More efficient screening for preclinical Alzheimer's clinical trials using mixture of experts Authors: Oliver Langford, Rema Raman, Paul Aisen, Reisa Sperling, Keith Johnson, Doris Molina‐Henry, Pallavi Sachdev, David Li, Gustavo A. Jimenez‐Maggiora, Sterling C. Johnson, Joel B Braunstein, Colin Birkenbihl, Madison Cuppels, Rachel F. Buckley, Robert A. Rissman, Michael C. Donohue Published: 2026 ABSTRACT AbstractINTRODUCTIONBlood plasma biomarkers identifying Alzheimer's disease (AD) neuropathology offer accessible and scalable alternatives to lumbar puncture and positron emission tomography (PET) scans, with potential efficiency gains in clinical trial recruitment.METHODSWe evaluated the impact of a blood‐based screening algorithm on recruitment for the AHEAD 3‐45 trial testing lecanemab in preclinical AD. The algorithm was developed during initial screening without blood plasma and subsequently deployed using a Mixture of Experts prediction model to estimate amyloid PET positivity; analyses reflect prospective enrichment and model characterization.RESULTSThe algorithm incorporating amyloid beta A β42/ A β40 and subsequently adding percent phosphorylated tau 217 (%p‐tau217), reduced ineligibility on amyloid PET from 71% to 31%. Latent class analysis identified low, intermediate, and high amyloid groups. Model‐based analyses indicated %p‐tau217 predicts High amyloid group, whereas A β42/ A β40 was more specific for the low group.DISCUSSIONBlood plasma screening reduced participant and site burden, while preserving enrichment for amyloid positivity for a preclinical AD trial. INTRODUCTION Approved amyloid lowering therapies (e.g. lecanemab and donanemab) have been shown to slow the clinical progression of symptomatic Alzheimer's disease (AD) by about 30%. 1 , 2 Clinical trials are currently underway to test the hypothesis that treating AD pathology at an earlier, pre‐symptomatic stage will lead to even better efficacy. 3 , 4 , 5 These “Preclinical AD” trials require screening large numbers of cognitively unimpaired older adults to identify those with amyloid plaque accumulation. Previous cognitively unimpaired clinical trial screening data indicate that the prevalence of amyloid accumulation in individuals 65 years of age or older is ≈30%, yielding screen fail rates of more than 70% when coupled with other trial criteria. 6 The current in vivo gold standard test to detect the presence of brain amyloid is positron emission tomography (PET) using amyloid tracers, which is expensive and time consuming. Much progress has been made in recent years to advance blood‐based biomarkers to detect amyloid accumulation, facilitating the identification of candidates for theapeutic intervention. Promising blood plasma biomarkers, such as phosphorylated tau 217 (p‐tau217) and amyloid beta A β42/ A β40, have been used in combination with other variables to develop prediction models for amyloid status. Some authors have used logistic regression to predict binary amyloid status 7 , 8 , 9 and others compare logistic regression to machine learning models like random forests, support vector machine, k‐nearest neighbors, extreme gradient boosting, and gradient boosting machine. 10 , 11 Each of these models are based on predicting binary classification of the disease. We adopted a continuous modeling approach for amyloid PET Centiloid levels to better capture the continuum of disease and improve prediction. Given that brain amyloid PET data often exhibit a bimodal (or even multi‐modal) distribution, it suggests the presence of latent subgroups with qualitatively different amyloid burden states. Although this could be approached using a two‐class Gaussian mixture model or logistic regression with binarized amyloid PET status, these approaches impose important limitations. Gaussian mixture models assume fixed class membership probabilities independent of clinical covariates, and logistic regression (while flexible in modeling a single decision boundary) cannot capture subgroup‐specific relationships between predictors and outcome. Mixture of Experts (MoE) models address these limitations by jointly learning (1) multiple expert models, each representing a distinct latent subgroup with its own predictor–outcome relationship, and (2) a gating function that allows covariates to inform class membership probabilities. 