Enhancing statistical power in temporal biomarker discovery through representative shapelet mining.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Data Science

Background:

• Temporal biomarker discovery in longitudinal data relies on identifying recurring subsequences called shapelets.
• Current methods struggle with redundant and overlapping shapelets, leading to impractical and underpowered analyses.
• Pre- or post-processing of shapelets has not sufficiently improved their power or utility.

Purpose of the Study:

• To introduce a novel method for temporal biomarker discovery that addresses limitations of existing approaches.
• To retrieve statistically significant shapelets that are structurally diverse and manageable in quantity.
• To enhance the statistical power and practical utility of temporal biomarker discovery.

Main Methods:

• Developed Statistically Significant Submodular Subset Shapelet Mining (S5M).
• Employed submodular optimization to prune non-representative shapelets and maximize structural diversity.
• Validated S5M on simulated and real-world longitudinal datasets.

Main Results:

• S5M successfully retrieves short subsequences that occur in the data and are statistically significantly associated with the phenotype.
• The method ensures a manageable quantity of shapelets while maximizing structural diversity.
• S5M demonstrated increased statistical power and utility compared to state-of-the-art methods.
• Identified temporal patterns in Sequential Organ Failure Assessment scores associated with in-ICU mortality in severe organ failure patients.

Conclusions:

• S5M offers a powerful and practical approach to temporal biomarker discovery.
• The method enhances the interpretability and utility of identified temporal patterns.
• S5M is available as an option in the S3M Python package.