IDT Routine Immunization Defaulter Prediction Model (v0.4.0)
Model Overview
The idt-defaulter-0.4.0 is a calibrated machine learning model engineered for the Immunization Defaulter Tracker (IDT) platform. It predicts the probability that an infant or young child will miss their next scheduled routine immunization dose ($Y=1$ defaulter within grace period).
Developed to support community health workers (CHWs) and primary healthcare facilities in resource-constrained environments (specifically modeled for Ethiopian routine immunization schedules: Pentavalent, Measles, Rota, PCV, OPV, and BCG).
Clinical Advisory & Regulatory Notice
Advisory Clinical Decision Support Only: This model was validated on research data. Predictions and TreeSHAP feature explanations are designed to assist healthcare workers in prioritizing proactive reminders and home outreach. They must never override clinical diagnosis, acute pediatric triage, or immediate vaccination contraindications.
Model Specifications
| Parameter | Specification |
|---|---|
| Model Algorithm | XGBoost Classifier (612 decision trees, max depth = 5, learning rate = 0.03) |
| Calibration Method | Platt Scaling (Sigmoid CalibratedClassifierCV) |
| Total Features | 62 Tabular Features (Demographic, Household, Prior Doses, Geographic Access, Workload) |
| Prediction Horizon | 7-day lead time ($T_j = S_j - 7$ days prior to scheduled dose date) |
| Primary Metric | ROC-AUC (0.84), Brier Score (< 0.12) |
| Target Population | Children 0–24 months undergoing routine EPI immunization |
Capacity-Constrained Quintile Thresholds
In resource-constrained settings, public health facilities face staffing caps. The model implements equity-stratified decision thresholds:
| Household Wealth Quintile | Operating Threshold | Action Priority |
|---|---|---|
| Global Default | 0.2122 |
Standard queue threshold |
| Quintile Q1 (Poorest / Highest Need) | 0.3249 |
Highest priority CHW home visit |
| Quintile Q2 | 0.2628 |
Proactive SMS + CHW follow-up |
| Quintile Q3 | 0.2185 |
Clinic reminder SMS |
| Quintile Q4 | 0.1690 |
Automated reminder |
| Quintile Q5 (Wealthiest) | 0.1140 |
Routine monitoring |
Strict Zero-PII Privacy Guarantee
To uphold patient confidentiality and global health data governance standards:
- No Direct Identifiers: The model feature schema strictly rejects names, phone numbers, national identification numbers, and exact GPS coordinates.
- Geographic Anonymization: Location is abstracted to administrative woreda identifiers and travel time/distance buckets.
Quickstart & Usage
1. Load with Python and Scikit-Learn
import joblib
import pandas as pd
from huggingface_hub import hf_hub_download
# Download model artifact
model_path = hf_hub_download(repo_id="YOUR_USERNAME/idt-defaulter-0.4.0", filename="model.joblib")
model = joblib.load(model_path)
# Predict probabilities on 62-feature matrix X
risk_probabilities = model.predict_proba(X)[:, 1]
print("Defaulter Risk Probabilities:", risk_probabilities[:5])
2. Verify Cryptographic Integrity
import hashlib
def verify_sha256(filepath, expected_hash):
h = hashlib.sha256()
with open(filepath, "rb") as f:
while chunk := f.read(8192):
h.update(chunk)
assert h.hexdigest() == expected_hash, "Hash mismatch!"
verify_sha256("model.joblib", "e2e6acac503459f48500326127a99eda2443aa09e9bed47218a1edd9ed4675d2")
Training Data & Cutoff
- Data Cutoff Date: 2026-10-27
- Trained At: 2026-11-09T08:33:50+00:00
- Antigens Evaluated: BCG, OPV, Penta (1-3), PCV (1-3), Rota (1-2), Measles (1-2).
- License: Apache 2.0