Evolutionary Multiobjective Clustering and Its Applications to Patient Stratification.

Area of Science:

• Computational Biology
• Medical Informatics
• Machine Learning

Background:

• Patient stratification is crucial for personalized medicine and effective disease treatment.
• Existing clustering algorithms face challenges with noisy data, high dimensionality, and limited interpretability.
• Current methods often rely on single internal evaluation functions, which are not robust across diverse datasets.

Purpose of the Study:

• To propose a novel multiobjective framework for patient stratification using clustering.
• To address limitations of existing algorithms, including feature selection and density evaluation.
• To enhance the discovery of disease subtypes for improved treatment strategies.

Main Methods:

• Developed a multiobjective clustering algorithm by fast search and find of density peaks (MO-Clu-FSFD).
• Employed a parameter candidate population evolved under multiple objectives for automatic feature selection and density evaluation.
• Utilized five cluster validity indices (compactness, separation, Calinski-Harabasz, Davies-Bouldin, Dunn) as objective functions.
• Applied a multiobjective differential evolution algorithm based on decomposition for simultaneous optimization.

Main Results:

• The proposed MO-Clu-FSFD algorithm demonstrated superior or competitive performance against 45 other algorithms across 94 diverse datasets.
• Experiments included real patient stratification, synthetic, and medical datasets, validating the algorithm's effectiveness.
• Analysis confirmed the algorithm's robustness, including time complexity, convergence, and parameter sensitivity.

Conclusions:

• The novel multiobjective framework offers a robust and effective solution for patient stratification and disease subtype discovery.
• This approach enhances the capabilities of clustering algorithms in medical research and personalized medicine.
• The framework's ability to handle multiple objectives provides more reliable clustering quality assessment.