Introduction

Alzheimer’s disease (AD) is a progressive and fatal neurodegenerative disorder and is the most common cause of dementia in the elderly population. An estimator for the global number of people with dementia is 55 million in 2019, and this number is expected to increase to 78 million by 2030 and to 139 million by 2050. The total estimated annual worldwide cost of dementia was over 1.3 trillion US dollars in 2021, and this figure is projected to be 2.8 trillion US dollars by 2030 [1]. More than 6 million adults aged 65 and older in the United States (US) are living with Alzheimer’s disease in 2021, and this number is projected to rise to nearly 13 million in 2050 [2]. In 2021, Alzheimer’s disease and other dementias are expected to cost the US $355 billion, and this amount could rise to $1.1 trillion by 2050. The financial burden of AD globally is potentially huge for human society.

Laboratory and clinical evidence support the concept that an imbalance between the production and clearance of Aβ protein in the brain is an early and often initiating factor in AD [3,4,5]. Research and clinical studies on Aβ removal and its potential for slowing the progression of cognitive and functional decline in patients with Alzheimer’s disease has led to the development of multiple disease-modifying therapies, including monoclonal antibodies directed to bind and remove Aβ plaques in the brain, such as aducanumab [6], lecanemab [7], donanemab [8], and gantenerumab [9]. In June of 2021, the US Food and Drug Administration (FDA) conditionally approved aducanumab (Aduhelm™), through the FDA-Accelerated Approval Program, as the first potentially disease-modifying medicine to treat patients with Alzheimer’s disease based on biomarker endpoint results in Phase 3 trials [10]. This Accelerated Approval triggered a worldwide discussion in the amyloid therapeutic area for Alzheimer’s disease, and several biopharmaceutical sponsors are developing similar strategies for potential regulatory submissions for their amyloid therapeutic agents including lecanemab [7], donanemab [8], and gantenerumab [9]. From the scientific perspective, a unified analysis of published clinical trials’ data in this field may provide some insight to answer important scientific questions concerning amyloid therapies for Alzheimer’s disease. For this article, we have extracted currently available, published summary statistics for clinical trials for aducanumab, lecanemab, donanemab, and gantenerumab. We then perform appropriate meta-analyses to assess relationships among such summary statistics for amyloid removal rate, Amyloid-Related Imaging Abnormality-Edema (ARIA-E) rate, and clinical cognitive response as efficacy, and these relationships are potentially informative for understanding the role of dose for amyloid therapies. However, the results for these relationships are not intended to have a causal interpretation nor to pertain to whether the amyloid removal rate might be a surrogate endpoint for clinical cognitive response, mainly because they are based on summary statistics for the clinical trials in the meta-analyses rather than patient-level data.

Methods

Data Extraction Methods for Analysis

Analysis datasets were created from published studies as presented in Table 1. Summary statistics (mean and standard errors, ARIA-E rate) were used for all analyses in this article. Available numerical results of variables of interest (e.g., mean amyloid reduction from baseline and standard error, numerator and denominator for ARIA-E rate, mean Clinical Dementia-Rating Scale Sum of Boxes (CDR-SB) score change from baseline and standard error) were used directly from the published materials. For numerical results that were not available in the published materials, we extracted the needed information from figures or graphs that were available in the published materials using measurement tools built in Nuance Power PDF Advanced© software version 3.0 [11]. Some studies only reported positron emission tomography standardized uptake value ratio (PET SUVR) values and standard errors of SUVR values for amyloid reduction. For these studies, SUVR values and standard errors were converted to amyloid centiloid values and their corresponding standard errors.

