PP1010 - Detecting Human Frequency-Following Responses Using an Artificial Neural Network

Abstract:
This study explores the use of artificial neural networks (ANNs) to detect frequency-following responses (FFRs) elicited by speech intonation, offering a novel approach to enhancing the analysis of neural encoding in audiology. By identifying optimal configurations of inputs, hidden neurons, and sweeps, the ANN achieved high prediction accuracy, demonstrating its potential for more efficient FFR detection compared to traditional methods. These findings provide a foundation for applying deep learning models in auditory processing assessments, with significant implications for improving diagnostic accuracy and patient outcomes in audiology practice.

Summary:
Objectives
The primary objective of this study was to evaluate the efficacy of a three-layer artificial neural network (ANN) in detecting frequency-following responses (FFRs) elicited by the English vowel /i/ with a rising intonation. The study aimed to systematically explore the influence of model architecture, input parameters, and the number of sweeps on the ANN’s prediction accuracy. Additionally, the research sought to establish optimal configurations that could be applied in clinical diagnostics and auditory processing assessments.

Rationale
FFRs are neural signals that encode acoustic features like speech intonation, providing insight into the auditory system's neural encoding of sound. While traditional machine learning methods have been employed to classify FFRs, the potential of deep learning, particularly neural networks, in this domain has not been fully explored. Given the success of deep learning models in other fields, this study addresses the gap by investigating the utility of an ANN in enhancing FFR detection, which has potential clinical implications for improving diagnostic accuracy in patients with auditory processing disorders.

Design
In this study, FFR recordings were collected from 60 normal-hearing adults in response to a rising intonation of the English vowel /i/. These recordings were used to train and test a three-layer ANN model. The input data consisted of F0 estimates derived from the spectral domain, which reflect the pitch-related features of the FFRs. To assess the model's performance, key parameters were systematically varied: the number of inputs (ranging from 1 to 16), hidden neurons (ranging from 1 to 16), and the number of sweeps (from 1 to 7000 sweeps) included in the recordings. The model’s prediction accuracy was analyzed under these varying conditions to identify optimal configurations for accurate FFR classification.

Results
The ANN’s prediction accuracy was influenced by the combination of inputs, hidden neurons, and sweeps. The optimal configuration involved using 6–8 inputs and 4–6 hidden neurons, achieving an accuracy rate of approximately 84%. Notably, improving the signal-to-noise ratio by including 100 or more sweeps enhanced the model's performance. Beyond these optimal ranges, further increasing the number of inputs or hidden neurons contributed little to the accuracy improvement. The model’s performance plateaued after reaching these configurations. These results indicate that the ANN’s performance is on par with, and in some cases surpasses, traditional machine learning approaches for FFR detection.

Brief Summary of Clinical Takeaways:
The findings demonstrate the effectiveness of using an ANN for detecting FFRs in response to speech intonation, providing a novel approach to FFR analysis in auditory research. The study highlights the importance of optimizing input parameters and hidden neuron counts to achieve accurate FFR classification. These results form a foundation for future studies investigating the broader application of deep learning in auditory processing assessments. The ANN-based approach could ultimately enhance clinical diagnostics, offering audiologists an advanced tool for evaluating auditory processing and speech perception in patients, particularly those with hearing or auditory processing disorders. Moreover, this research opens avenues for developing real-time neurofeedback systems and further improving diagnostic methodologies in audiology.