Inspiration
A devastating chat with the mother of a Thalassemia sufferer in Mumbai served as the impetus for Hemo Connect. She talked about the restless nights she spent frantically phoning blood banks at two in the morning to request O-negative blood for her daughter, who is eight years old, to have an emergency transfusion. "If only we could know in advance," she added, "maybe my child wouldn't have to suffer." This incident brought to light a terrible reality: a patient with Thalassemia in India risks a possible blood shortage crisis every two minutes. In light of Blood Warriors' goal to eradicate Thalassemia by 2035, we came to the conclusion that anticipating and averting these crises before they arise was more important than just recruiting additional donors. Finding out that 40% of blood donors quit after the first year not because they don't care, but rather because the system doesn't effectively involve them deepened our motivation. We identified a chance to employ AI for behavioral prediction and retention in addition to matching, converting infrequent donations into devoted supporters.
What it does
Hemo Connect would be India's first predictive blood ecosystem that transforms Thalassemia care from reactive crisis management to proactive precision healthcare. Here's how it works: Predictive Blood Banking Our AI engine forecasts blood availability 2-4 weeks in advance using:
Donor Behavior Analysis: LSTM neural networks analyze behavioral patterns Seasonal Intelligence: Accounts for festivals, weather, and local events affecting donations Churn Prevention: Identifies donors likely to stop donating with bgreat accuracy
Intelligent Patient-Donor Matching Beyond basic ABO-Rh compatibility, our system:
Geographic Optimization: Minimizes travel time while ensuring quality Cross-matching Prediction: Prevents transfusion reactions using ML models Emergency Triage: AI-powered priority system for critical cases
WhatsApp-Powered Care Assistant Leveraging India's favorite messaging platform:
Medication Reminders: Personalized chelation therapy schedules Symptom Tracking: Early warning system for complications Family Coordination: Automated appointment scheduling and updates Multilingual Support: 12 Indian languages with cultural sensitivity
Community Health Network
Peer Matching: Connects patients with similar conditions and age groups Support Groups: AI identifies optimal group compositions for mental health Knowledge Sharing: Curated experience sharing between families
How we built it
Donor Prediction Model :
model = Sequential([
LSTM(128, return_sequences=True, input_shape=(timesteps, features)),
Dropout(0.3),
LSTM(64, return_sequences=False),
Dense(32, activation='relu'),
Dense(1, activation='sigmoid') # Probability of donation
])
Mathematical Foundation: The core prediction uses a modified LSTM with attention mechanism: P(donationt+k)=σ(Wh⋅ht+Wx⋅xt+b)P(donation_{t+k}) = \sigma(W_h \cdot h_t + W_x \cdot x_t + b)P(donationt+k)=σ(Wh⋅ht+Wx⋅xt+b) Where:
$h_t$ = hidden state incorporating historical patterns $x_t$ = current contextual features (weather, festivals, etc.) $k$ = prediction horizon (1-4 weeks)
- Matching Algorithm Our multi-criteria decision analysis uses weighted scoring: Score=∑i=1nwi⋅fi(donor,patient)Score = \sum_{i=1}^{n} w_i \cdot f_i(donor, patient)Score=∑i=1nwi⋅fi(donor,patient) Where factors include:
$f_1$ = Blood compatibility (weight: 0.4) $f_2$ = Geographic distance (weight: 0.2) $f_3$ = Donor reliability score (weight: 0.2) $f_4$ = Historical success rate (weight: 0.2)
Technology Stack
Backend: Python Django + FastAPI for ML serving AI/ML: TensorFlow 2.8, scikit-learn, pandas, numpy Database: PostgreSQL (patient data) + Redis (real-time caching) Integration: WhatsApp Business API, e-RraktKosh API Cloud: Microsoft Azure (ML Studio, Cognitive Services) Frontend: React.js for healthcare dashboard Mobile: Progressive Web App for cross-platform compatibility
Data Pipeline
Data Ingestion: Real-time sync with e-RraktKosh, Blood Warriors database Feature Engineering: 47 behavioral and contextual features extracted Model Training: Automated retraining every 24 hours with new data Prediction Serving: Sub-200ms response time via Redis caching Feedback Loop: Continuous learning from actual donation outcomes
