1D Convolutional Neural Networks for Detecting Nystagmus

Exploring the Potentials of Wearable Technologies in Managing Vestibular Hypofunction

The vestibular system is dedicated to gaze stabilization, postural balance, and spatial orientation; this makes vestibular function crucial for our ability to interact effectively with our environment. Vestibular hypofunction (VH) progresses over time, and it presents differently in its early and advanced stages. In the initial stages of VH, the effects of VH are mitigated using vestibular rehabilitation therapy (VRT), which can be facilitated with the aid of technology. At more advanced stages of VH, novel techniques that use wearable technologies for sensory augmentation and sensory substitution have been applied to manage VH. Despite this, the potential of assistive technologies for VH management remains underexplored over the past decades. Hence, in this review article, we present the state-of-the-art technologies for facilitating early-stage VRT and for managing advanced-stage VH. Also, challenges and strategies on how these technologies can be improved to enable long-term ambulatory and home use are presented.

Digital biomarkers from gaze tests for classification of central and peripheral lesions in acute vestibular syndrome

Acute vestibular syndrome (AVS) is characterised by a sudden vertigo, gait instability, nausea and nystagmus. Accurate and rapid triage of patients with AVS to differentiate central (potentially sinister) from peripheral (usually benign) root causes is a challenge faced across emergency medicine settings. While there exist bedside exams which can reliably differentiate serious cases, they are underused due to clinicians' general unfamiliarity and low confidence interpreting results. Nystagmus is a fundamental part of AVS and can facilitate triaging, but identification of relevant characteristics requires expertise. This work presents two quantitative digital biomarkers from nystagmus analysis, which capture diagnostically-relevant information. The directionality biomarker evaluates changes in direction to differentiate spontaneous and gaze-evoked (direction-changing) nystagmus, while the intensity differential biomarker describes changes in intensity across eccentric gaze tests. In order to evaluate biomarkers, 24 sets of three gaze tests (left, right, and primary) are analysed. Both novel biomarkers were found to perform well, particularly directionality which was a perfect classifier. Generally, the biomarkers matched or eclipsed the performance of quantitative nystagmus features found in the literature. They also surpassed the performance of a support vector machine classifier trained on the same dataset, which achieved an accuracy of 75%. In conclusion, these biomarkers simplify the diagnostic process for non-specialist clinicians, bridging the gap between emergency care and specialist evaluation, ultimately benefiting patients with AVS.

Neurological update: neuro-otology 2023

Much has changed since our last review of recent advances in neuro-otology 7 years ago. Unfortunately there are still not many practising neuro-otologists, so that most patients with vestibular problems need, in the first instance, to be evaluated and treated by neurologists whose special expertise is not neuro-otology. The areas we consider here are mostly those that almost any neurologist should be able to start managing: acute spontaneous vertigo in the Emergency Room-is it vestibular neuritis or posterior circulation stroke; recurrent spontaneous vertigo in the office-is it vestibular migraine or Meniere's disease and the most common vestibular problem of all-benign positional vertigo. Finally we consider the future: long-term vestibular monitoring and the impact of machine learning on vestibular diagnosis.

Updates in neuro-otology

PURPOSE OF REVIEW

Recent updates with clinical implications in the field of neuro-otology are reviewed.

RECENT FINDINGS

Important updates relating to several neuro-otologic disorders have been reported in recent years. For benign positional paroxysmal vertigo (BPPV), we provide updates on the characteristics and features of the short arm variant of posterior canal BPPV. For the acute vestibular syndrome, we report important updates on the use of video-oculography in clinical diagnosis. For autoimmune causes of neuro-otologic symptoms, we describe the clinical and paraclinical features of kelch-like protein 11 encephalitis, a newly-identified antibody associated disorder. For cerebellar ataxia, neuropathy, vestibular areflexia syndrome, we report recent genetic insights into this condition.

SUMMARY

This review summarizes important recent updates relating to four hot topics in neuro-otology.

Quantifying the direct cost benefits of vestibular telemetry using the CAVA system to diagnose the causes of dizziness

BACKGROUND

It can be challenging to diagnose the cause of a patient's dizziness. Patients face significant delays before receiving a correct diagnosis as they will undergo many diagnostic tests under several different medical specialities. As well as prolonging the suffering of patients, these problems place a significant financial burden on health services worldwide. We have developed a wearable medical device which has the potential to diagnose the cause of a patient's dizziness using vestibular telemetry captured over a thirty-day period. We sought to quantify the potential direct cost savings of an alternative diagnostic pathway using our diagnostic device.