12 In practical terms, this enables clinically relevant heterogeneity to be modeled explicitly, allowing predictors to influence both amyloid status and the probability of belonging to a given subpopulation. Related mixture‐based approaches have been applied successfully in biomedical and clinical prediction settings to model disease heterogeneity and improve risk stratification, motivating their use here for amyloid PET modeling. In this research, we systematically evaluate the impact of the blood plasma screening algorithm on recruitment in the AHEAD 3‐45 study testing lecanemab for preclinical AD. 4 AHEAD 3‐45 is a study composed of two randomized trials of lecanemab versus placebo: A3 for individuals with intermediate amyloid PET levels (20 to 40 Centiloids), and A45 for individuals with elevated amyloid PET levels (>40 Centiloids). We describe the derivation and implementation of the MoE algorithm, which was trained to predict amyloid PET levels using blood‐based biomarkers. Individuals predicted to be at higher risk for having intermediate or elevated amyloid levels advanced to a confirmatory amyloid PET for enrollment into the trial. Using this two‐stage screening approach, we were able to greatly reduce the number of costly and burdensome PET scans while maintaining a desired level of specificity. The algorithm was iteratively refined during trial enrollment to leverage advances in blood plasma biomarker development—such as the introduction of %p‐tau217—and to incorporate insights from its observed performance, as will be described. METHODS AHEAD study AHEAD 3‐45 is a Phase 3, international, multi‐center, double‐blind, placebo‐controlled study investigating the efficacy and safety of the monoclonal antibody lecanemab (BAN2401) in cognitively unimpaired individuals with evidence of amyloid pathology, defining preclinical AD. The study comprises two trials, A3 and A45, each with a duration of 216 weeks. 4 The A3 and A45 trials enrolled participants at different levels of amyloid pathology and have different primary outcomes. A45 is the larger trial, designed to recruit at least 1000 participants with baseline amyloid PET levels of 40 Centiloids or higher. A45 targets a later disease stage, and the population in this trial is considered more likely to experience cognitive decline within the 4‐year trial period. Consequently, the primary outcome for A45 is the Pre‐clinical Alzheimer's Cognitive Composite score (PACC5). 13 In contrast, the smaller A3 trial aimed to recruit 400 participants with amyloid PET levels between 20 and 40 Centiloids. Although both populations are thought to be at increased risk of cognitive decline, the A3 population was selected to be at an earlier stage of AD pathology and is considered less likely to experience observable cognitive decline over trial period. Therefore, the primary outcome for A3 is amyloid PET, with tau PET as the key secondary outcome. Informed consent was obtained for all participants in the AHEAD 3‐45 study. Evolution of AHEAD screening process The AHEAD study commenced trial recruitment in July 2020. Identifying eligible participants for the study required five intensive screening visits, as detailed in Table S1. The first visit consisted of many cognitive and functional assessments, as well as blood and urine samples for vitals and drug screening. If the participant met the inclusion/exclusion criteria, they proceeded to the second screening visit for a confirmatory amyloid PET scan. Based on data from the initial 1080 participants with both blood plasma and amyloid PET, we developed and deployed blood plasma Screening Algorithm Version 1 in February 2022. At the same time, the study team split screening Visit 1 into two visits, 1a and 1b (as shown in Table S1) so that only individuals who were likely to proceed to amyloid PET would undergo the additional cognitive and medical screening. The blood plasma screening algorithm was devised with the aim of reducing costs and burden associated with PET scans on sites and participants. By utilizing blood plasma screening, we aimed to minimize the number of assessments and procedures during the initial visit, therefore reducing the number of participants that carried out the assessments and procedures. The initial plasma‐based algorithm incorporated Aβ42 / Aβ40, age, and apolipoprotein E (APOE) to predict individual amyloid PET values. Eligibility to continue screening required a predicted Centiloid value of at least 11 Centiloids. As the trial continued screening, promising blood plasma biomarker %p‐tau217 (100 × (p‐tau217/np‐tau217)) was evaluated, with a significant increase in predictive accuracy. 