Table 1 Studies Included in the Analysis
Full size table

Initial Method for ARIA-E Analysis

The sample sizes and numbers of patients with ARIA-E are shown in Table 2 for the dose groups for the four compounds within the clinical trials, and they correspond to 22 treatment groups in all. For the lecanemab G000-201 study, there is initial pooling for the 2.5 mg/kg biweekly and 5 mg/kg monthly groups (since each only has one patient with ARIA-E), and for the aducanumab Prime study, there is initial pooling of the 1 mg/kg and 3 mg/kg groups (since they, respectively, only have 1 and 2 patients with ARIA-E). There is also initial pooling for the G000-201, Prime, Scarlet RoAD, and Trailblazer-alz clinical trials because all of their placebo groups have ARIA-E event rates that are less than 0.01 (via two or fewer patients with ARIA-E events). Subsequent to these initial poolings, meta-analysis is applied with a logistic regression model that includes study (for 301, 302, and all others) and dose group for the distinct treatment groups. Since study (with p = 0.79) is not a statistically significant factor in this model, a simpler model without study is applied. This simpler model invokes pooling of the placebo groups for all of the studies and the pooling of the aducanumab studies 301, 302, and Prime for the 6 mg/kg dose and the 10 mg/kg dose. For the logistic regression model for the 11 distinct dose groups that result from the previously indicated poolings, there is support for the additional pooling of the aducanumab 1 mg/kg and 3 mg/kg groups with the lecanemab 2.5 mg/kg biweekly and 5 mg/kg monthly groups (p = 0.318) and the pooling of the lecanemab 5 mg/kg biweekly group with the lecanemab 10 mg/kg monthly group (p = 0.059), although the latter pooling has further justification by both of these treatment groups corresponding to a monthly dose of 10 mg/kg. Thus, the initial meta-analyses of the ARIA-E rates for the 22 treatment groups in Table 2 with logistic regression models lead to the following nine dose groups for consideration with the meta-analyses for the amyloid removal rate. These dose groups are presented as the following: Control: pooled placebo groups across all studies; G105: gantenernumab Scarlet RoAD 105 mg dose group; G225: gantenernumab Scarlet RoAD 225 mg group; L2.5BW, L5M, A13: pooled group from lecanemab G000-201 study 2.5 mg/kg biweekly and 5 mg/kg monthly with aducanumab Prime study 1 mg/kg and 3 mg/kg dose groups; A6: pooled group from aducanumab Studies 301, 302, and Prime 6 mg/kg dose groups; A10: pooled group from aducanumab Studies 301, 302, and Prime 10 mg/kg dose groups; L5BW, L10M: pooled group from lecanemab G000-201 study 5 mg/kg biweekly and 10 mg/kg monthly dose groups; L10BW: lecanemab G000-201 study 10 mg/kg biweekly dose group; D: donanemab 1400 mg dose group.

Table 2 ARIA-E Rate for Dose Groups Within Clinical Trials for the Four Compounds
Full size table

Amyloid Reduction Analysis Method

Amyloid centiloid value change from baseline over time is meta-analyzed with a mixed effects model with repeated measurements for visits and random study effects [12]. The data structure for this analysis is shown in Table 3. As shown there, amyloid centiloid values are available at Week 26 and Week 54 for five of the 22 treatment groups; at Week 26 and Week 78 for six of the 22 treatment groups; at Week 48 and Week 72 for six of the 22 treatment groups; at Week 24, Week 52, and Week 76 for two of the 22 treatment groups; and at Week 20, Week 60, and Week 100 for three of the 22 treatment groups. Also, all 22 treatment groups have amyloid centiloid values at baseline. Since the 22 treatment groups have different schedules for the measurements of amyloid centiloid values, the role of the mixed model is to produce estimates for all of them at Week 26, Week 52, and Week 78 through a structure with potentially different linear trends for baseline to Week 26, Week 26 to Week 52, Week 52 to Week 78, and Week 78 to Week 100. Accordingly, the preliminary model has the following specification:

$$y_{ijt} = \alpha + s_{i} + \beta_{0} x_{ij0} + \beta_{0j} x_{j} + \beta_{1} w_{1ijt} + \beta_{2} w_{2ijt} + \beta_{3} w_{3ijt} + \beta_{4j} w_{1ijt} x_{j} + \beta_{5j} w_{2ijt} x_{j} + \beta_{6j} w_{3ijt} x_{j} + \varepsilon_{ijt} .$$
Table 3 Available Visits for Repeated Measurements for Treatment Groups Within Studies
Full size table

