Unsupervised Bayesian Inference to Fuse Biosignal Sensory Estimates for Personalizing Care

Area of Science:

• Biomedical Engineering
• Machine Learning
• Data Science

Background:

• Wearable sensors generate vast biosignal data, necessitating automated labeling for precision care.
• Clinical applications often lack expert labels, requiring trust in algorithms without ground truth.
• Evaluating and selecting optimal algorithms for diverse patient data is challenging.

Purpose of the Study:

• To develop unsupervised Bayesian models for fusing labels from multiple annotators (algorithms/experts).
• To estimate annotator bias and precision without ground truth, inferring underlying data truth.
• To improve upon existing strategies by incorporating annotator correlations and task difficulty.

Main Methods:

• Proposed two fully Bayesian generative models for label fusion.
• Employed Gibbs sampling for inference, incorporating annotator correlations and sensor data quality.
• Validated models on simulated and two public biomedical datasets.

Main Results:

• The proposed models outperform existing literature approaches in estimating ground truth.
• Demonstrated robustness with missing data, common in real-world biomedical applications.
• Showcased efficiency for real-time applications, supporting personalized care.

Conclusions:

• Novel Bayesian fusion models offer reliable estimation of biosignal data truth without ground truth.
• Incorporating annotator correlations and task difficulty significantly enhances estimation accuracy.
• These methods provide a robust foundation for real-time, personalized healthcare using sensor data.