AI Analysis of Wearable Medical Device Data
Case: Extracting Clinical Value from Terabytes of PPG and Accelerometer Data
You manufacture smartwatches and want to add atrial fibrillation (AF) detection from the PPG sensor. At first glance, the task is solvable, but reality brings corrections: motion noise, battery depletion, irregular artifacts. Without an AI system trained on thousands of labeled recordings, accuracy stays below 75%, which is unacceptable for medical use. We developed a pipeline that improved sensitivity to 95% with specificity of 93% — a 1.5x improvement over standard approaches. Here’s how we did it.
Why Wearable AI Models Require MLOps
A model that works perfectly on clean data starts making mistakes a month after a smartwatch firmware update due to signal drift. Without MLOps, tracking and retraining are impossible. We build pipelines with MLflow for logging, DVC for data versioning, and automatic retraining triggers when metrics drop. This ensures stable production quality — a key element of MLOps in healthcare. Additionally, on-device inference reduces latency and preserves data privacy.
What Data Do Wearable Devices Collect?
Consumer wearables (Apple Watch, Garmin, Polar) generate:
- Heart rate (PPG sensor, 1–5 Hz)
- SpO₂ (photoplethysmography)
- ECG (single-channel Lead I, Apple Watch Series 4+)
- Accelerometer (3-axis, 50–100 Hz): activity, steps, falls
- Galvanic skin response (EDA): stress
- Skin temperature (Fitbit Sense, Oura Ring)
Medical wearables provide more accurate data:
- CGM (Dexcom G7, FreeStyle Libre): glucose every 5 minutes
- Patch ECG monitors (iRhythm Zio, BioTel): continuous 14-day ECG
- Ambulatory BP: 24-hour blood pressure
- Smart inhalers: timing and technique
Specialized sensors include EMG patches for muscle activity assessment and orthopedic insoles with force plates for gait analysis.
How We Build Health Monitoring AI Algorithms
Heart Rate Variability (HRV)
HRV is a key marker of autonomic regulation. We use time-domain (RMSSD, SDNN) and frequency-domain (LF, HF, LF/HF) features fed into an LSTM model. LSTM provides a 15–20% accuracy gain over gradient boosting by capturing temporal dependencies. The model is trained on data from 500+ patients and detects risk of sudden cardiac death with 92% sensitivity.
Hypoglycemia Prediction from CGM
Glucose every 5 minutes + accelerometer + time of day → LSTM predicts glucose levels at 30 and 60 minutes. Prediction error does not exceed 10 mg/dL. For T1D patients, this enables warning hypoglycemia before symptoms appear. Reducing hypoglycemic episodes cuts emergency care costs.
Fall Detection from Accelerometer
Fall pattern: acceleration increase → impact → inactivity. We train a CNN on the 3-axis signal. The main challenge is false positives (jumps, fast movements). Solution: a personalized threshold adapted to age and individual movement patterns. False positive rate reduced from 12% to 3% — a 4x reduction in false alarms. This significant reduction in false alarms leads to substantial cost savings.
Atrial Fibrillation Detection from PPG
PPG is less informative than ECG but available in every smartwatch. A deep network analyzes the PPG waveform shape and rhythm regularity. After fine-tuning on 10,000 recordings, we achieved sensitivity 95% and specificity 93%. Learn more about atrial fibrillation.
Accuracy Comparison
| Metric | Before AI Model | After Our Model |
|---|---|---|
| AF sensitivity | 75% | 95% |
| Fall false positive rate | 12% | 3% |
| Glucose prediction error | 20 mg/dL | 10 mg/dL |
Algorithm Comparison for HRV Analysis
| Model | RMSE | Training Time |
|---|---|---|
| XGBoost | 8.2 ms | 15 min |
| LSTM | 5.1 ms | 2 hours |
| Transformer | 4.8 ms | 4 hours |
Ensuring Model Accuracy on Wearables
Accuracy depends on data quality and regular retraining. We use cross-validation and independent test sets. The PhysioNet (Goldberger et al., 2000) datasets are widely used for validation. Additionally, we integrate drift detection mechanisms — if the feature distribution changes, the system automatically triggers retraining. This maintains high performance even under changing operating conditions.
End-to-End AI System Development Stages
- Define clinical task — gather requirements, select sensors, assess regulatory class.
- Collect labeled data — synchronize with existing databases (PhysioNet, MIMIC) or run pilot studies.
- Feature engineering — extract HRV, spectral features, time windows.
- Model selection and training — experiment with LSTM, CNN, Transformer; hyperparameter search; ensemble.
- Validation and testing — cross-validation, independent test set, evaluate sensitivity/specificity.
- Deployment — on-device (TFLite, ONNX Runtime) for fast inference; cloud pipeline for retraining.
- Monitoring and retraining — track data drift, automatically update model on schedule.
To version your data and models, a typical DVC pipeline command might look like:
dvc run -n train -d data/ -d src/train.py -m metrics.json --outs models/model.pkl python src/train.py
Common Pitfalls
- Insufficient labeled data: medical AI needs at least 1000 labeled records per class.
- Poor signal quality: standard filters don't always remove artifacts; custom preprocessing required.
- Ignoring regulatory requirements: SaMD may need certification, delaying release.
- No MLOps: without monitoring, models degrade within 3–6 months.
What's Included in Our Work
- Trained model (ONNX/TFLite) with validation documentation
- Source code with MLOps pipeline (MLflow, DVC)
- Deployment instructions (Docker, Kubernetes)
- Customer team training (3–5 people, 2 days)
- 3 months of support and consulting
Our experience includes over 5 years working with medical IoT data and 30+ successful projects. The cost of development is calculated individually, depending on algorithm complexity and data volume. Request a data analysis — we will analyze your data free of charge and select an effective pipeline. Contact us to start your project.