Triaxial Accelerometer Sleep State Classification via Spectral Feature Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing algorithms for classifying sleep and wake states from actigraphy data, such as the Cole-Kripke algorithm, rely on proprietary measures of activity, limiting their usability and accuracy, especially when using widely available triaxial accelerometer data from smart devices.
Innovation Solution
A process involving triaxial accelerometer data is developed, using features like standard deviation and peak analysis, clustered with K-means, and modeled with a hidden Markov model using the Baum Welch algorithm to identify sleep and wake states, which can be applied to classify sleep-wake patterns robustly using a single day's data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional algorithms like Cole-Kripke are used for sleep-wake classification, then the classification process is simple, but the accuracy and sensitivity are limited due to reliance on proprietary activity measures
Solution Approach 1:
The patent transforms the input data representation by converting raw accelerometer signals into spectral features through Fast Fourier Transform (FFT). This parameter transformation extracts frequency-domain characteristics that better represent sleep-wake patterns, improving classification accuracy without requiring complex proprietary algorithms. The spectral features capture rhythmic patterns inherent in human activity that are more discriminative for sleep state classification.
Solution Approach 2:
The patent replaces traditional mechanical signal processing approaches with machine learning-based classification. Instead of using rule-based algorithms like Cole-Kripke that rely on proprietary activity measures, the system employs trained classifiers (support vector machines, random forests, or neural networks) that learn optimal decision boundaries from training data. This substitution enables the system to achieve superior accuracy using standard accelerometer data without proprietary measures.
2Measurement precision
If more features and complex modeling are used to improve classification accuracy, then the accuracy increases to 92.99%, but the computational complexity and processing time increase
Solution Approach 1:
The patent performs feature extraction and spectral analysis in advance before classification. By pre-computing FFT-based spectral features and storing them in a structured format, the system prepares the data for rapid classification. This preliminary processing separates the computationally intensive feature extraction from the real-time classification decision, enabling fast inference while maintaining high accuracy.
Solution Approach 2:
The patent divides the classification task into distinct stages: (1) signal preprocessing and FFT transformation, (2) spectral feature extraction, (3) feature normalization and selection, and (4) classification. This segmentation allows each stage to be optimized independently, with parallel processing potential in the feature extraction phase, reducing overall processing time while maintaining the benefits of complex feature engineering.
3Adaptability or versatility
If subject-specific normalization is applied to improve model generalization, then the cross-subject accuracy improves, but the processing complexity per subject increases
Solution Approach 1:
The patent applies z-score normalization to standardize spectral features across different subjects. By transforming features to have zero mean and unit variance for each subject, the system removes inter-subject variability in signal amplitude and scale. This parameter transformation enables the classification model to generalize across subjects without requiring subject-specific model training, achieving adaptability through simple statistical normalization rather than complex personalized modeling.
Data Source
AI summary
A device may obtain triaxial accelerometer training data for a plurality of subjects. A device may define a number of feature vectors for each time interval and each subject of the triaxial accelerometer data. A device may cluster the feature vectors of the triaxial training data into a number of clusters to obtain a cluster assignment for each of the feature vectors of each subject. A device may fit a hidden Markov model to the to the triaxial accelerometer training data cluster assignments. A device may identify at least one state for the subjects based on the cluster assignments.