Predicting amyloid-PET status in a memory clinic: The role of the novel antero-posterior index and visual rating scales

Introduction: Visual rating scales are increasingly utilized in clinical practice to assess atrophy in crucial brain regions among patients with cognitive disorders. However, their capacity to predict Alzheimer's disease (AD)- related pathology remains unexplored, particularly within a heterogeneous memory clinic population. This study aims to assess the accuracy of a novel visual rating assessment, the antero-posterior index (API) scale, in predicting amyloid-PET status. Furthermore, the study seeks to determine the optimal cohort-based cutoffs for the medial temporal atrophy (MTA) and parietal atrophy (PA) scales and to integrate the main visual rating scores into a predictive model. Methods: We conducted a retrospective analysis of brain MRI and high-resolution TC scans from 153 patients with cognitive disorders who had undergone amyloid-PET assessments due to suspected AD pathology in a real-world memory clinic setting. Results: The API scale (cutoff ≥ 1) exhibited the highest accuracy (AUC = 0.721) among the visual rating scales. The combination of the cohort-based MTA and PA threshold with the API yielded favorable accuracy (AUC = 0.787). Analyzing a cohort of MCI/Mild dementia patients below 75 years of age, the API scale and the predictive model improved their accuracy (AUC = 0.741 and 0.813, respectively), achieving excellent results in the early-onset population (AUC = 0.857 and 0.949, respectively). Conclusion: Our study emphasizes the significance of visual rating scales in predicting amyloid-PET positivity within a real-world memory clinic. Implementing the novel API scale, alongside our cohort-based MTA and PA thresholds, has the potential to substantially enhance diagnostic accuracy.


Introduction
Alzheimer's disease (AD) has the highest prevalence among all forms of dementia globally [1].
The aging population, coupled with advancements in diagnostic capabilities, is widely recognized as a significant contributor to the projected increase in AD prevalence [2,3].Consequently, there is a critical demand for accessible and efficient tools capable of accurately predicting AD-related pathology.
Brain imaging, endorsed by recent diagnostic criteria [4,5], is indispensable for AD diagnosis.Both brain computed tomography (CT) and magnetic resonance imaging (MRI) play pivotal roles in identifying AD hallmarks and facilitating advanced diagnostic assessments [6], including amyloid and tau positron emission tomography (PET) as well as cerebrospinal fluid (CSF) and innovative plasma biomarkers [7,8].
Visual rating scales and semi-automated to automated quantification tools have provided substantial support in detecting AD-related neurodegeneration.Although partially automatic and automatic methods have demonstrated impressive accuracy, their implementation in routine clinical practice poses challenges due to time constraints [9,10].
Several visual rating scales have been developed over the years to assess atrophy in crucial brain regions affected by neurodegeneration in Alzheimer's disease (AD), as highlighted in a recent review [11].Notably, the medial temporal atrophy (MTA) scale [12], parietal atrophy (PA) scale [13], and global cortical atrophy (GCA) scale [14], including the recent "frontal" subtype (GCA-F) [15], have emerged as reliable methods [6].Additionally, the Fazekas scale stands out as one of the most reliable methods for assessing the severity of white matter hyperintensities, which are frequently associated with AD [16,17].
Although these methods have exhibited high accuracy in specific contexts, such as distinguishing between AD and healthy subjects or in selective cohorts with stringent inclusion criteria (e.g., the research cohorts such as the Alzheimer's Disease Neuroimaging Initiative) [18,19], their validation in real-world memory clinics, where patients with diverse cognitive disorders, including those beyond AD, are encountered, is limited [20].
Given the distinct original objectives of these scales, significant efforts have been directed toward establishing optimal cut-offs for visual rating scales [21] and introducing novel imaging-based approaches [22] to enhance the prediction of amyloid-PET positivity.
Nevertheless, the potential utility of combining various visual rating scales and the role of a scale specifically designed for this purpose are areas that remain to be discovered.
In this study, our objective is to assess the diagnostic accuracy of visual rating scales in predicting amyloid PET positivity within a real memory clinic setting.
We aim to achieve this by evaluating the performance of the following approaches: (I) A newly developed visual rating scale, known as the antero-posterior index (API), which has been specifically designed for this purpose.(II) The utilization of cohort-based cut-offs for MTA, PA, and GCA-F scales.(III) The application of previously established normative values for MTA, PA, and GCA-F scales, including the widely recognized Ferreira criteria [18], and the Cotta Ramusino criteria [23], which were established through the analysis of an Italian population.

