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

The study was retrospective and came from the First People's Hospital of Foshan, so the model may not perform the same way in other settings. There was no external validation in an independent sample; performance came from repeated training and testing splits within the same dataset. All patients received Ginkgo biloba extract, systemic methylprednisolone, and methylcobalamin, and some also received lidocaine, dexamethasone, papaverine, or intratympanic dexamethasone, so the study cannot separate the effect of any one treatment. Missing data were imputed, and the authors state that if missingness was missing not at random, imputation could be biased. The recovery categories are based on change from intake to discharge audiometry, so intake hearing loss may be partly tied to the outcome; the authors tested this but some circularity may remain. For the all-class task, random forest accuracy was 59.61 ± 2.59%, although its AUC was 0.84 ± 0.01.

Do not read the model performance as evidence that any treatment works or that the model is ready for clinic. All patients received multiple treatments together, and the study did not test whether using the model changes patient outcomes.