62 Chapter 3 Threshold derivation SUV images were visually assessed as ‘positive’ or ‘negative’ by a trained and experienced nuclear medicine physician (BvB) who was blinded for clinical information, based on standards provided by the manufacturer (32). Next, we used different data-driven methods to obtain thresholds for amyloid positivity for both BPND and SUVr. First we used the R studio function normalmixEM to fit Gaussian mixture models (GMM) with 1-9 components. Bayesian information criterion (BIC) indicated a model with 2 components as being the most optimal fit to our data. A threshold was derived representing the mean of the calculated mu of both components. The calculated thresholds were similar when we used the proportions derived from visual assessment (24% and 76%) as a starting value for mixture weights. This resulted in cut-off points of 0.23 (BPND) and 1.34 (SUVr). Next, we used K-means clustering. We assumed the data consisted of two clusters. We derived two cut-off values, the first representing the 90th percentile of the cluster with low amyloid burden, and the second representing the 10th percentile of the cluster with high amyloid burden. The cut-off values were purely data-driven, and information about visual assessment of scans was not used for these thresholds. This resulted in a low threshold (0.19 BPND and 1.28 SUVr), and a high threshold (0.29 BPND and 1.43 SUVr). Subsequently, we took the area between the lower and higher thresholds derived by K-means clustering to operationalise a grey zone. Figure 1 shows a summary of all derived thresholds and visualizes the predefined grey zone.
RkJQdWJsaXNoZXIy MjY0ODMw