stunting-risk-scorer

S2.T1.2 · AIMS KTT Hackathon · Stunting Risk Heatmap Dashboard Author: Joseph Nyingi Wambua

---

Live Demo

▶ Interactive Dashboard

---

Model description

A logistic regression classifier (scikit-learn CalibratedClassifierCV with Platt scaling, 5-fold cross-validation) that scores each household's stunting risk on a continuous scale [0, 1].

Inputs (5 features):

Feature	Encoding
water_source	piped=0.0 → river_lake=1.0
sanitation_tier	improved=0.0 → open_defecation=1.0
income_band	high=0.0 → very_low=1.0
avg_meal_count	inverted: 5 meals=0.0 → 1 meal=1.0
children_under5	min(n/5, 1.0)

Output: risk_score in [0, 1]. Threshold ≥ 0.50 = high risk.

Performance (held-out 20% test set, n=60)

Metric	Value
AUC-ROC	0.935
Precision	0.839
Recall	0.867
F1	0.853

Feature importance (LR coefficients, standardised)

Feature	LR Coefficient
Sanitation access	+1.666
Water source quality	+1.536
Income level	+1.070
Meal frequency	+0.801
Children under 5	+0.516

Intended use

Monthly screening tool for Rwandan community health workers and Umudugudu chiefs. Outputs feed a printed A4 'sector page' listing the top-10 highest-risk households with anonymised IDs and one-line intervention hints.

Not intended for direct clinical diagnosis. Scores are risk indicators based on household socioeconomic factors, not measured height-for-age z-scores.

Training data

Synthetic NISR-style data generated with generate_data.py (seed 42).

• 2,500 households across 5 Rwandan districts (Nyarugenge, Gasabo, Kicukiro, Nyanza, Musanze)
• 300 gold-labelled households (150 positive, 150 negative)
• Features sampled conditional on district stunting baselines (NISR DHS 2019–20)

How to use

import joblib
import pandas as pd

# Load the model
model_data = joblib.load("scorer.pkl")
lr, scaler = model_data["lr"], model_data["scaler"]

# Featurize a household
WATER_RISK  = {"piped": 0.0, "protected_well": 0.33,
               "unprotected_well": 0.67, "river_lake": 1.0}
SANIT_RISK  = {"improved": 0.0, "basic": 0.33,
               "limited": 0.67, "open_defecation": 1.0}
INCOME_RISK = {"high": 0.0, "medium": 0.33, "low": 0.67, "very_low": 1.0}

def featurize(row):
    import numpy as np
    return np.array([
        WATER_RISK.get(row["water_source"], 0.5),
        SANIT_RISK.get(row["sanitation_tier"], 0.5),
        INCOME_RISK.get(row["income_band"], 0.5),
        1.0 - (float(row["avg_meal_count"]) - 1.0) / 4.0,
        min(int(row["children_under5"]) / 5.0, 1.0),
    ])

household = {
    "water_source": "unprotected_well",
    "sanitation_tier": "limited",
    "income_band": "low",
    "avg_meal_count": 2,
    "children_under5": 3,
}
feats = featurize(household).reshape(1, -1)
risk_score = lr.predict_proba(scaler.transform(feats))[0, 1]
print(f"Risk score: {risk_score:.3f}  →  {'HIGH RISK' if risk_score >= 0.5 else 'low risk'}")

Limitations

• Trained on synthetic data — not validated against real household surveys
• 300 gold labels is a small training set; real deployment requires DHS or CRVS ground-truth labels
• Threshold (0.50) should be re-calibrated for the specific operational context (recall vs. precision trade-off depends on health worker capacity)

Citation

@misc{nyingi2026stunting,
  author = {Nyingi, Joseph Wambua},
  title  = {S2.T1.2 Stunting Risk Heatmap Dashboard},
  year   = {2026},
  url    = {https://huggingface.co/Nyingi101/stunting-risk-scorer}
}