Multiple neural network response variability as a predictor of neural network accuracy for chromosome recognition

Area of Science:

• Genetics and Bioinformatics
• Computational Biology
• Machine Learning in Medicine

Background:

• Accurate human chromosome classification is crucial for genetic diagnostics.
• Neural networks show promise for automated chromosome classification.
• Variability in neural network training may indicate classification reliability.

Purpose of the Study:

• To investigate if learning variability from multiple neural networks can predict classification performance.
• To assess the utility of error variability as a confidence indicator for chromosome classification.

Main Methods:

• Trained 100 backpropagation neural networks on the Copenhagen chromosome data bank (8106 chromosomes).
• Calculated an error variability score for each chromosome based on network outputs.
• Analyzed the correlation between error variability and classification accuracy of an optimal network.

Main Results:

• A significant difference (p < 0.0001) was found in variability scores between correctly and incorrectly classified chromosomes.
• Using error variability as a threshold achieved a peak classification rate of 98.93% for scores < 0.35.
• Error variability scores ranged from 0.16 to 1.31 (mean 0.41, SD 0.12).

Conclusions:

• The error variability of multiple neural network responses serves as a reliable confidence indicator.
• This method can enhance the accuracy and reliability of automated human chromosome classification.
• Potential applications in genetic diagnostics and cytogenetics research.