Participants and study design
Data from a retrospective analysis (from January 2015 to December 2022) were utilized, involving 166 consecutive adult patients with cognitive disorders who were referred to the Center for Cognitive and Dementia Disorders (AUSL) and the Dementia Unit at the University Hospital of Parma, Italy.A comprehensive routine work-up for dementia was conducted for all patients, encompassing general and neurological examinations, routine blood tests, and structural imaging.Additionally, these patients underwent an assessment of brain amyloid status using amyloid-PET due to suspected AD pathology.
The inclusion criteria for this study were as follows: -Execution of MRI or high-spatial resolution CT Imaging for brain atrophy assessment.Both of these methods can be employed with high accuracy for the evaluation of visual rating scales [6,24].
• Availability of screening neuropsychological testing inclusive of the Mini Mental State Examination (MMSE) [25].• Execution of an Amyloid-PET assessment within one year from the date of structural imaging and neuropsychological assessment.
Out of the initial 166 patients, a total of 153 patients were selected for analysis (13 were excluded due to low quality of brain images).
The diagnosis of mild cognitive impairment (MCI) was performed according to the Petersen criteria [26], while the diagnosis of AD dementia was based on the International Working Group (IWG)-2 [5].Diagnoses of neurodegenerative diseases other than AD (e.g., frontotemporal dementia, dementia with Lewy bodies, Parkinson-dementia, corticobasal syndrome) were established according to respective clinical criteria during the clinical follow-up [27][28][29][30].
We assessed the performance of visual rating scales (MTA, PA, GCA-F, API) for amyloid-positivity prediction within the entire cohort.Additionally, to evaluate the predictive accuracy of these scores in a population commonly encountered during initial clinical assessments, we conducted a sub-analysis focusing on individuals affected by MCI and mild dementia [31,32].
Moreover, we performed sub-analyses specifically targeted on patients under 65 and under 75 years of age, according to age thresholds commonly applied on visual rating scales.We also chose this age cutoff given the importance of better evaluating the performance of visual rating scales in an early-onset cohort (i.e., < 65 years of age) and the previously observed weak correlation between the presence of neuritic plaques and dementia in individuals over 75 years [33][34][35].In older patients, this correlation might impact clinical symptoms and brain atrophy, which are not solely attributed to AD pathology but are also associated with coexisting conditions such as cerebrovascular lesions and dementia with Lewy bodies [36,37].
All subjects were enrolled after obtaining written informed consent, and the study was carried out in accordance with Helsinki principles and with approval from the local ethical committee.

Amyloid status assessment
PET scans were performed using a whole-body hybrid system, Discovery IQ (GE Healthcare), in the three-dimensional detection mode.A head holder was employed to minimize patient movements, and regular checks ensured head stability.Cerebral emission scans were conducted 90 min after intravenous injection of either 4 MBq/kg weight (240-360 MBq) of [18F]florbetaben or 2 MBq/kg weight (150-250 MBq) of [18F] flutemetamol.Each patient underwent 10-min frame acquisitions to obtain movement-free images.Amy-PET sinograms were reconstructed using a 3-D iterative algorithm, incorporating corrections for randomness, scatter, photon attenuation, resulting in images with a voxel size of 2 × 2 × 2 mm and a spatial resolution of approximately 5 mm full-width at half-maximum (FWHM) in the field of view center.
Blinded visual PET image assessments were performed by two trained and independent readers (M.S; L.R.).
In case of disagreement, the decision was reached through consensus.
Florbetaben PET images were analyzed using a semiquantitative visual scan assessment [38].Regional quantification of [18F]flutemetamol uptake was achieved using a fully automated PET method based on a composite standardized uptake value ratio (SUVR) threshold derived from an autopsy cohort [39].The SUVRs in the cerebral cortex were automatically generated and normalized to the pons using Cor-texID Suite software, yielding z-scores for each examined cerebral area.Amyloid + status was assigned for Florbetaben and Flutemetamol, according to previous studies [39,40].[18F] florbetaben and [18F] flutemetamol have demonstrated high concordance in recent studies [41].