METHODS

In this work, we identified the existing diagnostic pathways followed by patients reporting dizziness to their General Practitioner, and modelled the best and worst-case direct costs of providing a patient with a correct diagnosis. We estimated the potential cost of our alternative pathway, and calculated the cost savings this could provide to the NHS.

RESULTS

The results show that our alternative diagnostic pathway could reduce the time and direct cost associated with providing a correct diagnosis. We present a potential indicative cost-saving of between £631 and £1305, per patient.

CONCLUSION

Our alternative diagnostic pathway would reduce the time taken to correctly diagnose patients with vertigo. This in turn would facilitate faster access to targeted treatments, reduce unnecessary interventions, and reduce the suffering of patients. These improvements would also lead to other savings, such as reducing the amount of sick leave taken by patients to attend appointments, and freeing up of NHS time to see other patients.

Deep learning in acute vertigo diagnosis

Recent advances in artificial intelligence are transforming healthcare and there are increasing efforts to apply these breakthroughs to the diagnosis of acute vertigo. Because the diagnosis of vertigo relies on the analysis of eye movements, there are several unique considerations that must be made when implementing deep learning approaches to vertigo. This review discusses the need for diagnostic aids for acute vertigo, the techniques used to preprocess eye movement data and adapt deep learning models to vertigo, and summarizes and analyzes all published models to date.

Artificial intelligence for early stroke diagnosis in acute vestibular syndrome

Objective

Measuring the Vestibular-Ocular-Reflex (VOR) gains with the video head impulse test (vHIT) allows for accurate discrimination between peripheral and central causes of acute vestibular syndrome (AVS). In this study, we sought to investigate whether the accuracy of artificial intelligence (AI) based vestibular stroke classification applied in unprocessed vHIT data is comparable to VOR gain classification.

Methods

We performed a prospective study from July 2015 until April 2020 on all patients presenting at the emergency department (ED) with signs of an AVS. The patients underwent vHIT followed by a delayed MRI, which served as a gold standard for stroke confirmation. The MRI ground truth labels were then applied to train a recurrent neural network (long short-term memory architecture) that used eye- and head velocity time series extracted from the vHIT examinations.

Results

We assessed 57 AVS patients, 39 acute unilateral vestibulopathy patients (AUVP) and 18 stroke patients. The overall sensitivity, specificity and accuracy for detecting stroke with a VOR gain cut-off of 0.57 was 88.8, 92.3, and 91.2%, respectively. The trained neural network was able to classify strokes with a sensitivity of 87.7%, a specificity of 88.4%, and an accuracy of 87.9% based on the unprocessed vHIT data. The accuracy of these two methods was not significantly different (p = 0.09).

Conclusion

AI can accurately diagnose a vestibular stroke by using unprocessed vHIT time series. The quantification of eye- and head movements with the use of machine learning and AI can serve in the future for an automated diagnosis in ED patients with acute dizziness. The application of different neural network architectures can potentially further improve performance and enable direct inference from raw video recordings.

aEYE: A deep learning system for video nystagmus detection

Background

Nystagmus identification and interpretation is challenging for non-experts who lack specific training in neuro-ophthalmology or neuro-otology. This challenge is magnified when the task is performed via telemedicine. Deep learning models have not been heavily studied in video-based eye movement detection.

Methods

We developed, trained, and validated a deep-learning system (aEYE) to classify video recordings as normal or bearing at least two consecutive beats of nystagmus. The videos were retrospectively collected from a subset of the monocular (right eye) video-oculography (VOG) recording used in the Acute Video-oculography for Vertigo in Emergency Rooms for Rapid Triage (AVERT) clinical trial (#NCT02483429). Our model was derived from a preliminary dataset representing about 10% of the total AVERT videos (n = 435). The videos were trimmed into 10-sec clips sampled at 60 Hz with a resolution of 240 × 320 pixels. We then created 8 variations of the videos by altering the sampling rates (i.e., 30 Hz and 15 Hz) and image resolution (i.e., 60 × 80 pixels and 15 × 20 pixels). The dataset was labeled as "nystagmus" or "no nystagmus" by one expert provider. We then used a filtered image-based motion classification approach to develop aEYE. The model's performance at detecting nystagmus was calculated by using the area under the receiver-operating characteristic curve (AUROC), sensitivity, specificity, and accuracy.

