Devpost project

Inspiration -The silent epidemic of hidden hunger affects over 2 billion people globally - they consume enough calories but lack essential micronutrients, leading to irreversible health impacts. What struck us most was that this condition remains invisible until serious complications arise. When we discovered that 60.8% of our dataset showed hidden hunger risk, we realized the urgent need for early detection systems. The fact that traditional screening methods require expensive blood tests and laboratory analysis inspired us to create an accessible, data-driven solution using basic demographic and dietary information that any community health worker could collect.

What it does-Our Hidden Hunger Risk Prediction System identifies individuals at risk of micronutrient deficiency with 89.2% accuracy using just 9 simple inputs (age, gender, income, education, and basic nutrient intake estimates).

The system:

• Predicts risk with an F1 score of 0.7244 and 98% recall, ensuring minimal missed cases
• Identifies critical deficiencies - revealing that 82% lack Vitamin D, 79% lack Folate
• Segments populations for targeted interventions (e.g., 85% risk in low-income + primary education groups)
• Provides real-time monitoring through an interactive dashboard -The model transforms invisible malnutrition into visible, preventable risk scores that enable early intervention before health impacts occur.

How we built it

-1. Data Engineering -2. Model Development:We tested 6 different algorithms (Logistic Regression, Random Forest, XGBoost, Gradient Boosting, Extra Trees, SVM). Extra Trees emerged as the champion with optimal threshold tuning (0.391 vs default 0.5). -3. Validation Strategy: Used 80/20 stratified split and 5-fold cross-validation to ensure robustness. The model maintains consistent performance (F1: 0.7180 ± 0.0234 across folds).

Challenges we ran into

-Class Imbalance -Interpretability vs Performance -Threshold Optimization

Accomplishments that we're proud of

• 98% Recall Rate: Our model catches nearly all at-risk individuals, crucial for a health screening application
• Actionable Insights: Identified that Vitamin D deficiency affects 82% of at-risk individuals - a clear intervention priority
• Feature Engineering: Our 35 engineered features capture complex relationships that improved F1 score from 0.62 (baseline) to 0.72

What we learned

• Hidden Patterns: Socio-economic factors (education, income) are stronger predictors than individual nutrient levels
• Age Vulnerability: Risk follows a U-shaped curve - highest in children (0-5) and working adults (19-60)
• Co-occurrence Matters: Multiple deficiencies compound risk - low Vitamin D + low Folate = 91% risk
• Prevention Economics: Early detection through ML screening is far more cost-effective than treating advanced deficiency

What's next for Devpost project

• Field Validation: Deploy in 2-3 pilot communities to validate model performance with real-world data
• Mobile Integration: Develop Android/iOS apps for community health workers
• Expanded Features: Incorporate seasonal dietary patterns and local food availability
• Multi-class Prediction: Move from binary to risk severity levels (low/medium/high/critical)

Built With

• python