Machine Learning Flags Anxiety in Sudden Hearing Loss Patients, Study Finds

More than six in ten adults hospitalized with a first episode of sudden sensorineural hearing loss met the study’s threshold for anxiety symptoms, and a machine-learning model built to flag those patients before discharge showed moderate—but imperfect—predictive performance in its first temporal test. The findings, from a single-center retrospective prediction-model study at Beijing Anzhen Hospital in China, were published in Frontiers in Public Health.

The study included 512 adults treated between February 2022 and August 2023. Anxiety symptoms were defined using the Hamilton Anxiety Rating Scale, or HAM-A, with a score of 7 or higher marking the outcome of interest. Overall, 320 of the 512 patients—62.5 percent—met that criterion, underscoring how common psychological distress is during acute care for this condition. The researchers split 410 patients into a development cohort, further divided once into a training set of 287 and an internal model-selection set of 123, while 102 later patients formed a same-center temporal-validation cohort that was kept entirely out of feature selection, preprocessing, tuning, and algorithm choice.

Fourteen candidate predictors entered three feature-selection methods—LASSO, Boruta, and recursive feature elimination—and eight classifiers were compared. Six predictors survived into the primary analysis: hearing-loss severity, tinnitus, vertigo, sleep status, number of comorbidities, and treatment efficacy. The selected support vector machine achieved a ROC-AUC of 0.772 internally and 0.763 in the temporal validation, a level of discrimination the authors characterize as moderate.

The more sobering results came from the robustness analyses. When the researchers re-ran the entire pipeline through repeated nested cross-validation on all 410 development observations, the pooled ROC-AUC was 0.747 (95 percent CI 0.696–0.792), the PR-AUC was 0.822 (0.768–0.867), and the Brier score was 0.208 (0.196–0.221). Calibration was imperfect: calibration-in-the-large was 0.451 and the calibration slope 1.570, meaning predicted probabilities did not align cleanly with observed rates.

Algorithm choice itself proved unstable. In the nested validation, XGBoost was selected in 8 of 10 outer evaluations and random forest in the other 2; the support vector machine chosen in the primary workflow was never selected. Four predictors—hearing-loss severity, sleep status, number of comorbidities, and treatment efficacy—were retained in all outer evaluations, suggesting these variables are the most stable signals, while the specific classifier matters less than the primary analysis implied.

Sensitivity analyses probed the fragility of the outcome definition. Raising the HAM-A threshold from 7 to 8 reduced prevalence to 57.0 percent and lowered the temporal SVM ROC-AUC from 0.759 to 0.730. The team also tested overlapping-symptom exclusion models, logistic comparators, corrected calibration, a development-only transported threshold, and continuous HAM-A severity as alternative framings.

The authors are explicit about the limits of their work. They conclude that the findings support an investigational model-development framework rather than superiority of a specific SVM or an implementation-ready rule. Because the study was retrospective and single-center, geographic multicenter validation, a prospectively locked model and threshold, and prospective impact assessment are required before any clinical use could be considered.

For clinicians, the practical takeaway is less about deploying an algorithm tomorrow and more about the epidemiology: anxiety symptoms were highly prevalent in this hospitalized SSNHL population, and a small set of routinely collected clinical variables—hearing-loss severity, sleep disturbance, comorbidity count, and treatment response—consistently carried the predictive signal across validation schemes. That consistency may guide screening attention in acute care even as the modeling itself remains experimental.

Subject of Research: Machine-learning prediction of anxiety symptoms in adults hospitalized with sudden sensorineural hearing loss

Article Title: Development and temporal validation of an explainable machine-learning model for identifying anxiety symptoms in adults with sudden sensorineural hearing loss

Article References: Yu, M., Yang, F., & Yu, X. (2026). Development and temporal validation of an explainable machine-learning model for identifying anxiety symptoms in adults with sudden sensorineural hearing loss. Frontiers in Public Health, 14, Article 2003468. https://doi.org/10.3389/fpubh.2026.2003468

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.2003468

Keywords: anxiety symptoms, sudden sensorineural hearing loss, machine learning, support vector machine, temporal validation, HAM-A, nested cross-validation, prediction model, tinnitus, calibration, Development, temporal