Authors: Şeyma Öztürk¹; Halil Berkay Saldırım¹
¹Istanbul Medipol University
Background: Occupational audiometry detects high-frequency threshold shifts but rarely supports individual risk stratification before deterioration is established. We tested whether demographic, occupational and baseline audiometric data predict final high-frequency hearing risk in airport catering personnel.
Method: This retrospective analysis included 512 workers with baseline and final air-conduction thresholds at 0.25–8 kHz in both ears. Final high-frequency risk was defined as bilateral 4–8 kHz PTA ≥30 dB HL. Logistic regression, random forest and gradient boosting classifiers were evaluated with 5-fold stratified cross-validation. We compared a demographic/occupational model (age, sex, employment duration, department) with an expanded model adding baseline low-frequency PTA, high-frequency PTA, high–low frequency gap and high-frequency interaural asymmetry.
Results: Final high-frequency risk was present in 204/512 workers (39.8%), compared with 106/512 (20.7%) at baseline. Mean high-frequency PTA increased from 21.9 to 29.1 dB HL. Demographic/occupational variables alone showed moderate discrimination (best ROC-AUC 0.724). Adding baseline audiometric features markedly improved prediction: the preferred gradient boosting model achieved ROC-AUC 0.878, balanced accuracy 0.772, specificity 0.893, precision 0.801 and F1 0.719. Baseline high-frequency PTA was the dominant predictor, followed by employment duration. Observed risk increased from 4.7% to 89.8% across the lowest to highest predicted-risk quartiles.
Conclusion: Machine-learning models using routine occupational audiometry can stratify airport catering workers by high-frequency hearing risk. Baseline audiometric profile improves performance beyond demographic and job-related variables, supporting a screening tool for prioritizing follow-up audiometry, counselling and hearing-protection interventions. External validation is required before deployment.
