TriHSPAM: Triclustering heterogeneous longitudinal clinical data using sequential patterns

Triclustering has become a well-established approach for handling the complexities of three-way data analysis, aiming to uncover patterns with unexpectedly strong coherence across subsets of observations, features, and time-points. In biomedical settings, this technique facilitates the analysis of multivariate physiological signals, omic datasets, and clinical records to discern coherent responses and distinct patient subgroups. Despite its significant potential, existing triclustering algorithms face limitations when dealing with heterogeneous data that include mixed-type features. Key challenges include establishing robust coherence criteria, managing noise and missing data, and effectively addressing temporal complexities. This paper introduces TriHSPAM, a novel triclustering algorithm designed to address these challenges and provide insights into heterogeneous clinical data. A novel merit function is proposed to evaluate heterogeneous triclusters and guide the search process. TriHSPAM demonstrates remarkable flexibility, accommodating diverse data intricacies and capturing complex relationships within data. This algorithm incorporates noise-tolerant techniques to enhance reliability and robustness, effectively handling noisy and sparse real-world data. TriHSPAM adeptly captures meaningful temporal dynamics in three-way heterogeneous data, addressing temporal specificities, such as temporal contiguity and misalignments, while providing statistical significance guarantees. Experimental validation on both synthetic and real datasets confirms TriHSPAM’s efficacy in identifying coherent temporal subspaces within heterogeneous data, thereby guiding knowledge discovery from longitudinal studies. • Novel triclustering algorithm, TriHSPAM, to handle heterogenous clinical data • Solid principles to capture temporal dynamics in three-way data with mixed-type features • Discovery of heterogeneous triclusters with guarantees of statistical significance and robustness to noise and missing values.