Comparing an AI-driven Hearing Test With the Standard Hearing Test in Adults With Hearing Loss and Tinnitus

Pure-tone audiometry is the reference standard for measuring hearing thresholds, and the procedure recommended by the British Society of Audiology has been essentially unchanged for decades. It is reliable, but it is also slow: the level of each tone is chosen by a fixed rule rather than by what the previous answers have already established, so a substantial proportion of the presentations in a test carry little information about where the threshold lies. The cost falls on both the patients' and the audiologists' time, which is the binding constraint on access to hearing assessment in the NHS.

AI-PTA addresses this by replacing the fixed rule for choosing the next level. The procedure maintains a probability distribution over the threshold at each ear and frequency, updates it after every response using a psychometric model of hearing, and selects the level of the next tone to maximise the expected information gained about the threshold. Testing at each frequency stops when the standard deviation of that distribution falls below 4 dB and the participant has responded on an ascending run. The starting distribution is drawn from one of four sources, chosen before the test: the participant's own most recently measured audiogram, a clinic population of 75,452 adults seen at this trust, a general population of 19,051 respondents from NHANES (Centers for Disease Control and Prevention (CDC). National Center for Health Statistics (NCHS). National Health and Nutrition Examination Survey Data), or a uniform distribution carrying no assumption. Only aggregate tables are used; no individual record from either population is held by the system.

The stopping threshold of 4 dB was fixed before recruitment on the basis of in-silico evaluation across simulated listeners of six audiogram shapes, and is not adjusted during the study. Changing it, or any other parameter capable of moving a measured threshold, would make sessions recorded before and after the change non-comparable, so any such change requires the participants already tested to be treated as a pilot set.

Both procedures are delivered by an audiologist who presents each tone on the department's own calibrated audiometer and records the response in the software. The software instructs and records; it does not generate sound, and it does not decide a threshold that the audiologist cannot override. This is deliberate. The system is a decision support tool operating with a clinician in the loop, not an autonomous test, and every departure from the procedure it proposes is recorded with a reason so that the frequency of such departures can be reported.

Safety constraints are applied identically in both procedures and take precedence over the level selection. Above 80 dB HL the level rises in steps of no more than 5 dB, and no step may cross 80 dB HL without stopping on it, because loudness growth is not linear at high levels and an information calculation that proposes a large step at the top of the range would be correct about the information and wrong about the participant. Maximum output is capped per frequency at the audiometer's calibrated limit, and a clinician may enter lower limits for the equipment actually in the room. The first tone of every session is a familiarisation tone at 40 dB HL whatever the starting distribution suggests, and is excluded from the threshold estimate.

Masking, where the rules indicate it, follows the plateau method of the British Society of Audiology in both procedures. The three indication rules, the 10 dB steps of masking noise and the criterion for a plateau are identical; what differs is only how the threshold is re-established at each level of noise, which is the same difference the study is measuring everywhere else. Bone conduction is likewise handled identically in both.

The design is within-participant and counterbalanced. Each participant receives both procedures, so each acts as their own control and the comparison is not vulnerable to between-participant variation in hearing. The order is randomised to prevent an effect of fatigue, learning or familiarity being attributed to either procedure, and randomisation is stratified by the presence or absence of concurrent tinnitus because tinnitus is expected to affect how reliably a participant responds near threshold, and an imbalance between the groups would otherwise be indistinguishable from a difference between the procedures.

Every presentation is recorded: the level, the response, the state of the estimate at that point, and the time between the instruction appearing and the answer being recorded. Each session also records the version of the procedure that produced it, the version of the software, and the specific build, so that any recorded threshold can be interpreted against the behaviour that produced it. These records support the error analysis, the assessment of agreement between the procedures, and the examination of whether performance differs across age, sex and tinnitus status.

The study is an early-stage clinical evaluation. It is powered to compare the two procedures within participants and is not designed to establish clinical outcomes, cost-effectiveness, or performance in settings or populations other than the one described.