14 Consequently, Screening Algorithm 2 included %p‐tau217 as a predictor and was introduced in May 2023, with eligibility determined by meeting a predicted centiloid threshold of 18 or more. As the trial recruitment concluded, Screening Algorithm 3 was introduced with an elevated threshold of 40 and sites were required to only screen participants 65 years of age or older, to further favor enrollment into the larger A45 trial. For Algorithms 2 and 3, we also introduced alternative prediction models to address cases with missing plasma data. All three algorithm versions were developed and deployed prospectively during active recruitment for the study. A timeline of AHEAD 3‐45 recruitment and introduction of plasma biomarker screening can be seen in Figure S1, with details and results of the prediction models in the Table S2. All blood plasma biomarker, including APOE, were determined using liquid chromatography‐tandem mass spectrometry by C2N Diagnostics, LLC (St. Louis, MO, USA). Statistical methods Mixture of Experts (MoE) Standard mixture models divide heterogeneous data into more homogeneous normal distributions. MoE, an extension of the mixture model, allows predictors to influence both the latent class classification and the within‐class outcome regression. This provides valuable insights into which variables inform latent classes (gating predictors) and which explain variance in the regressions (expert predictors). Many clustering methods are based on identifying sub‐groups via Euclidean distance. However, MoE is a likelihood‐based model that offers additional information about the probability of being allocated to each latent class. The MoE method introduced by Murphy and Murphy 12 has been applied in various domains including education 15 and biology. 16 In our study, we adopt a finite mixture model with G latent classes, where yi represents the response variable for the ith individual and xi denotes the corresponding predictors. Each latent class is modeled by a class‐specific probability density function f(yi|xi), parameterized by θg(xi) for the gth class. The latent class membership is governed by probabilities τg(xi), which satisfy the constraint ∑g=1Gτg(xi)=1. Subject‐specific predictions are computed as a weighted sum of the expert models, where the weights are given by the latent class probabilities, allowing the model to flexibly capture heterogeneity in the data. fyi|xi=∑g=1Gτgxifyi|θgxi All models were fit using R package MoEClust. 17 Prediction model development The prediction models were developed using data reported by Rissman et al., 14 which included information on 1080 screened participants. These data included blood plasma Aβ42 / Aβ40 and %p‐tau217 results, age, APOE, and amyloid PET levels. The sample was then split into two datasets using stratified random sampling on AHEAD amyloid PET status. Participants with amyloid PET levels ≥ 20 Centiloids were classified as positive, whereas those with amyloid PET levels < 20 Centiloids were classified as negative, in line with the amyloid eligibility requirements of AHEAD. By performing stratified random sampling, we ensured approximately equal amyloid positivity prevalence in both the training and test samples, which was split using a 75:25 ratio. Our primary objective in developing the algorithm was to achieve the best possible model prediction accuracy. When constructing prediction models, there is a trade‐off between bias and variance that must be considered. 18 As we increase the number of tunable parameters in a model, it becomes more adaptable and can minimize error (or bias) in its fit to the training dataset. However, this flexibility also comes with a risk of overfitting, where the model performs well on the training data but poorly on unseen data. Conversely, reducing model flexibility can lead to a reduced variance in prediction values, making the model less sensitive to the specific training data used. However, this can result in underfitting, where the model fails to capture the underlying patterns in the data and makes poor predictions. To aid in determining the amount of optimal model flexibility, we performed 10‐fold cross‐validation on each set of tunable parameters. The MoE model has three main tunable parameters: the number of classes, predictors in the gating portion, and predictors in the expert models. We performed an exhaustive sweep on the number of latent classes between two and five and all linear combinations of the pre‐specified predictors in both the gating and expert models. This resulted in 1024 model definitions to be fit on each of the 10‐fold datasets, where performance was assessed on the out‐of‐bag sample for each fold. We validated the proposed final model against two independent datasets. The first of these was the AHEAD study test data (N = 271) split prior to fitting any candidate model. The second set of data was retrieved from the Wisconsin Registry for Alzheimer's Prevention (WRAP) study 19 (N = 202), included to assess generalizability of the algorithm beyond the AHEAD screening