Structural brain imaging assessment
Sixty-eight patients underwent brain MRI scans using a 3 Tesla scanner (GE Healthcare Discovery MR 750) equipped with an 8-channel head coil, at the Neuroradiology Department of the University-Hospital of Parma.The imaging protocol included 3D T1 weighted (BRAVO, section thickness 0.9 mm, TR/TE 12.36/5.18ms, flip angle 13 • ), fluid- attenuated inversion recovery (FLAIR), gradient-echo (GRE) and diffusion-weighted imaging (DWI).
CT scans of 85 patients were performed on a 64-slice CT scanner (Somatom Definition Edge 128, Siemens Healthineers, Forchheim, Germany) equipped with a 48-mm wide detector (64 detectors each 0.6 mm wide along the z-axis), with high-resolution coronal and sagittal reconstructions.
We also utilized Cotta Ramusino's cut-off criteria for comparison: MTA >1 below age 60, MTA >2 between 60 and 80, and MTA >3 beyond age 80 were deemed pathological.Similarly, PA >1 below age 60 and PA >2 beyond age 60 were considered pathological.The same PA criteria were applied for evaluating GCA-F.[23].
Regarding the assessment of the Fazekas score, values equal to or >2 were considered pathological for individuals under 70 years of age, while a score of 3 was considered pathological regardless of age [16].
In addition, we introduced our novel visual rating scale, the API scale representing an index for antero-posterior atrophy, which was defined by the delta between PA and GCA-F.A value ≥1 was defined as pathological (Fig. 1), according to the highest Youden index value.
The visual rating scales of each subject were rated twice by two trained readers (A.Z, F.M., with 4 and 3 years of experience in the neuroimaging field, respectively), who were blinded to the clinical information of the participants.In case of discordance, a third reader (M.S., with 14 years of experience in neuroimaging) was consulted.Intra and inter-rater concordance measures were employed, as described in the statistical analysis section.

Statistical analysis
All quantitative data were expressed as mean and standard deviation (SD).Differences in demographics, education, and cognition were analyzed using the Kruskal-Wallis test.
Each visual rating scale was initially dichotomized as normal or pathological according to Ferreira and Cotta Ramusino criteria.
To determine the cut-offs derived from the cohort, the visual rating scales were considered as continuous variables in the binomial logistic regression with amyloid status as the dependent variable.
The value of cohort-based cut-off with the highest Youden index was selected to evaluate the single performance and create a scoring model predictive of amyloid-PET status.
The first model included the MTA, PA, and API scales, while the second model additionally included the Fazekas score.A score of 1.5 was assigned to variables with an AUC > 0.70, 1 for those with an AUC > 0.65, and 0.5 for variables with an AUC > 0.55.The amyloid prediction performance of the models was then tested considering them as independent variables in the binomial logistic regression.
The aim was to evaluate (1) sensitivity and specificity values, (2) positive and negative predictive values (PPV -NPV), and (3) receiver operating characteristic (ROC) curve analyses along with their respective areas under the curve (AUC) and intervals of confidence of 95%.
The intra and inter-observer concordance were evaluated with Cohen's kappa (k) coefficient of agreement.All analyses were conducted using the open-source statistical software Jamovi v. 2.3.21.0.

