This article originally appeared in the Fall 2025 issue of On The Level.
Vestibular disorders affect the inner ear and brain areas that control balance and spatial orientation. They can cause dizziness, vertigo, unsteadiness, and other symptoms that can be confusing and overwhelming.
These conditions are notoriously hard to diagnose because symptoms often overlap between different disorders. For example, someone with Menière’s disease, vestibular migraine, or persistent postural-perceptual dizziness (PPPD) might all report similar sensations, even though the underlying causes and treatments are different.
Traditionally, doctors rely on a combination of patient history, symptom descriptions, physical exams, and special tests to make a diagnosis. This process takes skill, experience, and time—and it’s not unusual for patients to see several specialists before getting the right answer.
The Goal of the Study
Researchers wanted to see if they could build a machine learning (ML) tool—a type of artificial intelligence—that could help doctors sort through the clues and narrow down the possibilities for six common vestibular disorders:
- Menière’s Disease (MD) – often includes vertigo, hearing loss, and tinnitus.
- Benign Paroxysmal Positional Vertigo (BPPV) – brief vertigo triggered by head position changes.
- Vestibulopathy (VEST) – damage to the balance organs in the inner ear.
- Hemodynamic Orthostatic Dizziness (HOD) – dizziness related to changes in blood pressure when standing up.
- Vestibular Migraine (VM) – migraine attacks that include dizziness or vertigo.
- Persistent Postural-Perceptual Dizziness (PPPD) – ongoing dizziness or unsteadiness not caused by a current structural problem in the ear.
How the Model Was Built
The team started with a large pool of information from real patient histories—details like symptoms, triggers, and test results. They used a two-step process to decide which pieces of information to include:
- Computer analysis to find features most helpful for distinguishing between disorders.
- Input from medical experts to make sure the chosen features were practical and made sense clinically.
This produced 50 key “features” for the model to learn from.
They then trained a type of ML algorithm called CatBoost, which is especially good at handling complex medical data. The model was tuned to be:
Very sensitive to common, less invasive-to-treat disorders like BPPV and vestibular migraine (to catch as many cases as possible).
Very specific for conditions like Menière’s and HOD, where unnecessary or incorrect treatments could be harmful.
How Well It Worked
When tested, the model’s overall accuracy was 88.4%. Of all the cases:
- 60.9% were exactly right.
- 27.5% were close—pointing to a similar disorder.
- 11.6% were wrong.
Why This Matters
This tool could help doctors—especially those who aren’t vestibular specialists—make faster, more accurate diagnoses, leading to better treatment decisions and fewer unnecessary tests. It could also be a useful teaching aid for new clinicians learning to evaluate dizziness.
Limitations
While promising, the study has some important caveats:
- Limited scope – It only covers six vestibular disorders. Many others (like acoustic neuroma or labyrinthitis) aren’t included.
- Data source – The model was trained on a specific dataset. Its performance might change if used in other clinics or populations.
- Not a replacement for doctors – It’s meant to assist, not replace, a full clinical evaluation. Human judgment remains essential, especially for rare or complicated cases.
- Accuracy gaps – About 40% of cases were either close-but-not-quite or wrong, so results must be interpreted with caution.
Bottom Line
The study shows that a carefully designed machine learning tool, built with both computer algorithms and medical expertise, can be a helpful partner in diagnosing vestibular disorders. It’s not perfect and won’t replace the need for an experienced clinician, but it could speed up diagnosis, reduce errors, and improve care for people struggling with dizziness and balance problems.
Source
Callejas Pastor, C.A., Ryu, H.T., Joo, J.S. et al. Clinical decision support for vestibular diagnosis: large-scale machine learning with lived experience coaching. npj Digit. Med. 8, 487 (2025). https://doi.org/10.1038/s41746-025-01880-z
Expert Perspective – Limitations of the Study
VeDA Board President and Associate Professor of Neurology and Otolaryngology-Head and Neck Surgery at Johns Hopkins University, Dr. Amir Kheradmand, MD, offered this perspective:
“This tool shows promise, but it’s important to remember how its performance was measured. The model was trained and evaluated against the opinions of vestibular specialists, rather than confirmed diagnoses. In that sense, it reflects how closely the system can mirror an expert’s clinical judgment based on the information available, rather than how it performs against a definitive diagnostic gold standard.”
