End-to-End Machine Learning Pipeline for PPG Signal Analysis

Stage 4: Feature Extraction vs. Representation Learning

Handcrafted Feature Engineering

Traditional ML approaches for PPG extract domain-specific features from preprocessed signals. The three main feature categories are:

Time-domain features: Inter-beat interval (IBI) statistics (mean, standard deviation, RMSSD, pNN50), pulse amplitude, pulse width at half maximum, systolic rise time, diastolic decay time, and their variability measures. These features are computationally inexpensive and clinically interpretable.

Frequency-domain features: Power spectral density in cardiac band (0.5-4 Hz), spectral peak frequency (heart rate estimate), spectral entropy, ratio of harmonic power to total power (signal quality indicator), and HRV spectral features (LF power 0.04-0.15 Hz, HF power 0.15-0.4 Hz, LF/HF ratio). Welch's periodogram or autoregressive spectral estimation are standard methods.

Morphological features: Pulse wave features derived from individual beat waveforms, including the augmentation index (ratio of late systolic to early systolic peak amplitudes), reflection index, stiffness index, crest time, and the second derivative pulse wave analysis indices (a, b, c, d, e waves). These features are particularly informative for vascular health assessment and blood pressure estimation.

For a comprehensive treatment of PPG feature extraction, see our dedicated feature extraction guide.

End-to-End Deep Learning

End-to-end deep learning models learn feature representations directly from raw or minimally preprocessed PPG signals. The dominant architectures for PPG analysis are:

1D Convolutional Neural Networks (1D-CNNs): Extract local temporal patterns through convolutional kernels. A typical architecture uses 3-5 convolutional layers with kernel sizes of 5-15 samples, batch normalization, ReLU activation, and max pooling, followed by fully connected layers. ResNet-style skip connections improve gradient flow for deeper architectures. Hannun et al. (2019) demonstrated the effectiveness of a deep neural network for ECG rhythm classification. Trained on 91,232 single-lead ECGs from 53,549 patients, it classified 12 rhythm classes with an average ROC AUC of 0.97 against a cardiologist committee (doi:10.1038/s41591-018-0268-3).

Recurrent Neural Networks (LSTMs, GRUs): Capture long-range temporal dependencies across cardiac cycles. Bidirectional LSTMs are effective for offline analysis where future context is available. For real-time applications, unidirectional LSTMs or GRUs with 64-256 hidden units per layer are common.

CNN-LSTM Hybrids: Use CNN layers for local feature extraction and LSTM layers for temporal modeling across windows. This architecture has shown strong results for heart rate estimation (Biswas et al., 2019, doi:10.1109/TBCAS.2019.2892297) and AF detection from PPG.

Transformer-based Models: Self-attention mechanisms can capture long-range dependencies without the sequential processing bottleneck of RNNs. PPG transformers typically use patch embedding (dividing the signal into overlapping patches), positional encoding, and multi-head self-attention.

Temporal Convolutional Networks (TCNs): Dilated causal convolutions provide exponentially growing receptive fields without increasing parameter count. TCNs achieve comparable or superior performance to LSTMs for PPG time series while being more parallelizable during training. For comparison, Reiss et al. (2019, doi:10.3390/s19143079) used a CNN on time-frequency spectra of PPG and accelerometer signals. On their large PPG-DaLiA daily-life dataset, it cut mean absolute error by 31% versus classical methods.

Stage 5: Model Training and Validation

Loss Functions

Loss function selection depends on the task type:

Regression tasks (heart rate, blood pressure, SpO2): Mean squared error (MSE) or mean absolute error (MAE) loss. MAE is preferred for PPG regression because it is more robust to outliers, which are common due to residual motion artifacts. Huber loss (smooth L1) provides a compromise, behaving like MSE for small errors and MAE for large errors, with the transition threshold typically set at 5-10 BPM for heart rate estimation.

Classification tasks (AF detection, rhythm classification): Binary or categorical cross-entropy. For imbalanced datasets (typical for disease detection), focal loss (Lin et al., 2017, doi:10.1109/ICCV.2017.324) down-weights well-classified examples, focusing training on hard cases. The focal loss parameter $\gamma$ of 2.0 is a common starting point, with $\alpha$ set proportional to the inverse class frequency.

Multi-task learning: Combining heart rate estimation and signal quality assessment in a single model with shared feature extraction layers and task-specific heads can improve both tasks through implicit regularization. Weight the losses proportionally to their scale and relative importance (typically $\lambda_{\mathrm{SQA}} = 0.3$, $\lambda_{\mathrm{HR}} = 1.0$).

Cross-Validation Strategy

Subject-level cross-validation is mandatory for PPG models to obtain realistic generalization estimates. The standard approaches are:

Leave-one-subject-out (LOSO): Train on N-1 subjects, evaluate on the held-out subject, repeat N times. Provides the most complete use of limited data but is computationally expensive for large datasets and can produce high-variance estimates if individual subjects are unrepresentative.

Grouped k-fold: Partition subjects into k groups (typically k=5 or 10), maintaining all data from each subject within a single fold. Provides a balance between LOSO and simple k-fold, with reduced computational cost and stable estimates. For PPG datasets with 30+ subjects, grouped 5-fold cross-validation is recommended.

