Artificial intelligence in dizziness and vertigo: A review of machine learning and deep learning applications for multimodal diagnosis and management

Dizziness/vertigo is among the most common symptoms in clinical practice, affecting approximately 15%–20% of adults each year. Its etiologies are complex and involve multiple systems, including the central nervous system, peripheral vestibular system, and psychological factors. Because symptoms are highly subjective and key physical signs are difficult to capture, diagnosis remains heavily dependent on clinicians’ experience, particularly in emergency settings, where central causes such as posterior circulation stroke are at risk of being missed. Artificial intelligence (AI), especially machine learning (ML) and deep learning (DL), can automatically extract latent features from high-dimensional and multimodal clinical data, providing a new approach for the diagnosis and management of dizziness/vertigo disorders. This review summarizes recent advances in AI applications for dizziness/vertigo disorders, covering multimodal data sources such as medical imaging, videonystagmography (VNG), video head impulse test (vHIT), postural and gait assessment, and structured clinical information. It also summarizes AI applications in benign paroxysmal positional vertigo, Ménière’s disease, vestibular migraine, posterior circulation stroke, orthostatic intolerance, and related disorders. Current studies have shown promising performance of AI in multiple scenarios. For example, a DL-based nystagmus pattern classification model achieved an accuracy of 94.91% using VNG videos; an LSTM model based on raw vHIT time-series data achieved an accuracy of 87.9% for stroke detection in acute vestibular syndrome; and a CNN-based automated inner ear MRI segmentation model enabled objective quantification of endolymphatic hydrops. However, current studies are still limited by small sample sizes, a high proportion of single-center datasets, inconsistent annotation, insufficient model interpretability, and a lack of prospective validation, which restrict their clinical translation. Future research should strengthen the development of standardized multicenter datasets, advance explainable AI, federated learning, and multimodal large models, and promote the expansion of AI from diagnostic assistance to treatment decision support and long-term management.