context; WRAP was selected as it represents an independently recruited cohort with different recruitment characteristics, providing a more stringent test of external validity than the AHEAD test set alone. In Figure S2 we see comparable area under the curve (AUC) performances 0.90 on both datasets. Model inference and variable importance The following analyses were conducted post‐enrollment to characterize the the final deployed screening prediction model. To further understand the deployed model, we explored the use of SHAP (SHapley Additive exPlanations) values, a commonly employed technique in machine learning to bridge the gap between black‐box models and model inference. SHAP values are particularly useful for understanding the impact of individual predictors on the model output. Moreover, they can be utilized in aggregation, providing a reliable measure of covariate importance using the median absolute SHAP value. For this work, we utilized the R package kernelshap, 20 which implements the method proposed by Covert and Lee. 21 RESULTS AHEAD 3‐45 study blood plasma screening results The AHEAD 3‐45 study commenced without blood‐based plasma screening, resulting in over 70% of participants undergoing PET scans being deemed ineligible due to insufficient amyloid levels. As the study introduced iterations of the algorithm at the start of the screening process, we observed an increase in the proportion of individuals excluded, by the algorithm prior to PET imaging. This approach was designed to reduce the number of negative PET scans and decrease participant burden for those who were most likely to be amyloid negative. The proportion of PET‐ineligible individuals declined steadily, reaching a low of 14%, as shown in Table 1. Figure 1 illustrates the monthly amyloid PET screen failure rate during the study enrollment period, with a clear downward trend over time corresponding to the implementation of the plasma screening algorithm (green line). TABLE 1AHEAD trial eligibility rates by the screening algorithm version.Algorithm version (AV)ModelEligibility cutoffPlasma screenedPlasma eligible a PET ineligible a A3 eligibleA45 eligibleA3 proportion b No algorithmNo algorithmNANANA1237 (71.00%)153 (9.00%)346 (20.00%)0.31AV1 Aβ42/Aβ40+Age+ Num APOE ε4 alleles > 11594033.28%566 (50.00%)182 (16.00%)385 (34.00%)0.32AV2 %p‐tau217+ Aβ42/Aβ40+Age+ Num APOE ε4 alleles > 18821522.59%278 (31.00%)189 (21.00%)442 (49.00%)0.30AV3 %p‐tau217+ Aβ42/Aβ40+Age+ Num APOE ε4 alleles > 40261112.22%16 (14.00%)10 (9.00%)90 (78.00%)0.10 Note: Algorithm version is updated in stages, starting with no plasma screening. Model refers to the variables used to determine eligibility for a PET scan. Eligibility cutoff is the model prediction binary threshold for assessing eligibility for a PET scan. Plasma algorithm columns contain the number screened and proportion eligibile to continue. Amyloid eligibility is the results from the PET scan.Abbreviations: APOE, apolipoprotein E; AV, Algorithm version; PET, Positron Emission Tomography. a Some participants became ineligible after plasma screening but before undergoing amyloid Positron Emission Tomography (PET) scans, resulting in fewerPET scans than plasma‐eligible participants. b To achieve target enrollment completion for both trials (A3/A45) at the same time, the proportion of A3 enrollment needed to be 4001400=0.29. John Wiley & Sons, Ltd. AHEAD trial eligibility rates by the screening algorithm version. Aβ42/Aβ40+Age+ Num APOE ε4 alleles %p‐tau217+ Aβ42/Aβ40+Age+ Num APOE ε4 alleles %p‐tau217+ Aβ42/Aβ40+Age+ Num APOE ε4 alleles Note: Algorithm version is updated in stages, starting with no plasma screening. Model refers to the variables used to determine eligibility for a PET scan. Eligibility cutoff is the model prediction binary threshold for assessing eligibility for a PET scan. Plasma algorithm columns contain the number screened and proportion eligibile to continue. Amyloid eligibility is the results from the PET scan. Abbreviations: APOE, apolipoprotein E; AV, Algorithm version; PET, Positron Emission Tomography. Some participants became ineligible after plasma screening but before undergoing amyloid Positron Emission Tomography (PET) scans, resulting in fewerPET scans than plasma‐eligible participants. To achieve target enrollment completion for both trials (A3/A45) at the same time, the proportion of A3 enrollment needed to be 4001400=0.29. Timeline of screen fail rate due to insufficient amyloid positron emission tomography (PET) levels by month. The black line represents pre‐plasma screening algorithm, with the green line representing those who screened using plasma results. The lines run parallel