Applying Machine Learning Models to Predict Occupational Noise-Induced Hearing Loss: A Systematic Review and Meta-analysis.

Abstract

Objectives: To map and critically appraise prediction models for occupational noise-induced hearing loss (ONIHL) using machine learning (ML) approaches and to evaluate reporting quality and risk of bias.

Methods: Seven databases were systematically searched. Reporting quality was assessed using the TRIPOD-AI checklist, and risk of bias and applicability were evaluated using the PROBAST tool. A meta-analysis was performed to evaluate ML model performance for predicting ONIHL using the area under the curve (AUC), with subgroup analysis by different algorithms and ONIHL definitions.

Results: Sixteen studies were included. Support vector machine (SVM) was the most frequently used algorithm, followed by random forest and artificial neural network (ANN). Age, noise level, and exposure duration were the most common input variables, with little consideration of noise type or its complex temporal structure, and most studies used categorical hearing loss status as the outcome. PROBAST assessment revealed 56% (9/16) of studies at high risk of bias, 38% (6/16) at low risk, and 6% (1/16) at unclear risk; applicability concerns were common (31% at high risk). Meta-analysis of 29 ML models yielded a pooled AUC of 0.788 (95% CI: 0.761-0.813; I2  = 93.4%), while six logistic regression models yielded 0.811 (95% CI: 0.743-0.865; I2  = 87.4%). Subgroup analysis showed that XGBoost and ANN algorithms achieved the highest AUC (both 0.846), and the ONIHL definition was a major source of heterogeneity.

Conclusions: ML models demonstrate potential for predicting ONIHL. However, due to high risk of bias, substantial heterogeneity, and lack of external validation, current evidence should be interpreted with caution. More well-designed studies are needed.