The validation fold should be used for hyperparameter tuning and early stopping, while the test fold provides the final unbiased performance estimate. Never tune hyperparameters on the test fold.

Data Augmentation

PPG data augmentation increases effective training set size and improves generalization. Effective augmentation strategies include:

Temporal augmentation: Random time warping (stretching/compressing by 5-15%), random cropping within the window, and jitter (adding small Gaussian noise with $\sigma$ = 1-5% of signal amplitude).

Amplitude augmentation: Random scaling (0.8-1.2x), random DC offset, and simulated baseline wander (adding low-frequency sinusoids at 0.1-0.3 Hz).

Physiological augmentation: Simulating heart rate changes by resampling individual beats to target different IBIs. This is particularly valuable for augmenting heart rate ranges underrepresented in the training data.

Channel augmentation: For multi-channel PPG, randomly dropping channels during training (channel dropout) improves robustness to channel-specific noise and sensor failures.

Um et al. (2017) showed that data augmentation of wearable accelerometer data raised CNN classification of Parkinson's disease motor states from 77.54% to 86.88% (doi:10.1145/3136755.3136817). Similar techniques apply to PPG, although gains must be confirmed for each task.

Stage 6: Deployment and Inference

Model Compression for Edge Deployment

Wearable PPG devices have severe computational constraints: ARM Cortex-M4 processors at 64-128 MHz, 256 KB - 1 MB SRAM, and power budgets of 1-10 mW for the inference engine. Deploying deep learning models on such platforms requires aggressive compression:

Quantization: Converting 32-bit floating-point weights and activations to 8-bit integers (INT8 quantization) reduces model size by 4x and inference latency by 2-4x with minimal accuracy degradation (typically less than 0.5% for well-designed architectures). Post-training quantization with representative calibration data is the simplest approach; quantization-aware training provides slightly better accuracy at the cost of training complexity. TensorFlow Lite Micro and CMSIS-NN provide optimized INT8 inference kernels for ARM Cortex-M processors.

Pruning: Removing weights below a magnitude threshold can reduce model size by 50-90% with moderate accuracy impact. Structured pruning (removing entire channels or layers) is more hardware-friendly than unstructured (individual weight) pruning because it maintains dense tensor operations.

Knowledge distillation: Training a small "student" model to mimic the outputs of a large "teacher" model. For PPG, a ResNet-34 teacher model can distill into a 3-layer CNN student with 10-50x fewer parameters and similar accuracy.

Latency and Throughput

Real-time PPG applications require inference within the analysis window update period. For a sliding window with 1-second advance at 25 Hz sample rate, the model must complete inference within 1 second (including all preprocessing). On a Cortex-M4 at 80 MHz, a typical 3-layer 1D-CNN with INT8 quantization completes inference in 5-50 ms for 250-sample (10-second) input windows, leaving ample headroom for preprocessing.

For cloud-based processing (where PPG data is transmitted from the device to a server), latency requirements are relaxed but network reliability becomes a concern. Hybrid architectures that perform basic processing (beat detection, quality assessment) on-device and transmit summary features for complex inference (arrhythmia classification, BP estimation) offer a practical compromise.

Monitoring and Drift Detection

Deployed PPG models require ongoing monitoring to detect performance degradation. Model confidence (softmax entropy for classifiers, prediction interval width for regressors) should be tracked over time. Increasing confidence degradation may indicate sensor aging, firmware updates affecting signal characteristics, or population drift (the model encountering physiological patterns not represented in training data).

For a deeper understanding of the algorithms that underpin these ML pipelines, explore our PPG signal processing algorithms reference and the foundational concepts in our PPG technology overview.

References

• Elgendi, M. (2016). Optimal signal quality index for photoplethysmogram signals. Bioengineering, 3(4), 21. https://doi.org/10.3390/bioengineering3040021
• Elgendi, M., Norton, I., Brearley, M., Abbott, D., & Schuurmans, D. (2013). Systolic peak detection in acceleration photoplethysmograms measured from emergency responders in tropical conditions. PLoS ONE, 8(10), e76585. https://doi.org/10.1371/journal.pone.0076585
• Hannun, A. Y., Rajpurkar, P., Haghpanahi, M., Tison, G. H., Bourn, C., Turakhia, M. P., & Ng, A. Y. (2019). Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature Medicine, 25(1), 65-69. https://doi.org/10.1038/s41591-018-0268-3
• Reiss, A., Indlekofer, I., Schmidt, P., & Van Laerhoven, K. (2019). Deep PPG: Large-scale heart rate estimation with convolutional neural networks. Sensors, 19(14), 3079. https://doi.org/10.3390/s19143079
• Um, T. T., Pfister, F. M. J., Pichler, D., Endo, S., Lang, M., Hirche, S., Fietzek, U., & Kulić, D. (2017). Data augmentation of wearable sensor data for Parkinson's disease monitoring using convolutional neural networks. In Proceedings of the 19th ACM International Conference on Multimodal Interaction (pp. 216-220). https://doi.org/10.1145/3136755.3136817