due to delays in site authorization to use blood plasma screening. The numbers at each month represent the number of PET scans performed that month. The vertical dashed lines indicate the months where specific versions of the algorithm were introduced (algorithm Version 1–3). Algorithm Version 1, introduced in February 2022, used an eligibility cutoff >11 predicted Centiloids, a threshold selected to prioritize sensitivity in order to detect intermediate levels of amyloid. This resulted in a 20% reduction in screen failures due to insufficient (<20 Centiloids) amyloid levels, meaning that 50% of participants who received an amyloid PET scan were amyloid eligible for the study. Among those who were PET eligible for AHEAD, ≈32% were eligible for A3 and 68% for A45, which was close to the target ratio sample sizes for A3 and A45 (29:61). In May 2023, blood plasma biomarker %p‐tau217 was integrated into the screening algorithm Version 2. The expected improvement in model performance led to the implementation of an eligibility cutoff of >18 predicted Centiloids. This update further reduced screen fails due to insufficient amyloid levels (<20 Centiloids) from 50% with algorithm Version 1 to 31% with algorithm Version 2. The trial amyloid eligibility inclusion ratio for the A3/A45 trial continued to remain relatively stable, ≈3:7. As the study neared its target enrollment, the eligibility cutoff was raised to >40 predicted Centiloids to prioritize enrolling more participants in the A45 trial. This adjustment further reduced screen failures caused by insufficient amyloid PET levels (<20 Centiloids), with a failure rate of 14%. As intended, the substantial increase in the cutoff led to an unbalanced ratio of A3 to A45 trial eligible participants in the final randomized sample, resulting in the A45 trial being favored at a ratio of 1:9. Mixture of experts model interpretation Latent classes The final model, optimized for predictive purposes, identified three classes: low, intermediate, and high amyloid PET levels. These labels were determined after observing Figure 2 and the MoE final model classification of the data. Most participants, 59%, were classified as low amyloid, 13% were classified as intermediate amyloid, and 28% were classified as high amyloid. Moving from top to bottom in Figure 2, we observe a trend in the number of APOE ε4 alleles affecting participant classification. The top row predominantly contains low amyloid participants (78% blue points), whereas the number of high amyloid participants (red points) increases as we move from one APOE ε4 allele (46%) to two APOE ε4 alleles (78%). Within each column of Figure 2, the relationship between the continuous predictor (Aβ42 / Aβ40, age, and %p‐tau217) and amyloid PET is depicted. The %p‐tau217 exhibits a positive linear relationship, with Aβ42 / Aβ40 showing an exponential decay‐type relationship in which lower values indicate higher amyloid PET levels. Training data scatter plots of the continuous variables (amyloid beta [Aβ]42 / Aβ40, age, and %p‐tau217) versus amyloid positron emission tomography (PET) Centiloid by number of apolipoprotein E (APOE) ε4 alleles. The colors in the figure represent posterior latent class labels, indicating the most probable class for each participant based on the data. The inferred groups were assigned using the posterior latent class labels. Figure 3 shows a histogram of the amyloid PET data used for training a prediction model. The histogram exhibits a bimodal distribution, with peaks at 0 and 60 Centiloids. Notably, the transition from low to high amyloid PET levels in this dataset exhibits a heavy tail that is not fully accounted for by a two‐class model, hence the presence of an intermediate class. The model's ability to separate latent classes is further evident in the appendix Figure S3, where the gating/posterior probability histograms highlight the separation. Histogram of the amyloid positron emission tomography (PET) Centiloid with cluster density curves overlaid. The dashed line represents the study eligibility threshold at amyloid PET 20 Centiloids. Variable importance Figure 4 is a graphical representation of the final model, which is made up of three latent classes represented by boxes. To determine a participant's class assignment, the final model incorporates predictors %p‐tau217, Aβ42 / Aβ40, and APOE ε4. This section of the model, known as the gating network, determines the weight (or mixing proportion) assigned to each class in the final prediction. The relative weight of each variable is indicated by the line width. Notably, Aβ42 / Aβ40 exhibits a higher weight in classifying the low group of amyloid compared to %p‐tau217, suggesting its value for specificity (true negative rate). Conversely, %p‐tau217 demonstrates a higher weight in classifying the high group of amyloid, indicating its sensitivity (true positive rate). Node diagram representation of the mixture of experts model. Circles represent the data, including predictors and response. Squares represent the three latent classes that have been inferred to be low, intermediate, and high. The arrows represent the connection directions, with line widths reflecting the relative weight. For each class, there is a subcomponent model called an expert that predicts amyloid PET levels. Given a class membership, the expert model predicts the expected amyloid PET Centiloid. For the expert models, the final model indicated age and %p‐tau217 as predictors, with the latter having the most weight on the final model prediction. %p‐tau217 importance is further emphasized by the fact that it is the only predictor present in both the gating and expert sub models. Figure 5 presents information regarding the variable importance based on the SHAP values. The overall variable importance is summarized in the left plot, where %p‐tau217 is ranked as the most important variable, followed by Aβ42 / Aβ40, APOE ε4, and then age. The scatter plots on the right illustrate the predictor's impact on the amyloid PET value. Because the model prediction is a weighted sum of class weights and expert regression predictions, it is not immediately apparent how a predictor contributes to the amyloid PET prediction. Therefore, the scatter plots not only demonstrate this relationship, which is non‐linear in some cases, but also reflects the magnitude of effect, with stronger predictors having SHAP values further away from 0. For these data we see %p‐tau217 effect ranging from low amyloid levels to very high, with the other variables having a much more modest, but still important, impact. (Left) Bar graph sorted by variable importance. Variable importance determined by calculating the median absolute SHAP (SHapley Additive exPlanations) metric. (Right) SHAP scatter plots showing the impact of each covariate on model prediction of amyloid positron emission tomography (PET) Centiloid. Each point represents an observation, with its position on the y‐axis indicating the SHAP value (i.e., the contribution of the covariate to the prediction). Positive SHAP values indicate a contribution toward higher predicted amyloid level, whereas negative values indicate a contribution toward a lower amyloid level. The gray line is a generalized additive model trend line. Model validation We validated the final model against two independent datasets, the AHEAD study test data (N = 271) split prior to fitting any candidate model and WRAP study data (N = 202, a subset with %p‐tau217 and requisite PET and other measures). WRAP is an observational longitudinal study enrolling participants with family history of AD that aims to identify early cognitive decline and understanding contributing factors. In Figure S2 we see AUC performances >0.90 for the model prediction, with the AHEAD test data showing higher performance compared to WRAP. Classification performance at an eligibility threshold of 18 Centiloids is reported in Table S3, with both datasets demonstrating comparable specificity, accuracy, and positive and negative predictive values. Differences in AUC and sensitivity between the two datasets may in part reflect the use of different amyloid PET tracers, with AHEAD using NAV‐4694 and WRAP using Pittsburgh compound‐B (PiB), potentially affecting the reference standard against which model predictions are evaluated. DISCUSSION The reduction in screen failures due to subthreshold amyloid levels highlights the value of incorporating blood plasma screening to improve clinical trial recruitment efficiency even for the very early stage of preclinical AD. By using blood plasma biomarker testing as an initial screening step, the number of assessments during the first visit was reduced, thereby decreasing burden on both sites and participants. Notably, our blood plasma screening algorithm preserved the desired cohort enrollment for the AHEAD trials, namely A3 (intermediate amyloid levels) and A45 (elevated amyloid levels). Screen failures due to insufficient levels of amyloid significantly reduced as we iterated on the blood plasma screening algorithm versions. The initial model, which included Aβ42 / Aβ40, age, and APOE, reduced PET ineligibility rates from 70% to 50%. The introduction of %p‐tau217 further reduced ineligibility rates to 30%. As expected, the introduction of blood plasma screening resulted in high screen failure rates at the first visit, substantially reducing the overall burden of the screening process. We employed MoE, an ensemble parametric model, to capture the heterogeneity commonly observed in amyloid PET levels. Although MoE models have not yet been widely adopted in clinical prediction, the broader family of mixture models on which they are founded has been applied in diagnosis, including use of Gaussian mixture models to derive data‐driven diagnostic cutoffs for AD biomarkers in a meta‐analysis of nearly 20,000 participants. 