Automated Loudness Growth Prediction From EEG Signals Using Autoencoder and Multi-Target Regression

• Published in: IEEE access Vol. 13; pp. 106561 - 106572
• Main Authors: Rama Harshita, D., Tiwari, Nitya, Padole, Himanshu, Nataraj, K. S.
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Ablation; Accuracy; Artificial neural networks; Auditory brainstem response; Autoencoders; Brain modeling; Configuration management; Convolution; convolutional autoencoder; Convolutional neural networks; Correlation coefficients; Electroencephalography; Estimation; Feature extraction; Hearing aids; Impact analysis; Loudness; loudness growth prediction; multi-target regression; Psychoacoustic models; Psychoacoustics; Regression models
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3580833

Accurately assessing loudness perception is crucial for optimizing hearing aid fittings, especially for individuals who are unable to perform subjective tests. This study presents an automated method for estimating frequency-specific loudness growth curves using tone-burst auditory brainstem responses (ABRs), which are subsets of EEG (electroencephalography) signals. Unlike traditional methods that rely on manually engineered features, the proposed method uses a convolutional autoencoder to learn latent representations of ABR signals, reducing dimensionality while preserving critical auditory information. The extracted features are mapped to psychoacoustic loudness growth estimates using a multi-target regression model based on a convolutional neural network. An ablation study was conducted to analyze the impact of different autoencoder configurations on feature extraction performance. The results demonstrate strong predictive consistency, with high Pearson correlation coefficients (PCC <inline-formula> <tex-math notation="LaTeX">\geq 0.9 </tex-math></inline-formula>) and low mean square errors (MSE <inline-formula> <tex-math notation="LaTeX">\leq 0.0011 </tex-math></inline-formula>) across different stimulus frequencies and subjects.