README.md

---
pretty_name: "Blood Donation Registry — Synthetic Donors, Prevalence & Compatibility"
license: cc-by-4.0
language:
  - en
size_categories:
  - 10K<n<100K
task_categories:
  - tabular-classification
  - tabular-regression
tags:
  - blood-donation
  - transfusion
  - healthcare-operations
  - decision-support
  - risk-scoring
  - calibration
  - thresholding
  - synthetic
  - tabular-data
annotations_creators:
  - no-annotation
source_datasets:
  - synthetic
configs:
  - config_name: ml_ready
    data_files:
      - split: train
        path: "data/blood_donation_registry_ml_ready.csv"
  - config_name: population_distribution
    data_files:
      - split: train
        path: "data/blood_population_distribution.csv"
  - config_name: compatibility_lookup
    data_files:
      - split: train
        path: "data/blood_compatibility_lookup.csv"
  - config_name: data_dictionary
    data_files:
      - split: train
        path: "data/data_dictionary.csv"
---

# 🩸 Blood Donation Registry — Synthetic Donors, Prevalence & Compatibility

Synthetic, decision-focused tables for blood donation operations: donor eligibility/deferrals, donation history, rare blood types, country-level prevalence, and RBC transfusion compatibility (ABO/Rh).

**Synthetic data** (safe for experimentation and teaching)  
**Not clinical/medical ground truth** — do not use for real-world medical decisions.

---

## 📦 What’s inside

This repo provides **four loadable dataset configs** via `datasets.load_dataset`:

- **`ml_ready`** *(default / recommended)* → donor-level ML-ready table
- **`population_distribution`** → country-level blood type prevalence
- **`compatibility_lookup`** → RBC transfusion compatibility matrix (8×8)
- **`data_dictionary`** → column-level documentation (all files)

---

## 🚀 Quick start (Hugging Face)

```python
from datasets import load_dataset

repo_id = "tarekmasryo/blood-donation-registry-dataset"

donors = load_dataset(repo_id, "ml_ready")["train"].to_pandas()
pop    = load_dataset(repo_id, "population_distribution")["train"].to_pandas()
compat = load_dataset(repo_id, "compatibility_lookup")["train"].to_pandas()

# Example join: add country population + Rh- rate to each donor
donors = donors.merge(
    pop[["country_code", "population_size", "rh_negative_rate"]],
    on="country_code",
    how="left",
)
print(donors.shape)
print(donors.head(3))
```

> Tip: `load_dataset(repo_id)` will load the first config (here: `ml_ready`).

---

## 🗂️ Data files

### 1) `data/blood_donation_registry_ml_ready.csv` (30,000 rows × 27 columns)

Donor-level snapshot with a fixed reference date:

- `as_of_date = 2024-12-31`

**Key groups**

- **Identity & geography:** `donor_id` (unique), `country_code`, `region`
- **Profile:** `age`, `sex` (M/F), `bmi`, `smoker` (0/1), `chronic_condition_flag` (0/1)
- **Eligibility & deferrals:** `eligibility_status`, `eligible_to_donate` (0/1), `deferral_reason`
- **Donation behavior/history:** `donation_count_last_12m`, `lifetime_donation_count`, `first_donation_year`,
  `years_since_first_donation`, `last_donation_date`, `recency_days`, `is_regular_donor` (0/1),
  `donor_age_at_first_donation`, `preferred_site`
- **Blood context:** `blood_type` (8 types), `is_rare_type` (0/1), `blood_type_country_prevalence`
- **Engineered score (optional):** `donation_propensity_score`

**Outcome columns**

- `donated_next_6m` (0/1)
- `next_6m_donation_count` (0–3)

---

### 2) `data/blood_population_distribution.csv` (39 rows × 12 columns)

Country-level population + blood type prevalence:

- keys: `country_code`, `region`
- `population_size`
- proportions: `p_o_pos`, `p_o_neg`, `p_a_pos`, `p_a_neg`, `p_b_pos`, `p_b_neg`, `p_ab_pos`, `p_ab_neg`
- `rh_negative_rate`

**Integrity expectation:** blood type proportions sum to **1.0** per country.

---

### 3) `data/blood_compatibility_lookup.csv` (64 rows × 4 columns)

RBC transfusion compatibility matrix:

- `donor_blood_type`, `recipient_blood_type`
- `compatible_for_rbc_transfusion` (0/1)
- `compatibility_level`: `ideal | acceptable | incompatible`

---

### 4) `data/data_dictionary.csv`

Column-level documentation for all files:

- `file`, `column_name`, `type`, `description`, `allowed_values_or_range`, `missing_values`

---

## 🔗 Relationships

- `blood_donation_registry_ml_ready.country_code`
  ↔ `blood_population_distribution.country_code`

- `blood_donation_registry_ml_ready.blood_type_country_prevalence`
  is derived from the matching country prevalence proportions.

- `blood_compatibility_lookup`
  provides rule-based compatibility for donor/recipient blood type pairs.

---

## 🧪 Recommended tasks

### 1) Donation likelihood (binary classification)
- Target: `donated_next_6m`
- Suggested evaluation: ROC-AUC, PR-AUC, F1 + **calibration** (reliability curve / Brier score)

### 2) Donation frequency (count prediction)
- Target: `next_6m_donation_count`
- Suggested evaluation: MAE, RMSE (optional Poisson/ordinal baselines)

### 3) Decision policy under constraints
Turn probabilities into an outreach policy:
- choose an operating threshold under **capacity/budget**
- compare FP/FN tradeoffs
- validate stability across segments (region, rare types, eligibility)

### 4) Rare-blood operations analytics
- analyze `is_rare_type` by country prevalence
- explore compatibility-aware matching using the lookup matrix

---

## 🧯 Modeling notes (avoid leakage / shortcuts)

- `donated_next_6m` is derived from `next_6m_donation_count` → use **one** outcome as the target.
- `eligible_to_donate` overlaps with `eligibility_status` → keep one for simpler baselines.
- `eligible_to_donate == 0` implies `donated_next_6m == 0` → for behavior modeling, consider training on `eligible_to_donate == 1`.
- `donation_propensity_score` is engineered; exclude it for “feature-only” benchmark baselines.

---

## ✅ Data quality expectations

- `donor_id` is unique (no duplicate donors)
- no duplicate rows
- snapshot consistency (`as_of_date` fixed)
- `recency_days` aligns with `as_of_date - last_donation_date`
- country codes match the prevalence table keys
- compatibility lookup covers all 8×8 donor/recipient pairs

---

## 🧬 Synthetic data generation (high-level)

Records are simulated to reflect realistic constraints and patterns:
- eligibility rules and deferral reasons (age/BMI/health flags)
- donation history distributions and recency behavior
- country-level blood type prevalence used to derive per-donor prevalence context
- ABO/Rh compatibility rules encoded in the lookup table

Synthetic distributions may not match any specific real-world population.

---

## ⚠️ Limitations

- snapshot-style dataset (not a full longitudinal event log beyond history fields)
- synthetic distributions may differ from real operational settings
- engineered signals (e.g., `donation_propensity_score`) can act as shortcut features if used without care

---

## 🧾 Citation

Tarek Masryo. (2025). *Blood Donation Registry — Synthetic Donors, Prevalence & Compatibility*.

---

## 📜 License

CC BY 4.0 — attribution required.

---

## 👤 Author

Tarek Masryo