DATA ANALYTICS & MACHINE LEARNING

Give Life — Predict Blood Donations

An end-to-end data analytics and machine learning project using the UCI Blood Transfusion Service Center dataset. Investigates historical donor patterns, examines data quality, and develops a balanced Logistic Regression model to estimate historical blood donation outcomes.

Project Overview

Give Life — Predict Blood Donations investigates historical blood donation patterns and develops a classification model to estimate whether a donor made a blood donation in March 2007.

The project uses the Blood Transfusion Service Center dataset from the UCI Machine Learning Repository.

It combines exploratory data analysis, data quality investigation, feature analysis, model development, evaluation, and business interpretation.

The objective is to demonstrate how historical data can support evidence-based decisions about donor outreach prioritization.

Tools & Technologies

Python

Pandas

Scikit-learn

Matplotlib

Jupyter Notebook

Dataset

UCI Blood Transfusion Service Center

Target

Donation in March 2007

Business Problem

Blood donation services need to maintain reliable blood supplies while encouraging regular donations.

The business question:

Can historical donor information help estimate which donors made a donation in March 2007, and how might these estimates support donor outreach prioritization?

This is decision-support analysis, not a replacement for donor screening, medical eligibility procedures, or human judgment.

Dataset

The analysis uses the Blood Transfusion Service Center dataset from the UCI Machine Learning Repository.

Dataset summary
Measure	Value
Observations	748
Columns	5
Non-donors	570 (76.2%)
Donors	178 (23.8%)

Features:

Recency: Months since the most recent donation.
Frequency: Total recorded donations.
Monetary: Total blood volume donated in cubic centimeters.
Time: Months since the first donation.
Target: Whether the donor donated in March 2007.

Source: UCI Machine Learning Repository — Blood Transfusion Service Center

Data Quality Investigation

Data quality was investigated before modelling because the dataset contains repeated records and no donor identifier.

Exact-duplicate and conflict audit
Measure	Result
Exact duplicate rows identified using pandas duplicated()	215
Rows involved in exact-duplicate groups	275
Unique exact rows	533
Rows involved in duplicate predictor profiles	315
Conflicting feature groups	31
Rows involved in conflicting groups	200

The 215 exact duplicate rows identified using pandas duplicated() exclude the first occurrence in each duplicate group. This is why 275 rows are involved in exact-duplicate groups while 533 unique exact rows remain.

The dataset contains no donor identifier, so it cannot establish whether identical predictor profiles represent repeated records of the same donor or different donors. Duplicate records were retained rather than automatically removed.

Monetary = 250 × Frequency. Monetary is therefore mathematically determined by Frequency and is not an independent information source.

Exploratory Analysis

The approximate target correlations were:

Approximate target correlations
Feature	Correlation
Recency	-0.280
Frequency	+0.219
Monetary	+0.219
Time	-0.036

These are descriptive associations, not causal relationships.

Monetary and Frequency should not be interpreted as independent evidence because Monetary is mathematically determined by Frequency.

Modelling Approach

The project used binary classification to estimate whether a donor made a donation in March 2007.

Stratified Split

A stratified train-test split produced 561 training observations and 187 test observations, using random_state=42.

Cross-Validation

Model evaluation used 5-fold StratifiedKFold cross-validation with shuffling enabled and random_state=123.

Training-Only Selection

Training-only cross-validation and out-of-fold predictions were used for model selection and threshold selection.

Binary Classification

The target represents whether a donor made a blood donation in March 2007.

Final Model

The final model was Balanced Logistic Regression — Configuration E.

Configuration E
Component	Configuration
Algorithm	Logistic Regression
Solver	liblinear
Class weighting	balanced
Random state	42
Features	Recency, Frequency, log(Monetary), Time
Monetary transformation	Logarithmic
StandardScaler	Not used
Selected threshold	0.35

Configuration E achieved cross-validation Average Precision of approximately 0.5401, compared with approximately 0.5376 for Configuration D with balanced class weighting.

The difference was marginal and does not demonstrate a large or definitive improvement.

Model Evaluation

The following results are from the held-out test set:

0.7902 ROC-AUC

0.5674 Average Precision

0.5294 Accuracy

0.3167 Precision

0.8636 Recall

0.4634 F1-score

Confusion matrix
	Predicted: No Donation	Predicted: Donation
Actual: No Donation	61	82
Actual: Donation	6	38

The model identified 38 of 44 actual positive cases, missed 6, and generated 82 false-positive predictions.

The high recall means the model captured most actual donors in the test set, but the lower precision means many predicted donations were false positives. The model is not highly accurate overall and does not guarantee future donors.

Business Insights and Recommendations

Recency

Recency was the strongest simple feature association.

Historical Frequency

Historical Frequency showed a positive association.

False-Positive Burden

High recall comes with a substantial false-positive follow-up burden.

Data Quality

Data quality and missing donor identifiers limit interpretation.

Proposed Applications

These are proposed applications, not demonstrated operational outcomes.

Predictions could support exploratory outreach prioritization.
Consider available outreach capacity and false-positive costs.
Conduct a controlled real-world pilot before operational use.
Improve donor identifiers and data collection.
Maintain human oversight and all established medical screening procedures.

Limitations and Transparency

The target is a historical March 2007 outcome, not a guaranteed future donation.
The dataset is relatively small (748 observations).
The target is imbalanced.
Duplicate predictor profiles with conflicting targets exist.
Donor identifiers are unavailable.
Precision is relatively low despite high recall.
The test set was used in earlier exploratory work and was not historically untouched throughout the entire project. Later model and threshold selection used training-only cross-validation and out-of-fold evaluation.
No real-world pilot has demonstrated increased donations, reduced costs, or improved operations.
The model does not assess medical eligibility or health status.

Conclusion

The project demonstrates an end-to-end analytical process combining data quality investigation, exploratory analysis, machine learning, model evaluation, interpretation, and business-oriented recommendations.

The final balanced Logistic Regression model achieved ROC-AUC 0.7902, Average Precision 0.5674, and recall 86.36% on the held-out test set at threshold 0.35.

Model results must be interpreted alongside data quality, operational capacity, and evidence limitations.

← Back to Projects