22 MoE models extend this framework by allowing patient‐level covariates to explicitly inform both the assignment of individuals to latent subgroups and the parameters of the prediction within each subgroup, a natural extension when clinical characteristics are expected to modify the diagnostic classification. The MoE approach enabled us to identify latent classes, which we classified as low, intermediate, and high amyloid groups. The final prediction model was validated using independent test data from the AHEAD test set and the WRAP study. With the latter WRAP validation included to assess generalizability of the algorithm beyond the AHEAD screening context, providing an independent external benchmark in a cohort with different recruitment characteristics. We used SHAP values to assess the relative importance of each covariate in predicting amyloid levels and in classifying latent classes. Blood plasma biomarkers (%p‐tau217 and Aβ42 / Aβ40) along with APOE were incorporated into the gating model to guide class assignment. %p‐tau217 was the most influential in identifying the high amyloid group, Aβ42/Aβ40 was more specific to the low amyloid group, and both contributed equally to the intermediate group classification. Compared to conventional regression or classification models, the MoE framework offers greater flexibility in capturing population‐level heterogeneity, enabling more nuanced subgroup identification and improving predictive accuracy for amyloid burden. To our knowledge, six anti‐amyloid studies (ALTITUDE‐AD, 23 START, 24 SKYLINE, 25 TRAILRUNNER ALZ3, 26 and TRAILBLAZER‐ALZ2 and 3 2 , 27 ), one anti‐tau study (Autonomy trial 28 ) and one dual‐target study (E2814 29 ) has incorporated blood plasma biomarkers into their screening protocols. TRAILBLAZER‐ALZ 3 adopted a decentralized approach to clinical trial recruitment, with many screening activities conducted remotely rather than at centralized study sites. 27 It is the first clinical trial to rely on plasma p‐tau217 as the sole biomarker to establish AD pathology, thereby reducing the need for reference procedures such as amyloid PET or cerebrospinal fluid testing. Although this approach streamlines recruitment and reduces participant burden, the absence of routine confirmatory testing precludes direct assessment of plasma p‐tau217 performance within the trial itself. The authors note that the plasma p‐tau217 eligibility threshold was selected to optimize positive predictive value for amyloid PET positivity. Bauer et al. used plasma p‑tau181 to rule out amyloid‑negative participants, a strategy that required high sensitivity. 25 Using a two‐step decision algorithm led to a modest reduction in screen failures at the PET scan stage. However, blood plasma pre‐screening achieved high sensitivity, which they verified in a retrospective sample. In contrast to Bauer et al., 25 our objective was to reduce costs and the burden on sites and participants, similar to the objectives of the TRAILBLAZER‑ALZ 3 trial. 27 We adopted a MoE model incorporating Aβ42/Aβ42, %p‐tau217, age, and APOE ε4 allele count, selected based on their availability and additive predictive value within the AHEAD screening infrastructure. Although %p‐tau217 alone demonstrates strong discriminative performance, the added value of combining plasma biomarkers with age and APOE has been explored by Rissman et al., 14 and the optimal combination of predictors for clinical trial screening remains an active area of investigation. It is also worth noting that inclusion of APOE may preferentially enrich the population proceeding to amyloid PET for APOE ε4 carriers, an important consideration for future screening algorithm design. The outcomes and insights from these blood plasma pre‐screening efforts are now informing our understanding of their utility in clinical trials, and our study helps bridge this gap by offering large‐scale trial evidence on the utility of blood‐based biomarkers in screening. A limitation of the study is the potential for untestable bias in the prediction model. The model was developed using data from a relatively homogeneous population, which differed from the more diverse population achieved over the course of screening. As recruitment efforts expanded to reflect the broader demographics of the U.S. population, early analyses have shown differences in biomarker distributions across racial and ethnic groups. 