Bleeding predisposition assessments in tonsillectomy/adenoidectomy patients using fuzzy interquartile encoded neural...

Area of Science:

• Machine Learning
• Data Preprocessing
• Biomedical Informatics

Background:

• Supervised feed-forward neural networks require effective feature preprocessing for optimal performance.
• Feature outliers and non-normalized feature spaces can degrade classifier accuracy.
• Existing methods may not adequately handle complex data distributions or improve discriminatory power.

Purpose of the Study:

• To introduce and evaluate fuzzy interquartile encoding (FIE) as a preprocessing strategy for neural networks.
• To assess FIE's impact on feature space normalization and robustness to outliers.
• To demonstrate FIE's effectiveness in improving classification accuracy on synthetic and real-world biomedical data.

Main Methods:

• Fuzzy set theory is applied to determine feature membership degrees within overlapping fuzzy sets at quartile boundaries.
• Original features are replaced by their membership values, creating a transformed feature space.
• The methodology is tested on synthetic datasets with varying distributions and two clinical datasets related to bleeding tendency.

Main Results:

• Fuzzy interquartile encoding consistently enhances the discriminatory power of classifiers across synthetic datasets.
• Application to a dataset of coagulation test results for tonsillectomy/adenoidectomy patients yielded an 11% improvement in classification accuracy.
• Analysis of a bleeding tendency questionnaire dataset showed an 18% increase in classification accuracy.

Conclusions:

• Fuzzy interquartile encoding is an effective preprocessing technique for supervised neural networks.
• FIE offers a normalizing effect on feature spaces and increased robustness to outliers.
• The method shows significant potential for improving diagnostic accuracy in clinical applications, such as predicting surgical bleeding risks.