Human Versus Computer Algorithmic Measurements of Caloric Response: Implications for Test Analysis

Abstract

Objectives: To quantify variation introduced to bithermal caloric test (BCT) analysis by data cleaning and determine how variation differs between examiners. Methods: Analysis of 435 consecutive BCTs performed by 6 examiners using identical protocols on adults with dizziness. Outcomes of total eye speed (TES) and unilateral weakness (UW%) were compared between examiner-modified tracings and automated algorithms. Results: Algorithms erroneously selected artifact in 9.7% of tests. Examiner cleaning resulted in a mean change in TES of (-)4deg/sec (95%CI 3-4, p<0.001) but no change in UW%. Limits of agreement (Bland-Altman analyses) for TES were (-)20 to (+)8deg/sec and for UW (-)10 to 10% and varied between examiners. Algorithms had 15% false negative and 2% false positive rates. Conclusions: Data cleaning may reduce the rate of false negative results. Differences in cleaning methods may produce test-retest and inter-individual variation and alter lab-derived normative values. Consensus is needed regarding optimal data cleaning methods.