30 Understanding how these differences affect model performance is critical for future validation and refinement. To address this limitation, future work will focus on retraining the prediction models using more diverse datasets and evaluating its performance across sub‐populations. Future work will also explore the use of a two cut‐point approach in diagnostic classification, which may offer greater flexibility than the single threshold currently employed. 31 , 32 This method could align more naturally with the three amyloid groups identified in this analysis (low, intermediate, and high), provide distinct boundaries that better reflect biology of the disease, and may help address classification uncertainty inherent to participants closest to a single eligibility threshold. In addition, analyses from the AHEAD study group will examine demographic characteristics of participants at each stage of the screening process, providing further insight into the broader applicability of plasma biomarker screening in clinical trial recruitment. We also plan to assess the test–retest reliability of key plasma biomarkers, including %p‐tau217 and Aβ42 / Aβ40, to ensure stability and reproducibility across time and assay conditions. CONCLUSIONS This research systematically assesses how blood plasma screening impacts recruitment efficiency, providing new insights into the practical utility of these biomarkers in a real‐world clinical trial setting. By applying advanced modeling techniques and validating against external datasets, we demonstrated the potential of blood plasma algorithms in reducing site workload and participant burden. Our work lays a foundation for more inclusive and scalable screening strategies with lower burden to the participant and the study site teams. Continued efforts to validate biomarker stability, refine diagnostic thresholds, and expand training datasets will be essential to advancing the utility of blood‐based screening in future trials. CONFLICT OF INTEREST STATEMENT Oliver Langford: Grant funding to the USC Epstein Family Alzheimer's Therapeutic Research Institute (ATRI) from Eisai, and National Institutes of Health (NIH). Rema Raman: Grant funding to ATRI from Eisai, the Alzheimer's Association, the American Heart Association, and NIH. Paul Aisen: Grants from NIH, the Alzheimer's Association, Lilly and Eisai, and consultation with Merck, Roche, Genentech, Abbvie, Biogen, ImmunoBrain Checkpoint, AltPep, Bristol Myers Squibb, Johnson & Johnson, New Amsterdam, and Neurimmune. Reisa Sperling: Grant funding to the Institution from Eisai and Eli Lilly, Alzheimer's Association, GHR Foundation, Gates Ventures, and NIH. She has served as a consultant or on scientific advisory boards for AbbVie, AC Immune, Acumen, Alector, Apellis, Biohaven, Bristol Myers Squibb, Genentech, Ionis, Janssen, Oligomerix, Prothena, Roche, and Vaxxinity over the past 3 years. Keith Johnson: Grant funding to the Institution from Eisai and Eli Lilly, Alzheimer's Association, GHR Foundation, and NIH. He has served as a consultant for Bristol‐Myers Squibb, Janssen, Merck, and Novartis. Doris Molina‐Henry: Grant funding to the Institution from Eisai and NIH. Pallavi Sachdev: Employee of Eisai. David Li: Employee of Eisai; spouse is fulltime employee of Merck & Co, Inc. Gustavo A. Jimenez‐Maggiora: Research support from NIH, the Alzheimer's Association, the American Heart Association, Gates Ventures, Eli Lilly, and Eisai. Sterling C. Johnson: Research Support from NIH; has served as an advisor to Enigma Biomedical, AlzPath, Alamar Bio, Lantheus, Eli Lilly, and Trillium Bio. Joel B Braunstein: Employee of C2N. Colin Jan Birkenbihl: Nothing to declare. Madison Cuppels: Nothing to declare. Rachel F. Buckley: Nothing to declare. Robert A. Rissman: Research support from the National Institute on Aging (NIA), Eisai, and the Alzheimer's Association; and is a consultant for Bioivt, Lexeo, Keystone Bio, Allyx, Biogen, DiamiR, Ionis, and PrecisionMed. Michael C. Donohue: Grant funding to ATRI from Eisai, and NIH; spouse is full time employee of Janssen Pharmaceuticals. Author disclosures are available in the Supporting Information. CONSENT STATEMENT Informed consent was obtained from all participants and their study partners. Requests for data related to this study should be directed to the corresponding author (O.L.). De‐identified AHEAD 3‐45 data are available from www.aheadstudydata.org. Supporting information Supporting information