Identifying density-based local outliers in medical multivariate circular data

Area of Science:

• Statistics
• Data Science
• Machine Learning

Background:

• Outlier detection is crucial in data analysis.
• Multivariate circular data presents unique challenges due to its bounded range.
• Existing methods may not adequately address circular data properties.

Purpose of the Study:

• To propose and evaluate a novel outlier detection algorithm for multivariate circular data.
• To extend the Local Outlier Factor (LOF) method for circular data.
• To assess the algorithm's performance using simulations and real-world medical data.

Main Methods:

• Extension of the Local Outlier Factor (LOF) algorithm.
• Utilizing two distinct circular distance metrics.
• Incorporating permutation testing for comprehensive analysis.
• Extensive simulation studies to evaluate performance.
• Application to medical datasets (X-ray beam projectors, eye data).

Main Results:

• The algorithm effectively detects outliers in multivariate circular data.
• Performance is positively correlated with the concentration parameter.
• Performance is negatively correlated with sample size.
• Demonstrated utility on medical imaging and biological datasets.

Conclusions:

• The proposed LOF extension is a viable method for outlier detection in multivariate circular data.
• Algorithm performance is sensitive to data concentration and sample size.
• Future work could explore extensions to spherical and cylindrical data types.