Final-year research project

A transparent proof of concept for maternal-risk triage.

The project combines leakage-aware evaluation, diverse machine-learning models, and explainability to classify three maternal-risk levels from routine physiological measurements.

Computer Science · FUOYE

Research title

Explainable AI for Maternal and Child Health Outcomes

ResearcherAdesanmi Anuoluwapo Gideon

Matric numberCSC/2022/81014

DepartmentComputer Science

SupervisorDr John Oloruntoba Awoyemi

InstitutionFederal University Oye-Ekiti

Study focusExplainable maternal-risk classification

“To develop an XAI system for classifying maternal health risk levels and explaining the contribution of maternal health measurements.”

Research aim

Scope boundary

The available dataset contains maternal physiology and risk labels only. It cannot support empirical claims about child or neonatal outcomes.

Research objectives

Three questions guide the implementation.

Design

Design a maternal health risk classification system using explainable artificial-intelligence techniques.

Implement

Build the system with appropriate machine-learning models and explanations for low, mid, and high-risk classification.

Evaluate

Measure performance using accuracy, precision, recall, F1, and class-sensitive diagnostic metrics.

Contribution to knowledge

The work connects model performance with the conditions under which that performance was measured.

Integrates leak detection, preprocessing, model comparison, ensemble selection, and interpretation in one auditable pipeline.
Demonstrates how duplicate profiles can inflate apparent accuracy in a small clinical dataset.
Pairs aggregate scores with class-wise confusion analysis, calibration, permutation importance, and SHAP.
Makes the weakness of mid-risk classification visible instead of allowing overall accuracy to conceal it.
States the maternal-only empirical scope despite the broader wording of the original project title.

What the prototype demonstrates

A compact screening model can be built from measurements available during routine antenatal care.

Six low-cost physiological inputs can drive three-class risk exploration.
Deterministic feature engineering captures pressure, glucose, temperature, and age interactions.
Algorithmic diversity can improve accuracy and ordinal agreement over individual candidates.
Global explanations identify systolic pressure, blood glucose, and temperature as influential signals.
Browser deployment can preserve privacy and eliminate paid inference infrastructure.

Study limitations

The deduplicated dataset contains only 452 observations.
Thirty-five predictor profiles carry conflicting target labels.
The holdout remains internal rather than an independent external cohort.
No patient, institution, or collection-time identifiers were available for stronger grouping.
Mid-risk recall is low at 15.8% for the selected ensemble.
Feature-importance methods describe prediction, not causation.
The dataset has no child, neonatal, mortality, or complication outcomes.

Recommendations

Collect a larger, more heterogeneous cohort from multiple care settings.
Validate the fitted workflow on a fully external group-isolated dataset.
Improve coverage and clinical justification for the mid-risk category.
Choose thresholds and referral rules on development data before external testing.
Ask obstetric experts to assess whether model explanations are clinically plausible.
Add child-specific measurements and outcomes before extending the system’s claims.

Use principles

The model supports inquiry; it does not replace care.

Private by design

The deployed prototype runs on the visitor’s device; entered measurements are not transmitted or stored remotely.

Evidence labelled

The robust holdout score and optimistic legacy benchmark are presented separately with their evaluation conditions.

Clinical caution

Outputs are research classifications only and require qualified clinical interpretation, especially for mid-risk patterns.

Research-use notice

This web app showcases a final-year computer-science project. It is not a medical device, has not undergone prospective clinical validation, and must not be used to diagnose, treat, or delay care.