For this model, \(y_{ijt}\) is the amyloid centiloid value change from baseline for study \(i = 1, 2, 3, 4, 5, 6\) for the studies in Table 1; dose group \(j = 1, 2, 3, 4, 5, 6, 7, 8, 9\) as summarized at the end of the previous section; at visit \(t = 20, 24, 26, 48, \ldots , 100\) as shown in Table 3; \(\alpha\) is the fixed intercept, and \(s_{i}\) is the random effect for study \(i\) on the intercept; the random study effects \(s_{i}\) are assumed to have independent \(N\left( {0,\sigma_{s}^{2} } \right)\) normal distributions. The baseline amyloid centiloid value is \(x_{ij0}\) for dose group \(j\) in Study \(i\), and \(\beta_{0}\) is its slope; \(x_{j}\) is the indicator variable for dose group \(j\) versus placebo, and \(\beta_{0j}\) is the corresponding treatment effect relative to placebo; \(w_{1ijt} , w_{2ijt} ,\) and \(w_{3ijt}\) are derived variables as shown in Table 4 for visit weeks for study \(i\) and dose group \(j\) at Week \(t,\) and they have coefficients \(\beta_{1} , \beta_{2} ,\) and \(\beta_{3}\) such that \((\beta_{1} + \beta_{2} + \beta_{3} )\) represents the linear trend between baseline and Week 26, \(\beta_{1}\) represents the linear trend between Week 26 and Week 52, \(\beta_{2}\) represents the linear trend between Week 52 and Week 78, and \(\beta_{3}\) represents the linear trend between Week 78 and Week 100; the \(w_{1ijt} x_{j} , w_{2ijt} x_{j}\), and \(w_{3ijt} x_{j}\) represent interaction terms with the \(x_{j} s\) for the corresponding dose groups, and they respectively have corresponding coefficients \(\beta_{4j} , \beta_{5j} ,\) and \(\beta_{6j}\) that represent the influence of the jth dose group relative to placebo on the linear trends that correspond to \(\beta_{1} ,\beta_{2}\), and \(\beta_{3}\). The random error terms \(\varepsilon_{ijt}\) are assumed to have independent \(N\left( {0,\sigma_{ei}^{2} } \right)\) distributions for the corresponding studies. For the model selection process with backward elimination, there is support for the removal from the model of \(\beta_{3} w_{3ijt}\), and the interaction terms for \(\beta_{5j} w_{2ijt} x_{j}\) and \(\beta_{6j} w_{3ijt} x_{j}\). Thus, the final model for amyloid centiloid value change from baseline is as follows:

$$\begin{array}{*{20}c} {y_{ijt} = \alpha + s_{i} + \beta_{0} x_{ij0} + \beta_{0j} x_{j} + \beta_{1} w_{1ijt} + \beta_{2} w_{2ijt} + \beta_{4j} w_{1ijt} x_{j} + \varepsilon_{ijt} } \\ \end{array}.$$
Table 4 Derived Visit Variables w1, w2, and w3 from Actual Visit Weeks for the Linear Model for Amyloid Centiloid Value Change from Baseline
Full size table

The analysis for this model is used to estimate predicted means for amyloid centiloid value change from baseline at Week 26, Week 52, and Week 78 for the 22 treatment groups in Table 3. Since the structure of Table 4 implies that the difference between the predicted mean at Week 78 and that at Week 52 for all 22 treatment groups is \(26\beta_{2}\), the analysis for the relationships for amyloid centiloid value change with ARIA-E and clinical response only addresses the roles for the predicted means of amyloid centiloid value change at Week 26 and at Week 52. Also, graphical displays are produced to describe how well the model predicts the observed means for amyloid reduction at the times for their determination.

