Predicting cognitive performance from three-electrode auditory brainstem responses using convolutional neural networks

Abstract

Age-related cognitive decline is a growing global concern, motivating the search for noninvasive, accessible biomarkers to support early detection and monitoring. Click-evoked auditory brainstem responses (ABRs), collected in routine clinical settings, offer a promising signal source. Building on prior evidence that ABRs relate to cognitive function, this study investigates whether raw human ABR waveforms can predict cognitive performance using deep learning without manual peak measurements. We used a dataset from 118 adults spanning a broad range of cognitive abilities, pairing each click-evoked ABR with a cognitive score. A 1-D convolutional neural network (CNN) was trained to learn time-series patterns directly from the raw signal. Model performance was evaluated using two, five, and ten-fold cross-validation and compared against traditional wave V metrics and a randomized-input baseline. After adjusting scores for age, the CNN achieved a mean area under the receiver operating characteristic curve of 0.77 ± 0.06, outperforming all benchmarks. To interpret model decisions, gradient-weighted class activation mapping (Grad-CAM) was applied. Three key latency windows were identified: 1.8–2.3, 3.2–4.0, and 4.9–6.5 ms. These correspond to canonical ABR waves I, III, and V, supporting the physiological relevance of the learned features and highlighting Wave III as a previously underutilized marker. Although limited to click stimuli from a single recording system, this work demonstrates that a CNN can extract meaningful features from raw ABRs collected using only three electrodes and predict cognitive status more accurately than traditional methods using the same dataset.

More information

Identifiers

ORCID for Yasmeen Hamza: orcid.org/0000-0003-3294-6629

Catalogue record

Export record

Altmetrics