A major study using AI and population data uncovers how hearing problems and psychological traits combine to drive tinnitus severity, paving the way for early screening and targeted support.
Study: Tinnitus risk factors and its evolution over time. Image Credit: Prostock-studio / Shutterstock
In a recent study in the journal Nature Communications, researchers in Canada and France leveraged machine learning models, specifically applying established algorithms, to investigate the presence (frequency) and severity of subjective tinnitus. The study used data from the UK Biobank (n = 192,993, tinnitus = 41,042) and revealed hearing health to be the variable most strongly associated with symptom presence and severity risk, while mood, sleep status, and neuroticism were significant predictors of tinnitus severity.
While the presence model couldn't predict future presence risk, the severity model showed good to excellent performance in predicting changes in tinnitus severity over time, particularly for individuals developing severe distress. Researchers validated the model on a cohort of volunteers from the Tinnitus Research Initiative database (n = 463) and further simplified the severity models to facilitate rapid deployment and screening for patients with severe tinnitus.
Background
Subjective tinnitus is an auditory condition in which individuals perceive buzzing, hissing, or other sound cues without the presence of an external acoustic stimulus. It is a relatively common condition, estimated to affect ~14% of all adult humans. While not bothersome to most individuals, subjective tinnitus can cause severe distress, cognitive difficulties, and socio-economic disturbances in some patients.
"As there is no cure to eliminate tinnitus perception—only palliative interventions aiming at reducing associated distress—enhancing prevention and clinical management through better identification of key risk factors is essential."
Several studies have attempted to elucidate the factors underlying tinnitus pathophysiology, but findings remain conflicting. So far, sensory deafferentation caused by excessive noise exposure, trauma, presbycusis, and ototoxic medication is assumed to be the primary cause of the condition, but a longitudinal examination of tinnitus’s risk factors remains lacking.
About the study
The present study aims to address these gaps in the literature by identifying risk factors capable of predicting the onset and evolution of subjective tinnitus. Study data were obtained from UK Biobank (UKB) participants between the ages of 40 and 69 years. Prospective participants were subjected to a baseline screening (V1, collected between April 2009 and November 2021) and a main follow-up visit (V2, collected between August 2012 and February 2023).
Tinnitus presence was evaluated using participant-completed questionnaires. Participants reporting tinnitus symptoms were subjected to additional questions to assess the severity of the condition.
For machine learning (ML) model development, features were selected a priori by consensus among the authors, based on their established relevance in the literature and comprised 101 features, including items relevant to hearing health, mood, physical health, and sociodemographics. The nonlinear iterative partial least squares (NIPALS) regression algorithm was used for model training due to its benefits in processing high-dimensional clinical datasets (notably its ability to manage multicollinearity and provide interpretable components).
Features were (risk) scored using both presence and severity as outcomes of interest. The final training dataset comprised 166,119 items for the presence model and 35,942 for the severity model. Adjusted risk scores were computed to examine tinnitus evolution independently of baseline status. Model performance was evaluated using explained variance (R²), AUC-ROCs, and Cohen's d effect sizes.
Study findings
Study findings identified 41,042 (21.3%) of UKB participants with tinnitus, 22.7% of whom experienced moderate to severe distress. Men (particularly older men) were demonstrated to experience the condition most frequently, while women experienced higher severity.
"Tinnitus presence and severity were associated with larger hearing deficits, be it self-reported hearing difficulties, speech-in-noise hearing difficulties self-reported and measured, or use of hearing aid or cochlear implants."
NIPALS risk score calculation revealed hearing health to be the strongest predictor of model performance, accurately differentiating patients from non-patients. Notably, presence risk scores failed to predict tinnitus progression over time, with AUC values below 0.60 and small effect sizes, suggesting that general health, environmental, and sociodemographic factors are not ideal traits in tinnitus investigations.
In contrast, severity risk scores showed better prognostic value, with ROC AUCs as high as 0.81 for predicting worsening severity, supporting the idea that simultaneously studying multiple traits (mood, sleep, hearing, and neuroticism) can elucidate the multifactorial nature of the condition.
The simplified model (called POST – 'Prediction Of the Severity of Tinnitus') also demonstrated moderate to excellent performance, especially in identifying individuals at high risk of developing severe distress, with a validation AUC of 0.94 in a subset of TRI participants at risk of progressing from non-distressing to severely distressing tinnitus, contingent on sufficient mood, sleep, hearing, and neuroticism data (6 items).
Conclusions
The present study identifies risk predictors of subjective tinnitus, using machine learning models to predict long-term severity outcomes but not presence trajectories. Hearing health was recognized as the most important predictor of both symptom presence and severity, highlighting its importance in future anti-tinnitus interventions. Meanwhile, mood, sleep, neuroticism, and hearing emerged as key modifiable factors influencing how tinnitus is experienced, offering actionable targets for clinical intervention.
The POST tool may assist clinicians in primary care and specialist settings in triaging patients most at risk for long-term severe tinnitus, helping to optimize resource allocation and guide early support.
Study limitations include the UK Biobank’s lack of ethnic diversity (91% White participants), exclusion of hyperacusis assessment, and the limited explained variance of the models (~12.5% for presence and ~9.2% for severity), which indicate the need for additional biological and environmental predictors in future research.