AI-Powered Sleep Quality Analysis from Sensor Data
Wrist accelerometer during sleep writes characteristic patterns for different phases: deep sleep—near total stillness with periodic micromovement, REM—rare movements with rising pulse, wakefulness—clear activity. Model task—correctly label 8 hours of this data.
Raw Data Sources
On iOS, sleep data available via HealthKit as HKCategoryType(.sleepAnalysis). Apple Watch and third-party trackers (Oura, Fitbit, Garmin) record there. But for raw accelerometer and GyroScope data for own ML classification—HealthKit doesn't provide retrospectively. Need background app with CMMotionManager writing data to disk.
class NightMotionRecorder {
private let motionManager = CMMotionManager()
private var dataBuffer: [(timestamp: Date, x: Double, y: Double, z: Double)] = []
func startNightRecording() {
guard motionManager.isAccelerometerAvailable else { return }
motionManager.accelerometerUpdateInterval = 1.0 / 25.0 // 25 Hz sufficient for sleep
motionManager.startAccelerometerUpdates(to: .main) { [weak self] data, error in
guard let data = data else { return }
self?.dataBuffer.append((
timestamp: Date(),
x: data.acceleration.x,
y: data.acceleration.y,
z: data.acceleration.z
))
}
}
}
25 Hz—balance between accuracy and battery drain. Sleep doesn't need 100 Hz accelerometer. Background mode for CMMotionManager requires UIBackgroundModes: motion in Info.plist—but iOS aggressively terminates background tasks. More reliable: BGProcessingTask for overnight data processing with partial recordings.
SpO2 from HealthKit
Pulse oximetry via HKQuantityType(.oxygenSaturation). Apple Watch Series 6+ writes SpO2 every 1-2 hours at night in background. Drop below 90%—sleep apnea sign. But Apple doesn't write continuous night SpO2 (battery reasons), so intermittent data requires careful interpolation.
Sleep Stage Classification Algorithm
Standard task—4 classes: Wake, Light NREM, Deep NREM (N3), REM. Clinically valid standard—polysomnography (PSG) with EEG. Accelerometer and heart rate sufficient for Wake/Sleep separation with ~85% accuracy, full 4-stage classification—60–75% PSG agreement.
Feature Engineering from Accelerometer
From raw 25 Hz signal per 30-second epochs compute:
- Activity count—sum of absolute acceleration vector changes (Cole-Kripke algorithm)
- ZCR (Zero Crossing Rate)—zero-crossing frequency, correlates with fine motor
-
ENMO (Euclidean Norm Minus One)—actigraphy standard,
sqrt(x²+y²+z²) - 1g, removes gravity - Angle z-axis—wrist angle characteristic for different sleep positions
From HR (if available) add:
- Resting HR vs current HR (delta)
- HRV (RMSSD from RR-intervals)—higher in REM than deep sleep
Model
Random Forest on these features gives reasonable baseline. For temporal context—LSTM over RF features: RF outputs feature vector for each 30-sec epoch, LSTM considers epoch sequence. Pattern from Stanford Sleep Lab research.
# Feature extraction for one epoch (30 sec, 750 samples at 25 Hz)
def extract_epoch_features(epoch_data):
x, y, z = epoch_data[:, 0], epoch_data[:, 1], epoch_data[:, 2]
enmo = np.maximum(np.sqrt(x**2 + y**2 + z**2) - 1, 0)
angle_z = np.arctan(z / np.sqrt(x**2 + y**2 + 1e-6)) * 180 / np.pi
return {
'enmo_mean': np.mean(enmo),
'enmo_std': np.std(enmo),
'enmo_max': np.max(enmo),
'angle_z_mean': np.mean(angle_z),
'angle_z_std': np.std(angle_z),
'activity_count': np.sum(np.abs(np.diff(enmo)))
}
Model Conversion and Deployment
For iOS: sklearn Random Forest → coremltools.converters.sklearn.convert() → .mlmodel. LSTM layer—separate CoreML neural net. Combine in CoreML Pipeline. For Android: both models via TFLite with NNAPI delegate for hardware acceleration.
UI: Result Interpretation
Hypnogram (sleep phase timeline chart)—standard display. Additionally: Sleep Score as aggregate metric, breakdown per phase duration, identified patterns (late sleep onset, frequent awakenings).
Tie recommendations to patterns: "Deep sleep N3 share dropped 18% to 9% over 5 nights" + specific advice, not "improve sleep quality".
Development Process
Choose data sources (HealthKit vs raw accelerometer). Develop night data processing pipeline. Feature engineering and classification model training. Conversion and app integration. UI components: hypnogram and recommendation. Test accuracy on real wearable data.
Timeframe Estimates
Wake/Sleep detector with actigraphy and basic UI—1–2 weeks. Full 4-stage classifier with hypnogram and personalized recommendations—3–5 weeks.