Optimal Classification of Respiratory Patterns From Manual Analyses Using Expectation-Maximization

Manual scoring (MS) of cardiorespiratory signals is the gold standard method for the analysis of respiratory data in sleep laboratories. In MS, trained, expert scorers characterize respiratory patterns by scrolling through a data record and visually identifying patterns. However, MS is limited by high intraand inter-scorer variability and subjectivity. A strategy to mitigate this is to analyze the same respiratory data multiple times and generate a consensus. This consensus is generally determined by a majority vote (MV), where the most frequent pattern is selected as the true pattern. This paper presents expectation-maximization pattern sequence (EM-PSEQ), a novel method based on EM that estimates the true patterns optimally. A simulation study examined the accuracies of EM-PSEQ, MV, and individual scorers (IS) as a function of the number of analyses. Accuracy was measured with the Fleiss κ statistic, and is reported as [κ_MDN = x; κ_P5 = y], where κ_MDN, the median value, is the expected accuracy, and κ_P5, the 5th percentile value, gives the minimum accuracy for 95% confidence. IS accuracy remained constant at [κ_MDN = 0.67; κ_P5 = 0.60] as the number of analyses increased. MV accuracy increased slowly with the number of analyses and plateaued at [κ_MDN = 0.78; κP5 = 0.76] after five analyses. In contrast, EM-PSEQ accuracy improved quickly, reaching an almost perfect value of [κ_MDN = 0.83; κ_P5 = 0.77] with four analyses, and perfect accuracy [κ_MDN = 1.00; κ_P5 = 0.99] after 25 analyses. EM-PSEQ performed much better than either MV or IS, and required only modest computational effort. Consequently, we believe EM-PSEQ will be a very valuable tool for clinical studies, as it can dramatically improve the accuracy of manual respiratory analysis with minimal additional cost.