Methods for Associations Between ARIA-E and Amyloid Reduction

Two methods are applied to evaluate the relationships between the amyloid reduction rate and the ARIA-E rate. For descriptive purposes, Spearman rank correlations are calculated between the ARIA-E rates in Table 2 for the 22 treatment groups and the predicted means at Week 26 and Week 52 for amyloid reduction from the methods described in the previous section. Additionally, a logistic regression model is applied to evaluate how well the variation among the ARIA-E rates for the 22 treatment groups can be explained by the variation among the predicted means for amyloid reduction at Week 26 and Week 52. The final model for this purpose includes the Week 52 predicted mean for amyloid reduction and an indicator variable which is equal to 1 for Study 301 or Study 302 and equal to 0 for all other studies. The indicator for Study 301 or Study 302 is included in the model because the predicted means for amyloid reduction at Week 52 are from a mixed model that includes study as a random effect; and the indicator variable invokes pooling for Studies 301 and 302 because of the similarity of their ARIA-E rates in Table 2 for placebo, aducanumab 6 mg/kg, and aducanumab 10 mg/kg. Graphical displays are produced to describe how well the logistic regression model explains the variation among the ARIA-E rates for the 22 treatment groups.

Clinical Response Analysis Method

Since the schedules for the measurements of CDR-SB for the 22 treatment groups are the same as those shown in Table 3 for the amyloid centiloid values, the first stage of analysis for CDR-SB change from baseline uses a mixed model with repeated measurements for visits and random study effects in ways that are similar to the previously described use of such models for amyloid centiloid value change from baseline. The structure of the resulting final model is essentially the same as that for the amyloid centiloid value change from baseline, except for the replacement of \(\beta_{2} w_{2ijt}\) by \(\tilde{\beta }_{2} \tilde{w}_{2ijt}\), where \(\tilde{w}_{2ijt} = 0\) if Week \(\le 52\) for the measurement of CDR-SB and \(\tilde{w}_{2ijt} = \left( {{\text{Week}} - 52} \right)\) if Week \(\ge 52\). From the analysis for this model, predicted means of CDR-SB change from baseline are estimated for Week 26, Week 52, and Week 78 for the 22 treatment groups in Table 3, with those at Week 78 being of primary clinical interest. Also, a graphical display is produced to describe how well this model predicted the observed means for CDR-SB change from baseline at the times for their determination for the 22 treatment groups.

For evaluation of the relationships between amyloid reduction and CDR-SB change from baseline, two methods are applied. For descriptive purposes, the Spearman rank correlation is produced between the predicted means at Week 52 for amyloid reduction and the predicted means at Week 78 for CDR-SB change from baseline for the 22 treatment groups in Table 3. Additionally, the previously described mixed model for CDR-SB change from baseline is modified to have the variables corresponding to dose groups (i.e., the \(x_{j}\) and the \(w_{1ijt} x_{j}\)) replaced by the predicted means at Week 52 for amyloid reduction and additionally to include the interactions of that explanatory variable with both the \(w_{1ijt}\) and the \(\tilde{w}_{2ijt}\) that correspond to weeks for measurements. How well this model explains the variation among the predicted means for CDR-SB change from baseline at Week 78 for the 22 treatment groups is described with a graphical display.

All analyses in this paper are performed using SAS 9.4 software [13]. Example SAS code using PROC MIXED procedure for the meta-analysis is included in Appendix.

Results

A total of 6 studies and 22 treatment groups within them are included in the meta-analyses. Detailed information for the studies and the treatment groups are provided in Tables 1 (and its cited references), 2, and 3.

