Differential diagnosis of vertigo using machine learning techniques

Abstract

Patients with vertigo can be categorised into syndromes, each with two main differential diagnoses. Syndromes include acute vestibular syndrome (AVS), typically caused by vestibular neuritis (VN) or posterior circulation stroke (PCS), and recurrent spontaneous vertigo (RSV), usually due to vestibular migraine (VM) or Menière’s disease (MD). For both AVS and RSV, clinicians use information from the history, bedside examination and investigations to determine the diagnosis, but this task can be challenging, particularly for non-experts. This thesis aimed to develop and validate machine learning models to differentiate VM from MD and VN from PCS and to tailor these models for clinicians with various levels of expertise.

For RSV patients, models used different combinations of data from history, examination and vestibular tests to simulate various clinical settings. These models separated VM and MD with accuracies of 97.8%, 94.5% and 92.3% for the neuro-otology clinic, general neurology/otolaryngology clinic and primary care respectively. Using a similar approach, models aimed at clinicians reviewing AVS patients in the Emergency Room differentiated VN and PCS with accuracies of 96.6% for those with neuro-otology support, 94.6% for non-experts with access to the video head impulse test (VHIT), and 88.8% for non-experts using history and basic bedside examination only. Finally, models that used only raw VHIT data to distinguish VN from PCS achieved 87.8% accuracy when evaluated on an external dataset, with performance superior to the optimal gain cut-off and not significantly different from that of experts.

The findings of this thesis demonstrate that machine learning models can accurately differentiate VM from MD and VN from PCS using clinical data. These models hold promise as diagnostic decision aids, suitable not only for subspecialists but also non-experts without neuro-otology expertise or resources, thereby bridging diagnostic gaps in current clinical practice.