Automated classification of benign paroxysmal positional vertigo from video-nystagmography using a delay-aware neural...

Area of Science:

• Neurology
• Medical Diagnostics
• Artificial Intelligence

Background:

• Nystagmus is a critical sign of vestibular disorders like benign paroxysmal positional vertigo (BPPV).
• Accurate BPPV diagnosis is vital for effective treatment with bedside maneuvers, improving patient outcomes and preventing unnecessary interventions.
• Current methods for identifying positional nystagmus from video nystagmography (VNG) can be subjective and difficult to standardize due to subtle or variable signals.

Purpose of the Study:

• To introduce DSF-BPPVNet, a novel delay-aware neural network architecture designed for classifying BPPV from VNG traces.
• To evaluate the performance of DSF-BPPVNet against existing deep learning models in a patient-independent setting.
• To analyze the explainability of the DSF-BPPVNet model, understanding its attribution patterns and temporal weighting.

Main Methods:

• Developed DSF-BPPVNet, a neural architecture incorporating temporal convolution, delayed-state feedback, and residual refinement.
• Trained and evaluated the model on 3,111 VNG traces from 705 patients using 5-fold cross-validation.
• Compared DSF-BPPVNet performance against established deep-learning baselines.

Main Results:

• DSF-BPPVNet demonstrated superior performance in patient-independent BPPV classification compared to baseline models.
• Achieved a significant F1-score of 0.819 ± 0.020.
• Explainability analyses provided insights into the model's decision-making process and temporal data interpretation.

Conclusions:

• DSF-BPPVNet offers a robust and accurate method for classifying BPPV from VNG data.
• The model's performance suggests potential for improved objective diagnosis of vestibular disorders.
• Further research into AI-driven diagnostic tools can enhance clinical practice for conditions like BPPV.