Pattern Dictionary Anomaly Detection for Wearable Physiological Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional anomaly detection techniques for wearable data fail to accurately and efficiently identify anomalous sequences due to their inability to consider the sequential structure of time series data and high computational complexity, especially when dealing with large sequence alphabets, leading to inconsistent characterization and ineffective health condition prediction.
Innovation Solution
A system utilizing a pattern dictionary and anomaly detection model to parse physiological data into parsed phrases, determine codelength, and compare against thresholds to generate alerts for anomalous data, with optional training and updating of the pattern dictionary based on user input and crowd-sourced data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional anomaly detection techniques are applied to wearable data, then the detection process can be performed, but the characterization is inconsistent and accuracy is insufficient
Solution Approach 1:
The patent segments the physiological time series data into discrete tokens based on a vocabulary of physiological patterns. Each token represents a segmented portion of the time series, allowing the system to analyze individual pattern units rather than treating the entire sequence as a monolith. This segmentation enables consistent characterization by breaking down complex temporal patterns into standardized, comparable units.
Solution Approach 2:
The patent transforms the time series data into a token sequence by changing the representation parameter from continuous temporal values to discrete symbolic tokens. This parameter transformation allows the anomaly detection model to work with standardized discrete inputs, improving both accuracy and consistency in characterizing anomalous patterns.
2Productivity
If conventional anomaly detection techniques are used, then processing can be performed, but computational complexity becomes excessively high
Solution Approach 1:
By segmenting the time series into discrete tokens, the patent reduces the computational complexity of analyzing physiological data. Instead of processing the entire continuous time series, the system only needs to process a manageable sequence of discrete tokens, significantly reducing the computational burden while maintaining detection effectiveness.
Solution Approach 2:
The patent creates a simplified token representation that copies the essential characteristics of the original time series data while discarding redundant temporal information. This token copy allows for efficient processing and comparison without requiring complex analysis of the original continuous data, reducing computational complexity while preserving detection accuracy.
3Measurement precision
If conventional techniques are applied to time series data, then analysis can be performed, but the sequential structure is not considered
Solution Approach 1:
The patent performs preliminary action by pre-defining a vocabulary of physiological patterns and pre-segmenting the time series data into tokens based on this vocabulary. This preliminary preparation establishes the sequential structure and pattern relationships before anomaly detection occurs, allowing the system to accurately characterize sequences without adding complex methodological steps during the actual detection process.
Data Source
AI summary
Systems and methods for enhancing anomaly detection using a pattern dictionary are disclosed. An example method includes receiving, from a wearable device, physiological data of the user, and parsing the physiological data into a set of parsed phrases having a number of parsed phrases by applying a pattern dictionary encoder using a pattern dictionary. Each parsed phrase represents a respective subsequence of the physiological data. The example method includes determining a codelength corresponding to the physiological data based on the set of parsed phrases, and comparing (i) the number of parsed phrases to a parsed phrase threshold, and (ii) the codelength to a codelength threshold using an anomaly detection model. Responsive to the number of parsed phrases exceeding the parsed phrase threshold or the codelength exceeding the codelength threshold, the example method includes generating an alert for display on a user interface indicating that the physiological data is anomalous.