Machine-Learning Models With Multiple Imputation With Sequential Nearest Neighbors Imputation for Predicting the Prognosis of Idiopathic Sudden Sensorineural Hearing Loss Patients.

objectivesThe prognosis of idiopathic sudden sensorineural hearing loss (ISSNHL) is heterogeneous. The study aimed to investigate the prognostic factors of ISSNHL and develop machine-learning models with multiple imputation to predict the prognosis of ISSNHL.

designA retrospective study was undertaken on a cohort of 600 patients with ISSNHL who underwent standardized treatment protocols. Clinical features, blood tests, concurrent symptoms, body measures, and audiometric features were collected. Missing values were imputed by multiple imputation with the sequential nearest neighbors algorithm. Six classifiers and four classification tasks were explored. Model performance was evaluated by accuracy and area under the receiver operating characteristic curve. Furthermore, a feature importance analysis was conducted to enhance model interpretability and streamline complexity.

resultsStatistically significant differences were observed across prognosis groups in terms of age, sex, days from onset to treatment, mean hearing threshold, vertigo, ear blockage, hearing curve types, loudness recruitment, auditory brainstem response, World Health Organization classification, distortion product evoked otoacoustic emission response, fibrinogen, cholesterol, tinnitus, hypertension, diabetes, and history of hearing loss. Notably, three machine-learning classifiers demonstrated robust performance across all classification tasks. The feature importance analysis illuminated the most pivotal prognostic factors for each classification model. In addition, the area under the receiver operating characteristic curve remained stable even after excluding the 14 to 24 least influential features from the random forest classifiers, facilitating the clinical practice of these models.

conclusionsIn addition to several factors (age, time from onset to treatment, vertigo, white blood cells, platelets, and fibrinogen) that have been previously reported, this study identified some novel clinical parameters as significant contributors to ISSNHL prognosis prediction, including distortion product evoked otoacoustic emission response, auditory brainstem response wave V interaural latency difference, mean hearing threshold of the contralateral ear, and body mass index. We strongly encourage further validation and expansion of our study's results by healthcare researchers, aiming to expedite their clinical application and improve patient outcomes.