Authors: Manuel Lazo-Maestre¹; MªAmparo Callejon-Leblic²
¹Hospital Universitario Virgen Macarena (HUVM); ²Hospital Universitario Virgen Macarena – Universidad de Sevilla (HUVM-US)
Background: The analysis of cortical auditory evoked potentials (CAEP) often requires expert interpretation to identify key components such as P1, N1, and P2. This work presents the development and evaluation of a MATLAB-based software system designed to assist in the analysis, feature extraction, and automated classification of these signals.
Method: A total of 742 recordings from clinical equipment were used to evaluate 33 machine learning algorithms, including support vector machines (SVM), ensemble methods, and neural networks. Signals were labeled by experts as “clear,” “inconclusive,” or “absent.” The system extracted latencies, amplitudes, signal-to-noise ratio (SNR), and correlation coefficients (CC). The dataset was split into training (80%), validation (10%), and testing (10%) subsets, with performance evaluated via confusion matrices and ROC analysis.
Results: Initial models showed accuracies above 80%, though a bias toward the “clear response” class was observed. Partial dependence analysis highlighted the critical importance of N1-P2 and N1 amplitudes for classification. By incorporating additional features such as SNR and CC, the accuracy improved significantly, with ensemble Bagged Trees models reaching up to 89% on the test set.
Conclusion: These results demonstrate that adding objective features like SNR and CC, combined with increasing dataset size, enhances the classification of cortical auditory responses, especially in ambiguous cases. The system provides a robust tool for automated clinical analysis, supporting more consistent physiological interpretation in audiological diagnostics.
