Random vector functional link with ε-insensitive Huber loss function for biomedical data classification.

Area of Science:

• Biomedical data analysis
• Machine learning
• Data science

Background:

• Biomedical data classification is crucial but challenged by noisy datasets.
• Conventional machine learning models struggle with data noise.
• Random Vector Functional Link (RVFL) models show promise but degrade with noise.

Purpose of the Study:

• To enhance the classification performance of RVFL models on noisy biomedical datasets.
• To introduce a novel RVFL model incorporating an ε-insensitive Huber loss function (ε-HRVFL).
• To evaluate the effectiveness of ε-HRVFL in handling feature noise in biomedical data.

Main Methods:

• The ε-HRVFL model's optimization problem was reformulated as a strongly convex minimization problem.
• Experiments were conducted on biomedical and non-biomedical datasets with varying label noise.
• Statistical comparison using Friedman test against Support Vector Machine and Extreme Learning Machine models.

Main Results:

• The proposed ε-HRVFL model achieved 98.1332% accuracy on non-biomedical datasets.
• The ε-HRVFL model demonstrated a best accuracy of 96.5229% on biomedical datasets.
• The ε-HRVFL model with a sigmoid activation function showed superior performance across all tested datasets.

Conclusions:

• The proposed ε-HRVFL model is effective for biomedical data classification, especially in the presence of noise.
• Future work can extend ε-HRVFL to multiclass problems and explore asymmetric Huber loss functions.
• The study highlights the potential of ε-HRVFL for robust biomedical data analysis.