Continuous Ambulatory Vestibular Assessment (CAVA): Development of a System to Provide an Automatic Diagnosis for Vestibular Conditions

This is a multicentre diagnostic accuracy study with a recruitment target of 255 adult participants. Recruitment will take place across ten sites in the UK, over two years. To be eligible, participants must have an existing diagnosis of Ménière's disease, vestibular migraine or BPPV. The study will recruit eighty-five participants with each condition. Each patient's diagnosis will form the "reference standard" which will later be compared with a automated computer diagnosis.

Each participant will wear the CAVA device near continuously for thirty days. The device will record all eye and head movements during this time, including during any episodes of vertigo. When a patient experiences an episode of vertigo, they will be asked to activate an event marker on the CAVA device and keep a written record of attack onset and duration in a diary. Although this manual logging will not be relied upon for algorithm development, the whole 30 days of data will be used to detect periods of eye movement associated with dizziness.

The investigators will develop computer algorithms that use the data captured by the CAVA devices to differentiate between three target conditions. These algorithms will build on the algorithms that have been developed during the current MRC grant which have produced excellent results.10,11,17. The algorithms use deep learning neural networks (DNNs) in a so-called 'end-to-end' approach. In this approach, a multi-layered network takes as input 'raw' time domain data and outputs a classification decision without using any explicit intermediate representations of the signal (such as signal entropy, power, skewness, kurtosis etc.). End-to-end approaches, which typically utilise alternating sequences of convolution and pooling layers, fully connected layers, auto-encoders and long short-term memory (LSTM) layers, have been shown to outperform traditional approaches in cases where large amounts of training data are available. However, they provide little insight into how they work (i.e., which features of the signal are important for classification). It will therefore also be considered to develop methods in which intermediate representations are generated and then use these input features to test a range of classifiers (e.g., boosted trees, random forests, logistic regression, support vector machines etc.).

The performance of the algorithms will be assessed using the 'leave-one-out cross-validation' technique, which is the usual way of evaluating classifiers.22 In this technique, the data from a single patient is set aside and the remaining data from the other patients is used to 'train' the classifier. A algorithm will be referred to that is trained to distinguish the three conditions as a 'classifier'. The classifier is then tested on the set aside data: the testing process consists of comparing the classifier's predicted diagnosis ('Ménière's disease', 'vestibular migraine' or 'BPPV') with the true diagnosis (the reference standard label). This process is repeated, removing a different patient's data each time, producing an automated diagnosis for all 255 patients. Although computationally expensive, this ensures that the classifier has not 'seen' during training the data from the patient it tests, which would bias its decision. It also closely matches the real-world scenario of a new patient presenting to their clinician and a diagnosis being provided on the basis of their experience of other patients.

The study will present the sensitivity and specificity for diagnosing each of the three target conditions and Receiver Operator Curves (ROCs), showing how the diagnostic accuracy changes for various classification thresholds.