Abstract
Purpose: To verify the accuracy of automated nystagmus detection algorithms.
Method: Video-oculography (VOG) plots were analyzed from consecutive patients with dizziness presenting to a neurology clinic. Data were recorded for 30 s in upright position with fixation block. For automated nystagmus detection, slow-phase algorithm parameters included mean and median slow-phase velocity (SPV), and slow-phase duration ratio. Quick-phase algorithm parameters included saccadic difference and saccadic ratio. For verification, two independent blinded assessors reviewed VOG traces and videos and coded presence or absence of nystagmus. Assessor consensus was used as reference standard. Accuracy of slow-phase and quick-phase algorithm parameters were compared, and ROC analysis was performed.
Results: Among 524 analyzed VOG traces, 99 were verified as nystagmus present and 425 were verified as nystagmus absent. Prevalence of nystagmus in the sample population was 18.9%. In ROC analysis, areas under the curve of individual algorithm parameters were 0.791-0.896. With optimal thresholds for determining presence or absence of nystagmus, algorithm sensitivity (70.7-87.9%), specificity (71.8-84.0%), and negative predictive value (91.7-96.4%) were ideal, but positive predictive value (38.8-53.4%) was not ideal. Combining algorithm parameters using logistic regression models mildly improved detection accuracy.
Conclusion: Both slow-phase and fast-phase algorithms were accurate for detecting nystagmus. Due to low positive predictive value, the utility of independent automated nystagmus detection systems is limited in clinical settings with low prevalence of nystagmus. Combining parameters using logistic regression models appears to improve detection accuracy, indicating that machine learning may potentially optimize the accuracy of future automated nystagmus detection systems.
Method: Video-oculography (VOG) plots were analyzed from consecutive patients with dizziness presenting to a neurology clinic. Data were recorded for 30 s in upright position with fixation block. For automated nystagmus detection, slow-phase algorithm parameters included mean and median slow-phase velocity (SPV), and slow-phase duration ratio. Quick-phase algorithm parameters included saccadic difference and saccadic ratio. For verification, two independent blinded assessors reviewed VOG traces and videos and coded presence or absence of nystagmus. Assessor consensus was used as reference standard. Accuracy of slow-phase and quick-phase algorithm parameters were compared, and ROC analysis was performed.
Results: Among 524 analyzed VOG traces, 99 were verified as nystagmus present and 425 were verified as nystagmus absent. Prevalence of nystagmus in the sample population was 18.9%. In ROC analysis, areas under the curve of individual algorithm parameters were 0.791-0.896. With optimal thresholds for determining presence or absence of nystagmus, algorithm sensitivity (70.7-87.9%), specificity (71.8-84.0%), and negative predictive value (91.7-96.4%) were ideal, but positive predictive value (38.8-53.4%) was not ideal. Combining algorithm parameters using logistic regression models mildly improved detection accuracy.
Conclusion: Both slow-phase and fast-phase algorithms were accurate for detecting nystagmus. Due to low positive predictive value, the utility of independent automated nystagmus detection systems is limited in clinical settings with low prevalence of nystagmus. Combining parameters using logistic regression models appears to improve detection accuracy, indicating that machine learning may potentially optimize the accuracy of future automated nystagmus detection systems.
Original language | American English |
---|---|
Journal | Journal of the Neurological Sciences |
State | Published - 28 Aug 2022 |