Portfolio showcase — AI-assisted maternal health risk screening with explainable machine learning.
MaternaAI is a clinical decision support prototype that screens maternal health risk from six physiological measurements and presents the result through a calm, transparent, bilingual interface.
Public portfolio repository: this repository intentionally contains documentation and project evidence only. Application source code, training notebooks, trained model artifacts, internal evaluation files, dataset copies, and the local assessment database are kept private.
Full project demo:
https://drive.google.com/file/d/1iV1us3YW6lIt7jeUDtGoG1jYIpvWgURh/view?usp=drivesdk
Before sharing this repository with recruiters, make sure the Google Drive video is set to Anyone with the link — Viewer. The current connector could not verify public access.
- End-to-end machine learning workflow for a 3-class maternal risk screening problem
- Two-stage XGBoost decision architecture
- Probability calibration and explicit separation between screening decision and model probability distribution
- Explainable AI using SHAP-based local explanations
- Flask-based web application with a responsive clinical-style interface
- Indonesian / English presentation layer
- Anonymous assessment flow with optional name or alias
- SQLite-backed screening history and analytics
- PDF export for saved screening results
- Model transparency page with performance and limitation disclosures
flowchart LR
A[6 physiological measurements] --> B[Input validation]
B --> C[Two-Stage XGBoost]
C --> D[Screening decision]
C --> E[Calibrated probabilities]
C --> F[SHAP explanation]
D --> G[Flask interface]
E --> G
F --> G
G --> H[(SQLite history)]
H --> I[PDF report]
The current V3 candidate is evaluated as an internal screening prototype, not a diagnostic system.
| Metric | Internal test result |
|---|---|
| Accuracy | 70.94% |
| Macro F1 | 72.01% |
| Mid Risk Recall | 83.58% |
| High Risk Recall | 85.45% |
| High Risk predicted directly as Low Risk | 0 cases |
These metrics are shown with their limitations rather than as claims of clinical readiness. The project emphasizes transparent error analysis, leakage-aware evaluation, calibration, and explainability.
Python · Flask · XGBoost · scikit-learn · SHAP · pandas · SQLite · HTML/CSS/JavaScript
The application is designed as a clinical intelligence workspace, not a generic ML dashboard.
- Overview — project context and screening entry point
- New Screening — six physiological measurements
- Result — screening decision and probability distribution shown separately
- Explainable AI — factors influencing the model output
- History — searchable stored assessments
- Analytics — summaries based on saved screening records
- Model Information — metrics, architecture, calibration, and limitations
- PDF Export — report generated from a saved assessment
- Bilingual UI — Indonesian and English presentation layer
MaternaAI is a personal portfolio project. This public repository is deliberately limited to project evidence and technical documentation so recruiters can evaluate the engineering decisions, UX, model behavior, and project scope without publishing the implementation itself.
The following are not included:
- application source code
- model-training notebooks
- trained model binaries
- raw or processed datasets
- threshold-search and internal evaluation artifacts
- SQLite assessment database
- internal implementation documentation
See PORTFOLIO_NOTICE.md for the repository-use notice.
MaternaAI is an educational and portfolio clinical decision support prototype. It does not provide a medical diagnosis and is not intended to replace professional clinical assessment.
Status: Portfolio showcase · V3 prototype · 2026