Machine learning can tell people with tinnitus from people without it in research datasets, with pooled accuracy of 86.2% across 24 EEG studies, but no AI tool is in routine clinical use and those figures come from lab conditions.
Tinnitus has no blood marker or single scan that confirms it, which is a large part of why diagnosis and treatment matching have relied on what people report and on clinical judgement. That gap is where AI and machine learning research has concentrated.
Reading brain activity: the EEG approach
One research direction uses electroencephalography (EEG), recordings of the brain's electrical activity, with algorithms trained to separate people with tinnitus from people without it.
A 2026 systematic review and meta-analysis gathered 24 such studies published between January 2010 and January 2026. Across them, pooled accuracy was 86.2%, and the area under the ROC curve, a standard measure of how well a test separates two groups, was 0.878. Models built on deep learning reported higher accuracy (89.0%) than those built on traditional machine learning (83.7%).
Individual studies show the range. A 2023 study recorded EEG from 129 people with tinnitus and 142 without; its classifier was 96% accurate on the data it was trained on and 94% on a held-back test set, and a second classifier separated low from high distress with 84% test accuracy.
The aim is not classification for its own sake but objective biomarkers: measurable brain-activity patterns that could one day add an objective measure to the questionnaires, such as the Tinnitus Handicap Inventory, that assessment relies on now. How tinnitus is measured covers those questionnaires.
Reading hearing test data
A separate direction applies machine learning to hearing tests, including frequencies above the 8 kHz limit of a standard audiogram. A 2024 study built a model from 509 high-frequency audiograms — 242 from people with tinnitus and 267 from people without — and reported 94.06% accuracy, with an area under the curve given as 97.06%. The model was trained on 70% of the audiograms and tested on the remaining 30%.
The suggestion is that some auditory changes linked to tinnitus sit in frequencies a standard test does not check. That is plausible, and it is the same idea behind hidden hearing loss: damage a standard audiogram does not show. The next section is about why the headline figure cannot yet carry more weight than that.
Why a lab accuracy figure is not a diagnostic test
Accuracy figures in this literature describe how well a model sorted one research sample. They do not tell you how often it would be right about a new patient in a clinic, and several features of how the studies are built push the numbers up.
Who is being compared. Most studies set people already diagnosed with tinnitus against people chosen because they do not have it. In the audiometry study, the control group was restricted to people whose audiograms fell within the normal hearing range, while the tinnitus group's audiograms showed abnormalities. A model can score well on that comparison partly by telling hearing loss from normal hearing. A clinic sees everyone in between.
Hearing loss hiding inside the signal. A 2025 study of 110 people with chronic tinnitus asked whether resting-state EEG could tell which ear the tinnitus was in. Most apparent differences turned out to reflect differences in hearing between the ears; after adjusting for hearing thresholds, no EEG feature held up. Separately, the study's machine learning models performed at chance when sorting all four groups (both ears, left, right, not localised) and near chance for two-way comparisons, with balanced accuracy of 57–63%. The authors estimated that around 335 participants per group would be needed to detect effects of the size they saw. The study compared people with tinnitus with each other, not with controls, but the lesson carries: tinnitus and hearing loss travel together, so a model that has not accounted for hearing may be partly detecting hearing instead.
Small samples and many attempts. A 2025 study reported 100% accuracy from EEG, using a main dataset of 36 people (20 with tinnitus, 16 without) and a public dataset of 37. The perfect score was the best of many combinations tried — five frequency bands, four microstate settings, eyes open and eyes closed, and several different classifiers. With a few dozen people and that many options, some combination will fit the sample very well.
Tested on the same source it was built from. The usual check is to hold back part of the dataset — 30% in the audiometry study, 20% in the prediction study below — and test on that. The held-back part comes from the same recruitment, the same equipment and the same population, so it is a check against memorising individual cases, not against a model that only works in the place it was built. The authors of the prediction study state it directly: the model lacks external validation, so its performance in other groups of patients is uncertain.
A pooled figure inherits all of this. The 86.2% is an average of what 24 studies reported, each with its own sample and method, so it carries their designs with it. A 2025 scoping review of 81 resting-state EEG studies found no single definitive EEG marker of tinnitus, and proposed validating candidate measures with machine learning on large, diverse populations as the way forward. That is a description of the work still to be done.
Accuracy also depends on who is being tested. A figure from a sample split roughly half and half between patients and controls says little about how often a positive result would be right in a clinic, where the mix of people is different. None of this makes the research worthless. It is how the early stage of a diagnostic test looks, and it is the same reasoning set out in the patterns that make a weak tinnitus result look strong.
Predicting who is at risk
A related question is whether models can predict how tinnitus will develop, not just whether it is present.
The strongest example so far is observational. A UK Biobank analysis of 192,993 people used health, mood, sleep and hearing information to model tinnitus severity, and that model predicted how tinnitus progressed over nine years. Hearing health was the main risk factor, and mood, neuroticism and sleep predicted severity. The UK Biobank tinnitus study covers the same dataset.
A study published in January 2026 is sometimes described in the same terms, and is different. It used the records of 301 people with hearing loss at one hospital, collected over a single period, and its best model identified who had clinically significant tinnitus with an area under the curve of 0.977 in its validation set. Its authors describe the design as retrospective, single-centre and cross-sectional: it identifies who already has significant tinnitus, and it cannot say who will go on to develop it. Hearing loss severity, age and sleep problems were its strongest predictors.
Why this is not yet part of clinical care
These are research and pilot studies, not tools used in an audiology appointment. The audiometry study's authors wrote that future work should expand dataset diversity and run clinical trials to validate the models' practical utility. Until a model has been tested on new patients, in other clinics, against the outcomes that matter, it is a promising result rather than a test.
What this could mean for tinnitus care
If the research matures, plausible uses include more objective confirmation alongside what patients report, earlier identification of people at risk of severe tinnitus, and better-matched treatment: using measurable brain or hearing patterns to predict which of sound therapy, CBT, TRT and other treatments a given person is more likely to respond to, in place of the trial and error of today. Each of those would need its own validation before it reached a clinic.