Information mining over heterogeneous and high-dimensional time-series data in clinical trials databases

Area of Science:

• Biomedical Data Analysis
• Computational Biology
• Clinical Informatics

Background:

• Clinical trial data often presents challenges like heterogeneity and high dimensionality in time series.
• Existing time series analysis methods frequently fail due to unmet assumptions regarding data length and interval consistency.
• These limitations are particularly pronounced in clinical trial datasets, necessitating revised data mining approaches.

Purpose of the Study:

• To propose a novel information mining approach tailored for heterogeneous, high-dimensional time series clinical trial data.
• To develop methods for identifying common and distinct patterns within complex biomedical datasets.
• To enhance the analysis of time series data by addressing limitations of current statistical techniques.

Main Methods:

• A two-step approach involving data mining on homogeneous subsets and pattern identification.
• Application of frequent itemset mining, clustering, and declustering techniques with novel distance metrics.
• Utilizing modified algorithms for effective declustering and feature selection in high-dimensional time series.

Main Results:

• Successful clustering identified strongly correlated analyte groups, validating known biological relationships and discovering novel ones.
• Declustering enabled effective feature selection, identifying a concise set of analytes modeling normal health.
• The proposed framework demonstrated efficacy on industry-sponsored clinical trials data.

Conclusions:

• The novel approach effectively mines heterogeneous, high-dimensional time series clinical trial data.
• The methods facilitate the discovery of significant analyte correlations and enable robust feature selection.
• This work provides a valuable framework for advancing biomedical data analysis and biomarker discovery.