Incerti, PV1, Kitterick PT1, Monaghan JM1, Zakis J1, Gavrilis J1, Croteau M1, Mishra D1, Boothalingam S1, Kwok C1, Rutledge KL1, Gilliver M1, Weiss M1, Milner S1, Flynn C1, Nakad P1, Chesnaye M1, Brewer S1,2, Ozonuk R1, Zhang VW1, Nguyen G1, Meng Q1, Oliver J1, , Edwards B1

1National Acoustic Laboratories, Sydney, Australia

2Hearing Australia, Sydney, Australia

Background. Advances in artificial intelligence (AI) and machine learning offer significant opportunities to improve the clinical relevance and personalisation of hearing aid fitting prescriptions.

Objective: To describe how machine-learning–based optimisation methods were applied in the development of the NAL-NL3 combined with large-scale real-world clinical fitting data and clinician insights was used to address clinician identified limitations identified in current prescriptive practice.

Methods NAL-NL3 Prescription uses a hybrid optimisation approach that combines broad exploration of possible gain settings with targeted refinement to arrive at clinically acceptable fittings. Global optimisation methods, including Differential Evolution, are used to explore a wide range of candidate gain solutions across different audiometric configurations, including high-frequency gain, reverse-slope losses, and mixed or conductive hearing losses and complex losses. These candidates are subsequently refined using local optimisation approaches, including gradient-based solvers and reinforcement learning–based refinement, to converge on solutions that maximise predicted speech intelligibility while remaining clinically and perceptually acceptable in terms of loudness, comfort, and compression. Machine-learning development was informed by large-scale real-world clinical datasets comprising over one million audiograms and real-ear measurements. Objective modelling assessed predicted speech intelligibility, loudness, sharpness, and compression behaviour, and clinician usability was evaluated through in-clinic fitting trials. Clinicians reported improved ease of matching targets and reduced need for post-fit fine-tuning.

Results Relative to NAL-NL2, the NAL-NL3 Prescription demonstrated small but consistent improvements in predicted speech intelligibility, minimal changes in overall loudness, and modest reductions in predicted sharpness. Gain was redistributed from low- and high-frequency extremes toward mid-frequency regions (approximately 1–4 kHz) critical for speech understanding, while compression ratios remained within predefined limits. Clinicians reported improved ease of matching targets and reduced need for post-fit fine-tuning. Compared with NAL-NL2, the NAL-NL3 Prescription produced small but consistent improvements in predicted speech intelligibility, with minimal changes in overall loudness and modest reductions in predicted sharpness. Prescribed gain was redistributed away from low- and high-frequency extremes toward mid-frequency regions most critical for speech understanding. Compression ratios were maintained within predefined design limits.

Conclusion NAL-NL3 illustrates how AI-driven optimisation and deep learning can be translated into clinically meaningful advances in evidence-based hearing aid prescription. This work demonstrates the potential of AI methodologies to enhance robustness, efficiency, and personalisation in real-world hearing care.