Wrist-PPG Peak Detection Using Machine-Learned Refinement

The patent applies preliminary action by performing signal preprocessing steps (noise filtering, baseline correction) before peak detection to prepare the signal for more accurate processing. This preliminary preparation enables subsequent machine learning algorithms to achieve higher measurement precision without requiring overly complex real-time processing during actual peak detection.

The patent introduces machine learning algorithms as intermediary components between the raw PPG signal and the final peak detection. These algorithms act as mediators that process the signal through learned patterns and features, achieving high measurement precision by translating complex signal patterns into accurate peak identifications without requiring direct complex signal manipulation.

The patent performs computationally intensive machine learning training and feature extraction during the preliminary offline phase using historical data and reference signals. Once trained, the models can make rapid predictions during online peak detection with minimal computational energy, thus achieving high measurement precision while controlling real-time energy consumption through this time separation of computational tasks.

The patent changes the computational approach by transitioning from raw signal processing to feature-based processing. Instead of processing every sample of the PPG signal directly, the system extracts key features (peak positions, morphological characteristics) and uses these simplified representations for machine learning classification, significantly reducing the computational energy required while maintaining high measurement precision.

The patent implements self-service by training machine learning models to automatically adapt to different signal morphologies without requiring manual configuration or threshold adjustment. The models learn from training data the characteristics of various signal types and automatically adjust their detection criteria during inference, providing high adaptability while maintaining ease of operation through automated decision-making.

The patent changes the operational parameters from fixed adaptive thresholds to learned model parameters. Instead of using manually set thresholds that must be adjusted for different signal types, the system uses machine learning models that have automatically learned optimal detection parameters from training data, providing both high adaptability to different morphologies and ease of operation through automated parameter selection.

The patent applies local quality by implementing adaptive filtering that treats different regions of the signal differently. Instead of applying uniform filtering across the entire signal, the system identifies and preserves locally important features (such as peaks with specific morphological characteristics) while filtering out noise in other regions. This selective approach maintains signal reliability by removing noise while preserving diagnostically important local details.

The patent incorporates feedback mechanisms where the machine learning models continuously evaluate detected peaks against learned patterns and reference signals. If a peak fails to match expected characteristics despite passing initial filtering, the feedback system can trigger additional processing or adjustment, ensuring reliable peak identification while minimizing information loss through iterative verification rather than aggressive initial filtering.