Study population characteristics
Our study included a population of 153 patients with an average age of 70.13 years (SD: ± 8.26).The patients had an average education duration of 10.49 years (±4.65), and 80 individuals (52.3%) were female.The average MMSE score was 23.55 (±4.72).Of the total patients, 102 (66.7%) tested positive for amyloid-PET, while 51 (33.3%) tested negative.Table 1 illustrates the demographic characteristics of these two groups.
There were no significant differences between the amyloid-positive and amyloid-negative groups in terms of sex (p = 0.843), education (p = 0.683), or MMSE score (p = 0.751).However, the amyloid-positive patients were significantly older than the amyloid-negative group (p = 0.026).Within the AD continuum, 62 patients (60.78%) were diagnosed with MCI.Among them, 52 had the amnestic type of MCI, while 1 had the posterior cortical variant, 3 had corticobasal syndrome, 3 had the logopenic variant, 2 had the behavioral frontal variant, and 1 had DLB with AD copathology.
Additionally, 40 patients (39.22%) were diagnosed with AD dementia.Among the AD dementia patients, 34 had the amnestic type, 1 had the posterior cortical variant, 2 had the behavioral frontal variant, 2 had the logopenic variant, and 1 had DLB with AD copathology.
In the group of amyloid-negative patients, the mean overall MTA score was 1.53 (± 0.81), with a score of 1.41 (± 0.85) for the right side and 1.66 (± 0.97) for the left side.The PA score was 1.43 (± 0.60), and the GCA-F score was 1.54 (± 1.15).The mean Fazekas score, which assessed white matter hyperintensities, was 1.05 (± 0.94), and the mean API was − 0.11 (± 0.73).In the amyloid-positive group, the average MTA score was 1.79 (± 0.82), with a score of 1.81 (± 0.91) for the right side and 1.78 (± 0.86) for the left side.The PA score had a mean value of 1.74 (± 0.60), and the GCA-F score was 1.11 (± 0.60).The mean Fazekas score was 1.27 (± 0.87), and the mean API was 0.54 (± 0.62).The difference between the amyloid-positive and amyloid-negative groups was significant for the mean MTA (p = 0.035), right MTA (p = 0.003), API (p < 0.001), and PA (p = 0.005) scales.However, no significant differences existed for the left MTA, Fazekas score, and GCA-F scales.

Predictive performance of cohort-based cut-offs and API within the global cohort
Based on a cohort-based cut-off of ≥1.5, the mean MTA scale achieved an AUC of 0.583 (0.500-0.665), with a sensitivity of 67.65% and a specificity of 49.02%.A better performance was observed with a PA cutoff of ≥2, which resulted in an AUC of 0.672 (0.591-0.752), a sensitivity of 71.57%, and a specificity of 62.75%.
On the other hand, GCA-F scale exhibited the worst predictive performance, with an AUC of 0.450 (0.355-0.544), a sensitivity of 89.22%, and a low specificity of 15.69%.
The API scale with a cut-off ≥1 showed the best predictive performance for amyloid-positive status, with a positive result leading to an AUC of 0.721 (0.644-0.795), a good specificity of 76.47%, and a lower sensitivity of 67.65%.
Within predictive model 1, by assigning 1.5 points when encountering a pathological value of the API scale (due to its superior performance), 1 point in the case of PA scale alteration (due to its intermediate performance), and 0.5 points for a pathological MTA scale value (due to its comparatively lower performance), we developed a model, achieving a good accuracy with an AUC of 0.787 (0.714-0.859).
It is worth noting that employing a cut-off of ≥2 for this model demonstrated a robust specificity of 80.39%, while coupled with a lower sensitivity of 62.75%.
Furthermore, the incorporation of the Fazekas score analysis into the predictive model 2, with the assignment of 0.5 points for values ≥2, did not substantially enhance the model's performance, yielding an AUC of 0.795 (0.718-0.863).
Table 2 displays the predictive performance of cohort-based cut-offs, API scale, and models 1 and 2 within the global cohort.Fig. 2 shows the ROC curves of cohort-based cut-offs, API scale and models 1 and 2.
The differences in terms of performance between the cohort of patients who underwent brain CT and brain MRI are displayed in the supplementary materials.

Predictive performance of cohort-based cut-offs and API scale within MCI/Mild dementia cohort among individuals aged below 75 and 65 years
In this cohort (N = 93; 57 amyloid+, 36 amyloid-), the mean MTA scale yielded an AUC of 0.553 (0.447-0.658), accompanied by a sensitivity of 57.89% and a specificity of 52.78%.A more robust performance was observed with a PA cut-off of ≥2, exhibiting an AUC of 0.703 (0.606-0.799), a sensitivity of 68.42%, and a specificity of 72.22%.Conversely, GCA-F scale displayed the least favorable predictive performance in this cohort, registering an AUC of 0.499 (0.426-0.571), a sensitivity of 85.98%, and notably low specificity of 13.89%.
The API scale still demonstrated the most robust predictive performance for an amyloid + status.A positive result correlated with an AUC of 0.741 (0.652-0.829), showing a specificity of 83.33% and a lower sensitivity of 64.91%.The predictive model 1 obtained by incorporating API, PA, and MTA scales achieved favorable accuracy, resulting in an AUC of 0.813 (0.726-0.899).Notably, when the model's cut-off was set at ≥2 for the model, it demonstrated high specificity of 88.89% and a lower sensitivity of 59.65%.The inclusion of the Fazekas score in predictive model 2, assigning 0.5 points if it is ≥2, did not change the  overall model's performance, yielding an AUC of 0.813 (0.723-0.900).Table 3 summarizes the predictive performance of cohort-based cut-offs, API scale, and models 1 and 2. Fig. 2. displays the different ROC curves.

