kmlShape: An Efficient Method to Cluster Longitudinal Data (Time-Series) According to Their Shapes.

Area of Science:

• Biostatistics
• Data Science
• Computational Biology

Background:

• Longitudinal data analysis involves repeated measurements over time.
• Traditional clustering methods group individuals with temporally close trajectories.
• These methods may miss similarities in trajectory shapes occurring at different times.

Purpose of the Study:

• To develop a novel algorithm for partitioning longitudinal data based on trajectory shape.
• To address the limitation of classical methods that focus on temporal proximity.
• To enable the identification of groups with similar patterns of change, irrespective of timing.

Main Methods:

• A new longitudinal data partitioning algorithm is presented, focusing on trajectory shapes.
• The algorithm utilizes shape-based comparisons instead of classical distance metrics.
• Two data simplification procedures are proposed to enhance applicability to high-dimensional datasets.

Main Results:

• The algorithm identified a distinct "rapid decline" patient group in Alzheimer's disease data, missed by traditional methods.
• Analysis of the female menstrual cycle revealed two luteinizing hormone peaks in 22% of women, challenging existing literature.
• Demonstrates the algorithm's ability to uncover novel patterns in complex biological data.

Conclusions:

• The proposed shape-based partitioning algorithm offers a powerful alternative for analyzing longitudinal data.
• It provides new insights into disease progression and biological cycles by focusing on pattern similarity.
• This method enhances the discovery of previously unidentified subgroups within longitudinal datasets.