Stratified locality-sensitive hashing for accelerated physiological time series retrieval

Area of Science:

• Biomedical Informatics
• Signal Processing
• Machine Learning

Background:

• Physiological waveform analysis is crucial for medical diagnosis.
• Existing methods like standard locality-sensitive hashing (LSH) have limitations in data perspective.
• Efficient retrieval of similar time series is computationally challenging.

Purpose of the Study:

• To introduce Stratified Locality-Sensitive Hashing (SLSH) for enhanced physiological waveform time series retrieval.
• To improve upon the sublinear retrieval times of standard LSH.
• To enable a more diverse and refined examination of waveform data.

Main Methods:

• Developed SLSH by incorporating multiple LSH families with varied distance functions (l1 and cosine).
• Implemented SLSH for analyzing arterial blood pressure time series data.
• Compared SLSH performance against standard LSH and linear search using the MIMIC2 database.

Main Results:

• SLSH achieved 1.7 times faster retrieval than standard LSH.
• SLSH demonstrated a 14 times speedup compared to linear search.
• A 5% decrease in accuracy was accepted as a trade-off for significant speed improvements.

Conclusions:

• SLSH offers a significant acceleration in retrieving similar physiological waveforms.
• The multi-perspective approach of SLSH enhances data examination capabilities.
• SLSH provides an efficient and effective solution for large-scale physiological waveform analysis.