Predicting depressive symptoms in older adults with hearing impairment development and validation of NHANES-based machine-learning models: A cross-sectional study

Clinical Snapshot

Key Findings

• P Value: Not reported for primary discrimination metric

• Effect Size: AUC 0.862 for logistic regression model

• Primary Outcome: Depressive symptoms (PHQ-9 ≥10) in 154/2123 participants (7.25%)

• Nnt Or Sensitivity: Brier score 0.053 indicating good calibration; sensitivity and specificity vary by threshold selection

• Confidence Interval: 95% CI 0.800-0.911

Clinical Application

High feasibility with simple 6-variable model and available online calculator, though requires audiometric confirmation of hearing loss Relevant to Australian aged care and primary care given similar demographic patterns, though external validation needed. Could inform Medicare-funded mental health assessments and hearing services under Australian Government Hearing Services Program Older adults (≥60 years) with hearing impairment in primary care and audiology settings

Abstract

Depressive symptoms frequently co-occur with hearing impairment in older adults, yet available prediction tools for this high-risk subgroup remain limited in practicality and applicability. We developed and internally validated prediction models for clinically relevant depressive symptoms among older adults with hearing impairment. We analyzed cross-sectional data from 5 National Health and Nutrition Examination Survey cycles (2005-2006, 2009-2010, 2011-2012, 2015-2016, and 2017-2018). Adults aged ≥60 years with hearing impairment were identified using the 2021 World Health Organization criterion (better-ear pure-tone average at 0.5/1/2/4 kHz ≥20 dB). Depressive symptoms were defined as a Patient Health Questionnaire-9 score ≥10. Candidate predictors included sociodemographic factors, comorbidities, lifestyle behaviors, and sensory/functional limitations; missing covariates were handled using multiple imputation. Participants were randomly split into training and validation sets (7:3). Predictors were selected using the Boruta algorithm followed by least absolute shrinkage and selection operator regression, yielding 6 predictors (age, poverty-income ratio, self-rated health, sleep disorder, memory difficulty, and physical functional limitation). Logistic regression and 6 machine-learning algorithms were developed. Model performance was assessed by discrimination (area under the receiver operating characteristic curve [AUC]), calibration (calibration curves and Brier score), and clinical utility (decision-curve analysis). The analytic sample included 2123 participants, of whom 154 (7.25%) had depressive symptoms. In the validation set, AUCs across models ranged from 0.799 to 0.862. Logistic regression achieved the highest discrimination (AUC 0.862, 95% CI 0.800-0.911) with good calibration (Brier score 0.053) and favorable net benefit across clinically relevant thresholds. A nomogram, a dynamic nomogram, and an online calculator were developed to facilitate individualized risk estimation. Sensitivity analyses showed broadly consistent results, except for attenuated discrimination in the severe hearing-loss subgroup. A parsimonious 6-variable model demonstrated strong discrimination, good calibration, and potential clinical utility for risk stratification and targeted screening among older adults with hearing impairment. External validation and prospective evaluation are warranted before clinical implementation.

References

1. 1.Shang, L., Jiang, S., & Dai, Q. (2026). Predicting depressive symptoms in older adults with hearing impairment development and validation of NHANES-based machine-learning models: A cross-sectional study. Medicine, 105(22), e49004. https://doi.org/10.1097/MD.0000000000049004