Improving Quality Control in Oscillometry: Repeatability, Efficiency, Feasibility and Accuracy

Abstract

RATIONALE: The current oscillometry acceptance criteria for measurement of respiratory impedance (Zrs) requires a minimum of three repeated measurements with a coefficient of variation (CV) of ≤ 10% in adults or ≤ 15% in young children at the lowest frequency resistance (R5). However, this acceptability criteria ignores all the other frequencies and the significant reactance (Xrs) component of Zrs. This thesis assessed novel algorithms that include variability in resistance (Rrs) and Xrs over a range of frequencies to improve the repeatability, efficiency, feasibility, and accuracy of oscillometry. It also explored if machine learning can be used to predict Chronic Obstructive Pulmonary Disease (COPD) severity using combinations of spirometry and oscillometry measures. METHODS: This thesis explored different automated weighted combination sums of Rrs or Zrs CVs across frequencies and sought the first three measurements out of all measurements with a CV ≤15% for young children and ≤10% for adults. Three different data sets were used, each including five to as many as 12 measurements per subject: 1) 550 five years old population representative children in Toronto (CHILD5Y), 2) 110 three to five years old children with wheeze (WESER) and 3) 818 adult clinic subjects with predominantly COPD (West Island Cohort, WIC). The repeatability, efficiency, and feasibility of the proposed Quality Control (QC) algorithm was first optimized using CHILD5Y and validated using WESER and WIC. Physiological variability and artifact distributions from CHILD5Y were also used to generate a computational model, which was employed to assess the accuracy of the proposed algorithm. Machine learning algorithms including Single Decision Tree, Bagged Decision Trees, Support Vector Machines and Gradient Boosting were assessed to predict COPD Assessment Test (CAT) scores based on oscillometry and spirometry inputs. RESULTS: It was found that using the proposed QC algorithm, Early, with an inverse frequency weighted sum of the Zrs outperformed current recommended criteria for CV, reducing the CV of the important outcome measures and achieving the best feasibility. Feasibility improved compared to no QC when restricting analysis to the first 5 measurements, from 64%, 61%, and 49% to 85%, 80% and 81%, while using all available measurements with QC improved feasibility to 94%, 91% and 82% with CHILD5Y, WESER and WIC, respectively. Early also improved efficiency by reducing the number of required measurements from 5.4±1.7 to 3.7(0.9). Accuracy was maintained when applying the Early algorithm, resulting in comparable Root Mean Square Error (RMSE) to no QC and compared to QC method based on two standard deviations. Accuracy was also maintained for the important oscillometry measures R5-19 and AX. It was found that using machine learning with combined spirometry and oscillometry measures outperforms the use of spirometry or oscillometry measures separately. CONCLUSION: Optimizing an automated CV algorithm based on Zrs across frequencies provided improved repeatability, efficiency and feasibility, while maintaining the measurement accuracy. Additionally, machine-based algorithms using combinations of spirometry and oscillometry measures show potential for patient screening and monitoring.