Diagnosis of Sensorineural Hearing Loss with Neural Networks versus Logistic Regression Modeling of Distortion Product Otoacoustic Emissions

We investigated whether modeling with artificial neural networks or logistic regression of distortion product otoacoustic emissions (DPOAE), across diverse frequencies, may achieve an accurate diagnosis of sensorineural hearing loss (SNHL) of cochlear origin. 256 ears (90 with SNHL and 166 with normal hearing) were evaluated with pure-tone audiometry, impedance audiometry, speech audiometry and DPOAE. Ears were split into training (n = 176) and validation (n = 80) sets. Input variables included gender, age, examination time, DPOAE intensity at <i>f</i><sub>2</sub> frequencies 593, 937, 1906, 3812 and 6031 Hz, and respective values corrected for noise levels. In the validation data set, an average network had an area under the receiver operating characteristic curve (AUC) of 0.86 (accuracy 84%). Logistic regressions including all these variables or those selected by backward elimination had AUC values of 0.91 and 0.92, respectively (accuracy 85% both). Eleven of 12 trained networks had better specificity than the backward elimination logistic regression, and the backward elimination logistic regression had a better sensitivity than 11 of the 12 networks. Both modeling approaches correctly identified all ears with sudden hearing loss, congenital hearing loss, head trauma, nuclear jaundice and ototoxicity, and 2–3 of 5 ears with acoustic trauma, but missed 1–3 of 3 ears with Ménière’s disease and 4–6 of 8 ears with abnormal pure-tone thresholds on audiometry which had no accompanying findings. For SNHL exceeding 45 dB HL on a pure-tone threshold, sensitivity was 83% (15/18) by neural networks and 84 or 94% (16/18 or 17/18) by logistic regression. Both neural-network-based analysis and logistic regression modeling of the DPOAE pattern across a range of frequencies offer promising approaches for the objective diagnosis of moderate and severe SNHL.