Authors: Prasad Darveshi¹; Afagh Farhadi¹; Samantha Hauser²; Andrew Sivaprakasam³; Michael Heinz¹
¹Purdue University; ²University of Pittsburgh; ³Indiana University School of Medicine
Background: Standard Auditory Brainstem Response (ABR) interpretation focuses on threshold estimation, often overlooking suprathreshold waveform changes indicative of hidden pathologies like cochlear synaptopathy and inner-hair-cell (IHC) dysfunction. This study investigated whether deep learning applied to ABR “waterfall plots” can detect morphology-based signatures of these pathologies in a chinchilla model.
Method: ABR datasets were collected from three groups: normal-hearing baseline, noise-induced temporary-threshold-shift (TTS/synaptopathy), and carboplatin-induced IHC dysfunction. Level-stacked ABR traces were converted into standardized waterfall images to preserve timing and inter-wave morphology. A two-stage pipeline using a fine-tuned EfficientNet-B0 was employed: Stage 1 classified ABRs as normal or abnormal; Stage 2 differentiated between TTS/synaptopathy and carboplatin/IHC dysfunction.
Results: The deep learning model successfully extracted diagnostic information from waveform morphology independent of threshold changes. Stage 1 achieved ~91.5% accuracy in detecting abnormal patterns. Stage 2 differentiated between the two types of hidden pathology with ~83% accuracy. Errors primarily occurred between the two abnormal groups, likely due to overlapping morphological features of neural and sensory damage.
Conclusion: A lightweight, morphology-aware deep-learning pipeline can identify subtle ABR differences associated with specific types of hidden cochlear pathology. These findings suggest that automated morphology analysis can effectively complement threshold-based interpretation, offering a potential path toward individualized diagnostic tools for identifying subclinical auditory disorders in human patients.
