Speech-Based Concussion Detection Using CNNs: A Comparative Study of Acoustic Feature Representations for Digital Health Monitoring
Conference
Rubaiat, R, Poellabauer, C. (2026). Speech-Based Concussion Detection Using CNNs: A Comparative Study of Acoustic Feature Representations for Digital Health Monitoring
. 458-463. 10.1109/CHASE69719.2026.00078
Rubaiat, R, Poellabauer, C. (2026). Speech-Based Concussion Detection Using CNNs: A Comparative Study of Acoustic Feature Representations for Digital Health Monitoring
. 458-463. 10.1109/CHASE69719.2026.00078
Speech-based analysis has emerged as a promising non-invasive approach for detecting neurological impairments such as concussion, which often presents with subtle and difficult-to-measure symptoms. In this study, we investigate the effectiveness of convolutional neural networks (CNNs) trained on multiple acoustic feature representations for concussion detection. Speech recordings from high school and college athletes performing a structured neurological speech task were transformed into diverse feature sets, including mel spectrograms, Mel-frequency cepstral coefficients (MFCCs), chroma features, spectral contrast, zero-crossing rate, and RMS energy. CNN models were trained separately on each representation and evaluated using accuracy, precision, recall, and F1-score.The results show that mel spectrogram-based models consistently outperform other feature representations, achieving the highest classification accuracy (85%) and balanced performance across evaluation metrics. Spectral contrast and RMS energy also produced comparatively strong results, while MFCCs and other spectral and temporal features yielded moderate but consistent performance, indicating that multiple acoustic properties contain information relevant to concussion detection. Limited data augmentation further improved performance in selected cases, suggesting the value of introducing controlled variability during training.These findings highlight the importance of feature representation in speech-based concussion assessment and support the feasibility of CNN-driven models as scalable tools for automated concussion screening. This work contributes to the development of non-invasive, accessible, and data-driven approaches for early detection and monitoring of concussion. Because the current evaluation is based on a limited dataset, a single speech task, and sample-level splitting, the findings should be interpreted as preliminary evidence for feature-representation effectiveness rather than as validation of a clinically deployable concussion detection system.
