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

Diagnosing the benign paroxysmal positional vertigo via 1D and deep-learning composite model

BACKGROUND

Benign Paroxysmal Positional Vertigo (BPPV) is the leading cause of vertigo, and its characteristic nystagmus induced by positional maneuvers makes it a good model for Artificial Intelligence (AI) diagnosis. However, during the testing procedure, up to 10 min of indivisible long-range temporal correlation data are produced, making the AI-informed real-time diagnosing unlikely in clinical practice.

METHODS

A combined 1D and Deep-Learning (DL) composite model was proposed. Two separate cohorts were recruited, with one for model generation and the other for evaluation of model's real-world generalizability. Eight features, including two head traces and three eye traces and their corresponding slow phase velocity (SPV) value, were served as the inputs. Three candidate models were tested, and a sensitivity study was conducted to determine the saliently important features.

RESULTS

The study included 2671 patients in the training cohort and 703 in the test cohort. A hybrid DL model achieved a micro-area under the receiver operating curve (AUROC) of 0.982 (95% CI 0.965, 0.994) and macro-AUROC of 0.965 (95% CI 0.898, 0.999) for overall classification. The highest accuracy was observed for right posterior BPPV, with an AUROC of 0.991 (95% CI 0.972, 1.000), followed by left posterior BPPV, with an AUROC of 0.979 (95% CI 0.940, 0.998), the lowest AUROC was 0.928 (95% CI 0.878, 0.966) for lateral BPPV. The SPV was consistently identified as the most predictive feature in the models. If the model process is carried out 100 times for a 10-min data, one single running takes 0.79 ± 0.06 s.

CONCLUSION

This study designed DL models which can accurately detect and categorize the subtype of BPPV, enabling a quick and straightforward diagnosis of BPPV in clinical setting. The critical feature identified in the model helps expand our understanding of this disorder.

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.

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.