Predictive performance of Ferreira and Cotta Ramusino cut-offs
Table 4 summarizes the predictive performance of pathological visual rating scales based on the Ferreira and Cotta Ramusino criteria for detecting amyloid-positive PET scans.
Each visual rating scale demonstrated low performance both in the global and MCI-mild dementia groups, with an area under the curve (AUC) of <0.60.
The different cut-off values for PA and GCA-F scales, as defined by the Ferreira [18] and Cotta Ramusino criteria [23], resulted in divergent results in terms of specificity and sensitivity.The Cotta Ramusino criteria showed higher specificity, with 96.08% for PA and 100% for GCA-F.Conversely, the Ferreira criteria demonstrated higher sensitivity, with 97.06% for PA and 89.22% for GCA-F.

Discussion
Our study aimed to investigate the role of visual rating scales in predicting amyloid-PET positivity in a real-world memory clinic setting.The cohort-based cut-off identified in our work demonstrated significantly better accuracy than previously established normative values, which exhibited either low specificity [18] or sensitivity [23].Additionally, we introduced a new visual rating assessment called the anteroposterior index, which exhibited the best performance among individual visual rating scores.Moreover, we developed a predictive model that combined the main visual rating scores, resulting in excellent performance, particularly within the early-onset population.
One of the primary aims of this study was to evaluate the role of previously described normative cut-offs of visual rating scales in predicting AD-related pathology.While their ability to distinguish patients in the AD continuum from controls has been established previously, their role within a real memory clinic setting, encompassing various types of cognitive disorders, remains unexplored.
Our results showed that two of the most widely used normative values were unable to accurately differentiate AD-related pathology from other types of cognitive disorders in the real-world setting.These values struggled with validation in cohorts composed of strict inclusion criteria and healthy subjects as a comparative group.Furthermore, the absence of an analysis of AD-related pathology both in patients and controls precludes the determination of optimal values in terms of specificity and sensitivity.
Given the low accuracy of previously published normative values, we aimed to identify the best cut-offs based on a heterogeneous cohort for routine clinical practice.
Our cohort included individuals with a spectrum of different neurodegenerative conditions (e.g., AD, frontotemporal dementia, corticobasal syndrome) and psychiatric disorders (e.g., major depression), providing a reliable representation of the real-world setting where visual rating scales can be applied for diagnostic purposes.
Given the relatively low mean age of our population (71.28 years) and the prevalence of parieto-occipital atrophy in this age range [42], it is unsurprising that visual rating scores focusing on parieto-occipital areas, such as the PA and API scales, exhibited the highest accuracy in our cohort.Notably, the API scale (cut-off ≥1) proposed in our study achieved the highest accuracy in both the global (AUC = 0.721) and MCI/Mild dementia cohort (AUC = 0.741), reaching remarkable accuracy within the early-onset population (AUC = 0.857).
The API score is easy to administer, highly concordant among observers, independent of age, and does not require advanced imaging techniques, rendering it suitable for use in non-specialized outpatient units.
Compared to the PA assessment (cut-off ≥2), the API scale demonstrated slightly lower sensitivity (67.65% vs 71.57%) but a notable increase in specificity (76.47% vs 62.75%).This is because the API considers not only the parieto-occipital regions but also indirectly reflects the predilection of posterior regions in the AD-related neurodegenerative process compared to the anterior areas [43].Consequently, the PA assessment might indicate pathology even in some disorders beyond AD, which can affect both posterior and anterior regions more uniformly, such as psychiatric [44] and toxic disorders [45].
Despite being considered the most accurate measure of AD-related pathology in prior studies [46], the MTA scale did not achieve acceptable accuracy even with the cohort-based cut-offs (AUC = 0.583).
The assessment of atrophy in the medial temporal region might suffer from low accuracy due to the cortical involvement's predilection in this age group (71.28 ± 7.74).Indeed, the hippocampal sparing subtype has been observed to have a predilection for a younger age at onset compared to the limbic predominant or typical subtypes, which are characterized by the involvement of medial-temporal regions and a higher age at onset [47].