ARIA-E Analysis Results

The observed ARIA-E rates across treatment groups are shown in the last column of Table 2. Within each study, the ARIA-E rate is dose exposure dependent, i.e., higher ARIA-E rates are evident at higher dose levels. The odds ratios (ORs) and their 95% confidence intervals from the meta-analysis with logistic regression are shown in Table 5 for the comparisons to placebo for each of the other eight dose groups.

Table 5 Odds Ratio (OR) of ARIA-E Event for Different Dose Groups
Full size table

Except for the low dose group L2.5BW, L5M, A13, the 95% confidence intervals for the odds ratios of ARIA-E event for the other dose groups, relative to the pooled placebo group, exclude 1. The OR for dose group A10 is 24.75, which is the highest among the dose groups included in this analysis. The OR for the donanemab dose group (D) is more than three times that of the OR for the lecanemab L10BW dose group, even though the total exposure seems similar (i.e., 20 mg/kg for a month) between these two dose groups. The reason for this difference could be the different dose schedules or the difference between the compounds. The lecanemab L10BW group has a biweekly dosing schedule, and the donanemab D group has a monthly dosing schedule titrated up to 1400 mg after 700 mg in the first 3 doses.

Amyloid Centiloid Value Reduction Analysis Results

The observed means for amyloid centiloid value reduction over time across the 22 treatment groups within the 6 studies are presented in Fig. 1. Within each study, the means for amyloid centiloid value reduction from baseline are dose exposure dependent, i.e., larger means for amyloid centiloid value reduction are evident at higher doses within each study. Also, the cumulative amyloid centiloid value reduction becomes larger over time within each treatment group. Figure 2 shows that the mixed model with repeated measurements for visits and random study effects provides predicted means for amyloid removal with reasonably good fit to the observed means for amyloid removal values. The model predicted means for amyloid reduction for the 9 dose groups at Week 26, Week 52, and Week 78 are presented in Table 6.

Fig. 1
figure 1

Observed means for amyloid centiloid value reduction over time across treatment groups in different studies

Full size image
Fig. 2
figure 2

Consistency of predicted mean amyloid reduction versus observed mean amyloid reduction. Labels for data points represent dose groups. The diagonal line is a reference line where observed equals to predicted values

Full size image
Table 6 Predicted Means for Amyloid Reduction over Time
Full size table

Consistent with the trends for the observed means for amyloid reduction, the donanemab D group has the largest predicted mean for amyloid reduction during the first 26 weeks; the aducanumab 10 mg/kg (A10) and 6 mg/kg (A6) dose groups have the second and third largest predicted means for amyloid reduction during the first 26 weeks, followed by the lecanemab 10 mg/kg biweekly group for which the very large standard error is due to the absence of an observed mean amyloid reduction at Week 26 in the G000-201 study. As might be expected, there is no apparent amyloid removal for the control group over time.

Relationship Between the Amyloid Removal Rate and the ARIA-E Rate

The Spearman rank correlation coefficients between the ARIA-E rates and the predicted least-squares (LS) means of amyloid centiloid value reduction at Week 26, Week 52, and the LS mean differences between Week 52 and Week 26 are − 0.69, − 0.79, and − 0.67, respectively, all with \(p < 0.001\). For the logistic regression analysis, the simplest model only includes the predicted means for amyloid centiloid value reduction at Week 52 since it nearly is a sufficient explanatory variable for the ARIA-E rate. Figure 3 shows the consistency of the model predicted ARIA-E rates with the observed ARIA-E rates for this logistic regression model. Alternatively, Fig. 4 presents the predicted ARIA-E rates from a logistic regression model that includes study (as pooled 301 and 302 versus the pooled other studies) and the predicted means for amyloid centiloid value reduction at Week 52. As shown in both Figs. 3 and 4, the observed ARIA-E rates for the A6 and A10 dose groups in the PRIME study are outliers. In contrast to the observed ARIA-E rates for the same dose groups (A6 and A10) in the 301 and 302 studies, the observed ARIA-E rates in the PRIME study are not stable due to the small sample sizes. More generally, a clear trend is evident in Figs. 3 and 4 for larger predicted means for amyloid centiloid value reduction at Week 52 to correspond to higher ARIA-E event rates.

