Learning multimorbidity patterns from electronic health records using Non-negative Matrix Factorisation.

Area of Science:

• Computational epidemiology
• Health informatics
• Biostatistics

Background:

• Multimorbidity, the co-occurrence of multiple diseases, is rising but remains poorly understood.
• Existing research often uses cross-sectional data, small datasets, or lacks robust validation, limiting insights into disease evolution.
• Understanding temporal patterns of multimorbidity is crucial for public health and clinical practice.

Purpose of the Study:

• To develop and validate a novel Non-negative Matrix Factorisation (NMF) approach for temporal phenotyping of multimorbidity.
• To introduce quantitative metrics for evaluating disease clusters and their trajectories over time.
• To demonstrate the utility of temporal disease clusters in mining multimorbidity networks and generating hypotheses.

Main Methods:

• Employed Non-negative Matrix Factorisation (NMF) for simultaneous mining of disease clusters and their temporal trajectories.
• Developed quantitative metrics for cluster and trajectory evaluation.
• Utilized a large-scale electronic health records (EHR) dataset comprising over 2 million patients.

Main Results:

• Successfully implemented a novel temporal phenotyping approach using NMF.
• Generated quantitative metrics for robust evaluation of identified multimorbidity patterns.
• Demonstrated the capability of the model to uncover temporal disease relationships and generate new hypotheses.

Conclusions:

• The novel NMF-based temporal phenotyping method offers a powerful tool for understanding multimorbidity dynamics.
• This approach enhances the analysis of disease co-occurrence and evolution over time.
• Findings provide a foundation for generating new hypotheses regarding the emergence and progression of complex disease patterns.