71 Grey zone amyloid burden We used different data-driven methods, such as Gaussian mixture modelling and K-means clustering, to derive cut-off values for amyloid positivity. We found thresholds of 0.19, 0.23 and 0.29 for BPND, and thresholds of 1.28, 1.34 and 1.43 for SUVr. Literature has generated inconsistent findings with respect to amyloid thresholds, ranging from 1.08-1.34 for SUVr, with 1.10 being reported most frequently (12, 18-23, 42-45). The large variability indicates that thresholds may to some extent rely on methodology, image processing pipeline used and study sample. For example, the partial volume correction method (46), and the choice of ROIs (47) affect the degree of amyloid burden. For this reason we used a commonly used ‘meta ROI’ (21, 24, 25), which is able to clearly distinguish AD patients from cognitively normal controls. However, small differences can be seen across studies (26, 48, 49). In addition, thresholds are dependent on sample characteristics (50). We aimed to minimize this effect with our choice of robust data-driven methods. Although our thresholds seemed substantially higher than the aforementioned thresholds, all corresponded equally well to visual assessment. We show that dichotomized BPND values may even correspond to visual assessment somewhat better, which is consistent with the findings of a previous [18F] flutemetamol PET study (13). All amyloid positivity thresholds predicted future memory decline, which is consistent with another study (51), although models with BPND thresholds and visual assessment seemingly resulted in a slightly better fit. Because of the underlying gradual association between amyloid burden and memory function, apparently the height of the threshold does not necessarily have a substantial effect on the association between amyloid positivity and memory function. Strengths of this study include that we used two measures of amyloid quantification, BPND and SUVr, and that we applied various data-driven approaches. BPND has been shown to be less sensitive to differences in flow and we found a good concordance with visual assessment. Using BPND and SUVr as continuous measures enabled us to thoroughly explore the grey zone, which is not possible with a strict binary division like visual assessment. Furthermore, we had a large, well-defined cohort, with a relatively long follow-up. Limitations include the lack of a gold standard such as pathology confirmation. Notwithstanding, we used visual assessment for comparison analyses which has been shown to correlate very well with pathology (12, 52). Furthermore, we used memory decline as outcome measure, as opposed to clinical progression to MCI or dementia. This might have led to less robust results because memory performance may be reversible, particularly in cases with limited amyloid burden. On the other hand, it might take a relatively long time before a substantial part of this sample shows clinical progression or cognitive impairment. Nevertheless, since we had cognitive data covering on average 3.8 years, our models should give an accurate estimation of memory slope. Using these methods, we were able to capture subtle decline, 3