Fig. 3
figure 3

Consistency of observed ARIA-E rates and predicted ARIA-E rates from the logistic regression model with the predicted mean for amyloid centiloid value reduction at Week 52 as the explanatory variable. Labels for data points represent dose groups. The diagonal line is a reference line where observed is equal to predicted values

Full size image
Fig. 4
figure 4

Prediction of ARIA-E rates from a model that includes predicted mean amyloid reduction at Week 52 and studies in the model. Pooled studies 301 and 302 (gray solid upper triangle) are compared with the other pooled studies (black solid circle)

Full size image

Relationship Between Amyloid Removal Rate and Clinical Efficacy Response

The observed means for longitudinal CDR-SB score changes from baseline for the 22 treatment groups are described in Fig. 5 where lower values of CDR-SB score change from baseline are preferable to higher values. An effective treatment is expected to slow the disease progression, with this being equivalent to a smaller observed mean CDR-SB score change from baseline. Thus, a more negative difference between a treated group and the placebo group (as treated – placebo) represents a better treatment effect. As shown in Fig. 6, the mixed model with repeated measurements for visits and random study effects produces predicted means for CDR-SB score change from baseline that tends to increase for all nine dose groups over time. The predicted LS mean CDR-SB score changes from baseline in the control and L2.5BW, L5M, and A13 dose groups have similar increases; and those increases are larger than those for the other dose groups, particularly Groups D, A10, and L10BW for which the increases are clearly smaller than for the control group. The previously noted interpretations are also well supported by the predicted LS means for CDR-SB score changes from baseline at Week 26, Week 52, and Week 78 that are shown in Table 7. How well larger predicted means for amyloid centiloid value reduction at Week 52 correspond to less decline for the predicted means of CDR-SB score changes from baseline at Week 78 is shown in Fig. 7. Dose groups D, A10, and L10BW have the smallest predicted means for CDR-SB score increase at Week 78 relative to the other dose groups, especially relative to the placebo control group. The trends in Fig. 7 indicate that the predicted mean for amyloid centiloid value reduction at Week 52 is strongly correlated with the clinical treatment response at Week 78 as represented by the predicted mean for CDR-SB score change from baseline. The Spearman rank correlation coefficient between the predicted mean for amyloid centiloid value reduction at Week 52 (LS mean) and the predicted mean for CDR-SB score change from baseline at Week 78 (LS mean) is 0.73 \((p < 0.001)\) for the 22 treatment groups in Table 3. Thus, larger amyloid reduction at Week 52 (as more negative change from baseline) corresponds to slower cognition decline in CDR-SB score change from baseline.

Fig. 5
figure 5

Longitudinal observed means CDR-SB score change from baseline across treatment groups in different studies

Full size image
Fig. 6
figure 6

Predicted means for CDR-SB score change from baseline over time for dose groups

Full size image
Table 7 Predicted Means for CDR-SB Score Change over Time Comparing to Control
Full size table
Fig. 7
figure 7

Predicted means for CDR-SB score change from baseline at Week 78 from the model where the predicted mean of amyloid reduction at Week 52 is an explanatory variable

