Self-Supervised Learning for Wearable Sensor Data Imputation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning analysis of wearable device sensor data is complicated by gaps created when devices are deactivated or removed, leading to incomplete datasets that hinder accurate health event detection and chronic condition monitoring.
Innovation Solution
A self-supervised learning method that generates masked data records by mimicking natural patterns of missingness in wearable device data, allowing for the training of machine learning models to impute missing data and improve downstream analysis tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If wearable devices are used to collect sensor data outside clinical settings, then data collection coverage and real-world applicability are improved, but data completeness deteriorates due to device removal or deactivation
Solution Approach 1:
The patent converts the harmful effect of missing data into a beneficial training signal by using the natural missingness patterns to create masked data records for self-supervised learning. The model learns to impute missing values by training on data where portions are systematically masked according to real-world missingness patterns, thereby turning the problem of incomplete data into a feature that enhances model robustness and generalizability to real-world deployment conditions.
2Ease of manufacture
If random masking is used for self-supervised learning, then training simplicity is improved, but effectiveness in handling real-world missing data patterns deteriorates
Solution Approach 1:
The patent applies local quality by using different masking strategies for different portions of the data based on the actual missingness patterns observed in each user's data. Instead of uniform random masking, the system identifies specific time periods or data segments where missingness occurs and applies masking accordingly, making the training process locally adapted to each user's real-world data characteristics rather than applying a global uniform approach.
3Productivity
If machine learning models are trained on complete data sets, then model convergence speed is improved, but performance on incomplete real-world data deteriorates
Solution Approach 1:
The patent implements preliminary action by pre-training the model on complete data sets to achieve fast convergence and establish baseline performance, then subsequently training on masked data records that simulate real-world missingness patterns. This two-stage approach allows the model to first learn from ideal complete data efficiently, then adapt to handle incomplete data scenarios, combining the benefits of fast convergence with robust real-world performance.
Data Source
AI summary
Disclosed is a method comprising accessing, by a machine learning system, a set of data records for a plurality of users, the data records representative of physical statistics measured for each of the plurality of users over a time period. At least a subset of the data records comprises patterns of missing data for at least a portion of the time period. The method also comprises generating a set of masked data records by masking a subset of the data records in accordance with a pattern of natural missingness from a data record. The method also comprises generating, by the machine learning system, a set of learned representations from at least the set of masked data records. Finally, the method comprises fine tuning, by the machine learning system, a machine learning model using the set of learned representations, the machine learning model configured to perform a downstream machine learning task.