding
 
 ## Intended Use
 
-This dataset is intended for evaluating binary classification and uncertainty quantification methods for 30-day readmission prediction.
+This dataset is intended for research on:
+
+- Binary clinical prediction
+- Hospital readmission prediction
+- Uncertainty quantification
+- Selective prediction
+- Risk-coverage evaluation
+- Decision-aware evaluation of machine learning models
+
+## Out-of-Scope Use
+
+This dataset should not be used for direct clinical decision-making, diagnosis, prognosis, or treatment allocation without external validation and appropriate clinical oversight.
 
 ## Limitations
 
-This dataset should not be used for clinical decision-making without further validation. Potential limitations include retrospective design, missingness, class imbalance, and dataset-specific biases.
+Important limitations include:
+
+- Retrospective observational data
+- Potential label noise in administrative readmission labels
+- Missingness in several clinical variables
+- Class imbalance in the 30-day readmission outcome
+- Potential dataset-specific biases
+- Limited generalizability to other hospitals, time periods, or patient populations
+
+## Ethical Considerations
+
+This dataset is derived from publicly available clinical data. Users should handle it responsibly and avoid using models trained on this dataset for clinical deployment without further validation.
+
+## Citation
+
+If you use this dataset, please cite the original Kaggle dataset:
+
+Diabetic Patients Readmission Prediction:
+https://www.kaggle.com/datasets/saurabhtayal/diabetic-patients-readmission-prediction
+
+You may also cite this dataset repository if using this processed version as part of an uncertainty quantification benchmark.
+
+## Additional Notes
+
+This dataset is part of a multi-dataset benchmark for evaluating uncertainty quantification methods in clinical prediction tasks.

train.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4af2d4be6198c5529e0676b4ce9450c42bc5e29efe6ff6c43956c146089ef151
-size 12839394
+oid sha256:a9c38594d3dea1ba0e16c3be2bf913ce32d95f1f2151ff05a64a01b72c8d8c3d
+size 664800886

train_raw.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4af2d4be6198c5529e0676b4ce9450c42bc5e29efe6ff6c43956c146089ef151
+size 12839394