Furthermore, the specificity of MTA could be compromised by the potential of psychiatric and non-AD-related disorders to cause detrimental effects on the hippocampal areas [48,49].
By combining the main visual rating assessments into a predictive score model that accounted for their distinct predictive roles, we achieved good accuracy in the global (AUC = 0.787) and MCI/Mild dementia (AUC = 0.813) cohorts, with excellent accuracy within the earlyonset population (AUC = 0.949).
Including white matter hyperintensities (WMH) assessment did not substantially improve diagnostic prediction in global and different subcohorts (AUC = 0.795 vs 0.787).This result could be related to the absence of brain regional assessment of WMH in the Fazekas score, suggesting the potential for future incorporation of topography distribution of WMH into the scale, considering its established link with AD [50].
Upon comparing the performance within the global cohort and the MCI/Mild dementia cohort, we observed that the accuracy was consistently higher for the API scale and the predictive model in the latter group.
This finding could have two main explanations.First, patients in the advanced stage of the disease (e.g., above moderate dementia stage) might lose the typical pattern of atrophy characterizing AD, showing widespread involvement of associative and neocortical areas, even in the frontal regions as described by Braak stages [51,52].Second, considering patients below 75 years of age might have mitigated the influence of co-pathology on brain atrophy, reinforcing the consideration of a cohort primarily affected by the main effects of AD-related pathology [36,37].
Therefore, both the API scale and the predictive score models could play a pivotal role in outpatient settings due to their fast and straightforward use, particularly in the early stages of the disease and among individuals below 75 years of age.
These scales can be used to identify a group of selected patients who would benefit from more advanced and expensive diagnostic assessments, such as CSF and amyloid-PET, thus partially addressing the increasing demand for assessments resulting from the growing prevalence of AD over the years, especially considering the ongoing availability of monoclonal treatment [53,54].
From this perspective, the potential impact of visual rating scales on predicting amyloid-PET results is significant.Amyloid-PET enables the in vivo detection of β-amyloid protein, a fundamental constituent of neuritic plaques.When combined with appropriate clinical presentation, it facilitates earlier AD diagnosis [55] and extends to evaluating the risk of progression in MCI patients [56].Furthermore, its importance in AD detection is underscored by the latest proposed criteria for AD diagnosis [57].
The primary limitation of our study is its relatively small sample size.While the API and score models achieved good predictive accuracy, their accuracy requires confirmation in larger naturalistic cohort.A larger sample size would allow for more robust subgroup analyses, further solidifying our findings in early-onset patients (< 65 years) or exclusively among MCI patients.
Furthermore, the utilization of brain CT and MRI may be considered a potential study limitation; however, it is noteworthy that several studies have consistently reported high and comparable accuracy with both techniques [6,24].
In conclusion, we acknowledge that, focusing on a posterior pattern of atrophy, the API scale may overlook the behavioral frontal variant of AD, which has an opposite pattern of atrophy.However, our cohort presented few cases of the frontal variant, respecting the low overall prevalence of this rare disorder [58].

Conclusion
Our study highlights the significance of visual rating scales in predicting amyloid-PET positivity within a real-world memory clinic context.Notably, the newly introduced antero-posterior index has emerged as the most effective predictor of AD-related pathology, holding substantial potential to enhance the AD diagnostic process, especially when integrated with the cohort-based cut-offs of MTA and PA scales and in the mildest stages of the disease.

Fig. 1 .
Fig. 1.On the left, a brain CT scan with GCA-F = 0 and GCA-P = 1, resulting in an API of 1.On the right, a brain MRI with GCA-F = 1 and GCA-P = 2, resulting in an API of 1.

Fig. 2 .
Fig. 2. ROC Curves of Cohort-Based Cut-offs and API scale in the Global Cohort (on the left) and the MCI/Early Dementia Cohort below 75 Years of Age (on the right).

Table 1
Summary of population characteristics.

Table 2
Predictive performance of cohort-based cut-offs, API scale and model 1 and 2 within the global cohort.

Table 3
Predictive performance of cohort-based cut-offs, API scale and model 1 and 2 in the MCI and Mild Dementia below 75 of age.

Table 4
Predictive performance of Ferreira and Cotta Ramusino cut-offs within the global cohort.