Full size image

Discussion and Conclusions

This article reports the results of meta-analyses based on summary statistics extracted from published data for clinical trials in amyloid therapeutics for patients with early Alzheimer’s disease. These meta-analyses evaluate the relationships among the amyloid removal rate in terms of amyloid centiloid value reduction, the ARIA-E rate, and the clinical efficacy response in CDR-SB score change from baseline in the target patient population. In general, the magnitude and the extent of brain Aβ plaques removal seem to depend on therapeutic agents and their dose exposure levels. For the reported meta-analyses, higher dose levels tend to provide more amyloid removal in terms of the observed means for amyloid centiloid value reduction for each of the therapeutic agents in the corresponding studies. Among the investigated amyloid therapeutic agents in this paper, donanemab and lecanemab 10 mg/kg biweekly have predicted means at Week 52 that correspond to more removal of brain Aβ plaques than the other agents, and their administration is at seemingly similar therapeutic dose exposures. Such predicted means for brain Aβ plaques removal in early Alzheimer’s disease patients are strongly correlated with higher rates of ARIA-E events, and the rates of ARIA-E events are larger for larger dose exposures. The predicted means for amyloid removal at Week 52 are also positively correlated with the predicted means for clinical efficacy response assessed at 78 weeks, i.e., more amyloid centiloid value reduction at Week 52 corresponds to better clinical response at Week 78 in terms of CDR-SB score change from baseline. However, for aducanemab at 10 mg/kg, the predicted means for the clinical efficacy response in terms of CDR-SB score change from baseline did not differ as much from placebo at Week 26 as at Week 52 or Week 78, even though the amyloid removal difference was substantial at Week 26. Such considerations suggest that the clinical efficacy response difference from placebo control for a therapeutic dose level of some compounds might be delayed to Week 52 or Week 78 after the appearance of substantial amyloid removal at Week 26.

For the meta-analyses to fit models to the summary statistics from the longitudinal data for amyloid centiloid value reduction and CDR-SB score change from baseline, the numerical values of estimated parameters vary somewhat among orders of data-sorting procedures. This consideration could suggest that the likelihood surface of the meta-analysis for this dataset is relatively flat so that the optimal estimates for parameters could not be obtained. Nevertheless, the predicted means from the mixed models for the amyloid centiloid value reductions and the CDR-SB score change from baseline for the dose groups are relatively similar for the alternative sorting procedures, and so are their interpretations. The analysis results in this paper are based on the dataset with sorting by Study, Treatment group, and Week variables, respectively. This sorting specification satisfies criteria for better model fitting statistics (smaller is better) in terms of Akaike Information Criterion (AIC), AIC corrected for small sample sizes (AICC), and Bayesian Information Criterion (BIC) versus other possible sorting specifications for the dataset.

A limitation of the results from the meta-analyses presented in this paper is that they are based on published summary statistics that are available from clinical trials information in the public domains. Also, only six studies are included in the analyses, with this scope being limited by the number of amyloid therapeutic agents currently ongoing investigation. As noted for Table 3, the schedules for assessments of amyloid removal and clinical efficacy differ for the six studies, and the extent of such differences increases the variability for the predicted means at unobserved time points. Nevertheless, consistent patterns are noted across different therapeutic agents with respect to relationships among the predicted means for the amyloid removal rate, the ARIA-E rate, and the predicted means for the clinical response in CDR-SB score change over time. Such relationships need further evaluation with additional clinical trials data for amyloid therapeutic agents as they become available in the future, especially from the datasets that will be produced from the currently ongoing large phase 3 clinical trials, such as the donanemab TRAILBLAZER-ALZ 2 study [ClinicalTrials.gov identifier: NCT04437511] [14], the lecanemab CLARITY study [ClinicalTrials.gov identifier:NCT03887455] [15], and the gantenernumab GRADUATE 1 & 2 studies [ClinicalTrials.gov identifier: NCT03443973, NCT03444870] [16, 17]. When available, subject level data from these and other studies may provide very useful information for further investigation of the relationships among the amyloid removal rate, the ARIA-E event rate, and the clinical efficacy response. As indicated at the end of the Introduction, the reported relationships for summary statistics from the studies included in the meta-analyses in this article are not intended to have a causal interpretation, although they are considered to be potentially informative for understanding the roles of the doses for the amyloid therapies in those studies. For future confirmatory studies, they shed light on the relevance for the balance between clinical efficacy response and safety in the ARIA-E rate for the choice of dose for amyloid therapies.