Building Anupraan: Our AI-Powered Health Monitoring System#

Posted on January 15, 2024

Anupraan represents our most ambitious project yet - combining IoT sensors, machine learning, and healthcare technology into a comprehensive health monitoring solution. This post covers our journey from concept to working prototype.

The Vision#

We wanted to build something that could genuinely improve people’s lives. Traditional health monitoring is reactive - you only know something’s wrong after it happens. Anupraan aims to be proactive, using AI to detect patterns and provide early warnings.

Hardware Design#

Sensor Selection#

After extensive research, we chose these sensors for their accuracy and reliability:

  • MAX30100: Pulse oximeter and heart rate sensor
  • MLX90614: Infrared temperature sensor (non-contact)
  • MPU6050: Accelerometer for movement and fall detection
  • ESP32: Dual-core microcontroller with WiFi/Bluetooth

PCB Design#

We designed a custom PCB to integrate all sensors cleanly:

# Main board layout
ESP32-WROOM-32 (center)
├── MAX30100 (top-left)
├── MLX90614 (top-right)  
├── MPU6050 (bottom-left)
└── Power management (bottom-right)

The board includes proper filtering, voltage regulation, and EMI shielding for medical-grade reliability.

Software Architecture#

Edge Processing Pipeline#

class HealthMonitor:
    def __init__(self):
        self.sensors = {
            'heart_rate': MAX30100(),
            'temperature': MLX90614(),
            'movement': MPU6050()
        }
        self.ml_model = AnomalyDetector()
    
    def collect_data(self):
        readings = {}
        for name, sensor in self.sensors.items():
            readings[name] = sensor.read()
        return readings
    
    def analyze(self, data):
        # Local ML inference
        anomaly_score = self.ml_model.predict(data)
        if anomaly_score > THRESHOLD:
            self.trigger_alert(data, anomaly_score)
        return anomaly_score

Machine Learning Model#

We trained a lightweight anomaly detection model using TensorFlow Lite:

# Model architecture
model = Sequential([
    LSTM(64, input_shape=(timesteps, features)),
    Dropout(0.2),
    Dense(32, activation='relu'),
    Dense(1, activation='sigmoid')
])

# Training data: 10,000+ hours of normal vital signs
# Anomaly detection using One-Class SVM approach

Calibration & Validation#

Sensor Calibration#

Each sensor requires careful calibration:

def calibrate_max30100():
    """Calibrate pulse oximeter using known reference values"""
    reference_values = [95, 98, 100]  # SpO2 percentages
    measured_values = []
    
    for ref in reference_values:
        # Use calibration device
        measured = sensor.read_spo2()
        measured_values.append(measured)
    
    # Calculate calibration coefficients
    calibration_factor = np.mean(reference_values) / np.mean(measured_values)
    return calibration_factor

Clinical Validation#

We’re working with healthcare professionals to validate our measurements against medical-grade equipment. Initial results show 95%+ accuracy for heart rate and 90%+ for SpO2.

Power Management#

Sleep Modes#

void enter_sleep_mode() {
    // Deep sleep for 30 seconds
    esp_sleep_enable_timer_wakeup(30 * 1000000);
    
    // Keep only essential sensors active
    gpio_deep_sleep_hold_en();
    
    // Enter deep sleep
    esp_deep_sleep_start();
}

Battery Life Optimization#

  • Active mode: 12 hours continuous monitoring
  • Sleep mode: 7+ days standby
  • Selective sensing: Only activate sensors when needed

Data Privacy & Security#

Local Processing#

All sensitive health data is processed locally on the ESP32. Only anonymized, aggregated data is sent to the cloud (with user consent).

Encryption#

from cryptography.fernet import Fernet

def encrypt_health_data(data):
    key = Fernet.generate_key()
    cipher = Fernet(key)
    encrypted_data = cipher.encrypt(json.dumps(data).encode())
    return encrypted_data, key

Current Challenges#

1. Sensor Drift#

Over time, sensor readings can drift. We’re implementing auto-calibration using known good values.

2. Battery Life#

Continuous monitoring is power-hungry. We’re exploring energy harvesting from body heat and movement.

3. Regulatory Compliance#

Medical devices require FDA approval. We’re consulting with regulatory experts to navigate this process.

What’s Working#

  • ✅ Sensor integration and basic readings
  • ✅ Local ML inference (95% accuracy)
  • ✅ Power management (7+ days battery)
  • ✅ Data encryption and privacy controls
  • ✅ Real-time anomaly detection

What’s Next#

  1. Mobile App: React Native companion app for data visualization
  2. Cloud Backend: FastAPI server with PostgreSQL database
  3. Clinical Trials: Partner with healthcare providers for validation
  4. Regulatory: Begin FDA approval process

Lessons Learned#

  1. Medical-grade accuracy is hard - consumer sensors need extensive calibration
  2. Privacy is paramount - health data security can’t be compromised
  3. Battery life matters - users won’t charge devices daily
  4. Validation takes time - clinical testing is essential for health tech

Get Involved#

Anupraan is open source! We’re looking for contributors in:

  • Embedded systems and IoT
  • Machine learning and data science
  • Mobile app development
  • Healthcare and medical expertise
  • Regulatory compliance

Check out our GitHub repository and join our Discord community.

Questions about the build? Want to contribute? Get in touch or join our development discussions.