Results

An ensemble between the ResNet-soft voting and the VGG-hard voting models had the best performing metrics. The AUROC, sensitivity, specificity, and accuracy were 0.86, 88.4, 74.2, and 82.7%, respectively. Our validated folds had an average AUROC, sensitivity, specificity, and accuracy of 0.86, 80.3, 80.9, and 80.4%, respectively. Models created from the compressed videos decreased in accuracy as image sampling rate decreased from 60 Hz to 15 Hz. There was only minimal change in the accuracy of nystagmus detection when decreasing image resolution and keeping sampling rate constant.

Conclusion

Deep learning is useful in detecting nystagmus in 60 Hz video recordings as well as videos with lower image resolutions and sampling rates, making it a potentially useful tool to aid future automated eye-movement enabled neurologic diagnosis.

A Deep Learning Model for Three-Dimensional Nystagmus Detection and Its Preliminary Application

Symptoms of vertigo are frequently reported and are usually accompanied by eye-movements called nystagmus. In this article, we designed a three-dimensional nystagmus recognition model and a benign paroxysmal positional vertigo automatic diagnosis system based on deep neural network architectures (Chinese Clinical Trials Registry ChiCTR-IOR-17010506). An object detection model was constructed to track the movement of the pupil centre. Convolutional neural network-based models were trained to detect nystagmus patterns in three dimensions. Our nystagmus detection models obtained high areas under the curve; 0.982 in horizontal tests, 0.893 in vertical tests, and 0.957 in torsional tests. Moreover, our automatic benign paroxysmal positional vertigo diagnosis system achieved a sensitivity of 0.8848, specificity of 0.8841, accuracy of 0.8845, and an F1 score of 0.8914. Compared with previous studies, our system provides a clinical reference, facilitates nystagmus detection and diagnosis, and it can be applied in real-world medical practices.

Current concepts in acute vestibular syndrome and video-oculography

PURPOSE OF REVIEW

We present here neuro-otological tests using portable video-oculography (VOG) and strategies assisting physicians in the process of decision making beyond the classical 'HINTS' testing battery at the bedside.

RECENT FINDINGS

Patients with acute vestibular syndrome (AVS) experience dizziness, gait unsteadiness and nausea/vomiting. A variety of causes can lead to this condition, including strokes. These patients cannot be adequately identified with the conventional approach by stratifying based on risk factors and symptom type. In addition to bedside methods such as HINTS and HINTS plus, quantitative methods for recording eye movements using VOG can augment the ability to diagnose and localize the lesion. In particular, the ability to identify and quantify the head impulse test (VOR gain, saccade metrics), nystagmus characteristics (waveform, beating direction and intensity), skew deviation, audiometry and lateropulsion expands our diagnostic capabilities. In addition to telemedicine, algorithms and artificial intelligence can be used to support emergency physicians and nonexperts in the future.

SUMMARY

VOG, telemedicine and artificial intelligence may assist physicians in the diagnostic process of AVS patients.

Detecting positional vertigo using an ensemble of 2D convolutional neural networks

•

We trained Deep Neural Networks to detect attacks of motion provoked dizziness.

•

2D Convolutional Deep Neural Networks outperform 1D network architectures.

•

Best results were provided by input features combining eye- and head-movement.

•

An ensemble of five networks outperformed each individual network alone.

The aim of the work presented here was to develop a system that can automatically identify attacks of dizziness occurring in patients suffering from positional vertigo, which occurs when sufferers move their head into certain positions. We used our novel medical device, CAVA, to record eye- and head-movement data continually for up to 30 days in patients diagnosed with a disorder called Benign Paroxysmal Positional Vertigo. Building upon our previous work, we describe a novel ensemble of five 2D Convolutional Neural Networks, using composite recognition features, including eye-movement data and three-channel accelerometer data. We achieve an F1 score of 0.63 across an 11-fold cross-fold validation experiment, demonstrating that the system can detect a few seconds of motion provoked dizziness from within over a 100 h of normal eye-movement data. We show that the system outperforms our previous 1D Neural Network approach, and that our ensemble classifier is superior to each of the individual networks it contains. We also demonstrate that our composite recognition features provide improved performance over results obtained using the individual data sources independently.