Machine Learning Closes the 3 AM Diagnostic Gap for Life-Threatening Vertigo

When a patient stumbles into the emergency department at 3 a.m. with severe vertigo, frontline physicians face a diagnostic nightmare: is this a benign inner ear problem or a life-threatening stroke masquerading as dizziness? Now, researchers from the University of Sydney and Royal Prince Alfred Hospital have developed machine learning algorithms that solve this critical triage challenge with expert-level accuracy—even when specialist neurologists are unavailable.

AI Matches Specialist Performance Across Resource Settings

The gradient-boosted tree models, published in the Journal of Neurology, were trained on 294 consecutive emergency department patients presenting between March 2018 and March 2023. The cohort included 163 patients with acute unilateral vestibulopathy (benign inner ear dysfunction) and 131 with imaging-confirmed posterior circulation stroke.

Lead researchers Chao Wang and Miriam Welgampola designed a three-tier system matching different clinical resource levels. The top-tier model achieved 96.6% accuracy (95% CI: 93.3–99.9%) using full neuro-otology testing. More importantly, the mid-tier model—requiring only structured history, basic examination, and a five-minute video head impulse test—achieved 94.6% accuracy, statistically equivalent to expert-performed HINTS examination. Even the most basic tier, using history and examination alone, reached 88.8% accuracy.

The researchers tested XGBoost, LightGBM, and CatBoost algorithms with five-fold stratified cross-validation across 203 input variables, including demographics, vertigo characteristics, cardiovascular risk factors, and various vestibular function tests. By lowering probability thresholds, sensitivity could be pushed above 99% in better-resourced tiers.

Addressing a High-Stakes Clinical Blind Spot

Acute vestibular syndrome represents 10–20% of emergency department dizziness presentations, yet posterior circulation strokes lack focal neurological deficits in two-thirds of cases. Early MRI misses approximately 15% of strokes within the first 48 hours, creating a dangerous diagnostic gap.

The bedside head impulse test emerged as the most influential feature in the top-tier model, followed by spontaneous nystagmus slow phase velocity and focal neurological symptoms. Notably, focal symptoms appeared in 65.6% of stroke patients versus only 3.0% of vestibulopathy patients, but their absence doesn't rule out stroke.

The models correctly classified some atypical cases where traditional HINTS examination failed, including two vestibulopathy patients without spontaneous nystagmus and two with misleading test-of-skew results. Persistent failures included two atypical vestibulopathy cases and one pontine stroke presenting without focal signs.

Democratizing Specialist-Level Diagnosis

The clinical significance extends beyond raw accuracy numbers. Frontline emergency physicians frequently avoid, misperform, or misinterpret the HINTS examination, particularly during overnight shifts when specialist support is unavailable. The video head impulse test takes approximately five minutes at bedside, is well-tolerated by patients, and yields objective quantified results that are easier to interpret than naked-eye examination.

For health systems, the implications are substantial. Missed strokes result in preventable disability and litigation, while false-positive stroke diagnoses drive expensive neuroimaging and potentially harmful reperfusion therapy. An AI-assisted triage system achieving 99%+ sensitivity could serve as a gatekeeper for advanced imaging and specialist consultation, optimizing resource allocation in a high-volume, high-stakes scenario.

The tiered architecture demonstrates practical deployment pathways across different resource settings, from tertiary centers to community hospitals, potentially integrating with existing video head impulse test devices or video-oculography systems already deployed in emergency departments.

Sources

1. AI Spots the Stroke Hiding Behind Sudden Dizziness With Expert-Level — Scienmag