Improving data splitting for classification applications in spectrochemical analyses employing a random-mutation...

Area of Science:

• Biomedical data analysis
• Machine learning for spectral data
• Chemometrics

Background:

• Data splitting is crucial for building robust classification models using spectral data in biomedical applications.
• Common data splitting techniques include random selection (RS) and Kennard-Stone (KS) algorithms.
• These methods divide samples into training and testing sets for model construction and validation.

Purpose of the Study:

• To introduce the Morais-Lima-Martin (MLM) algorithm as an improved method for data splitting in classification tasks.
• To evaluate the performance of the MLM algorithm against existing RS and KS methods.

Main Methods:

• The proposed Morais-Lima-Martin (MLM) algorithm modifies the Kennard-Stone (KS) algorithm by incorporating a random-mutation factor.
• Performance comparison involved simulated data and six real-world biospectroscopic datasets.
• Classification models were built using principal component analysis and linear discriminant analysis (PCA-LDA).

Main Results:

• The MLM algorithm demonstrated superior predictive performance compared to RS and KS algorithms.
• MLM resulted in more balanced classification outcomes, particularly in terms of sensitivity and specificity.
• RS yielded the poorest predictive results, while KS showed good accuracy but unbalanced sensitivity and specificity.

Conclusions:

• The Morais-Lima-Martin (MLM) algorithm shows significant potential as an advanced sample selection method for spectral data classification.
• MLM offers enhanced predictive capabilities and balanced performance metrics over traditional data splitting techniques.
• The MLM algorithm is available for MATLAB users.