Post-marketing safety signals and report-level ARIA-E stratification for lecanemab and donanemab: an integrated FAERS and WHO-VigiAccess pharmacovigilance study

Abstract

Background:

Anti-β-amyloid monoclonal antibodies provide a disease-modifying treatment approach for Alzheimer’s disease, but post-marketing safety concerns remain, particularly amyloid-related imaging abnormalities (ARIA). Real-world studies have mainly described adverse event spectra and disproportionality signals, whereas report-level amyloid-related imaging abnormalities with edema or effusion (ARIA-E) risk stratification remains less well characterized.

Objectives:

To compare post-marketing adverse event profiles of lecanemab (LEC) and donanemab (DON), identify report-level factors associated with ARIA-E, and develop a machine learning model for ARIA-E risk stratification.

Design:

Retrospective pharmacovigilance study using spontaneous reporting databases.

Methods:

Adverse event reports for lecanemab and donanemab were analyzed using the Food and Drug Administration Adverse Event Reporting System (FAERS) and WHO-VigiAccess. FAERS disproportionality analyses were performed at the Preferred Term and System Organ Class levels using frequency-based and Bayesian methods. Time-to-onset and serious-outcome patterns were further explored. Multivariable logistic regression was used to examine factors associated with report-level ARIA-E. An extreme gradient boosting (XGBoost) model was developed using temporally split FAERS datasets with nested cross-validation and temporal external validation.

Results:

FAERS included 2961 lecanemab and 1542 donanemab primary suspect reports; WHO-VigiAccess included 2525 lecanemab- and 1450 donanemab-related reports. Across databases, adverse event profiles were dominated by neurological events, with ARIA-related events representing the central safety signals. Serious outcomes clustered mainly in the early treatment period and attenuated over time. In complete-case regression, no overall difference in ARIA-E reporting odds was observed between donanemab and lecanemab, whereas a significant donanemab-by-body weight interaction was identified. The XGBoost model showed modest discrimination, with area under the receiver operating characteristic curve values of 0.677 in internal validation and 0.630 in temporal external validation. Calibration was suboptimal, and decision curve analysis suggested limited net benefit mainly within low-threshold ranges.

Conclusion:

ARIA-related and other neurological events remain the principal post-marketing safety concern for lecanemab and donanemab. The XGBoost model may support report-level risk prioritization, particularly during early treatment, but should not be used as a diagnostic substitute or a source of precise individualized risk estimates.

Plain language summary

What did this study find about safety reports for lecanemab and donanemab?

Lecanemab and donanemab are anti-amyloid antibodies used in Alzheimer disease, but both can be associated with ARIA, a brain imaging abnormality involving edema, effusion, or hemorrhagic changes. Using FAERS and WHO-VigiAccess, this study found that post-marketing reports for both drugs were dominated by neurological events, especially ARIA-related reports. Serious reports tended to occur early after treatment initiation, supporting careful monitoring during the early treatment period. A machine-learning model showed only moderate performance. It may help prioritize reports for closer review, but it should not replace clinical assessment, MRI-based monitoring, or established ARIA risk-management frameworks.

Introduction

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder and the leading cause of dementia in older adults. With global population aging, its burden has become an increasingly important public health challenge. According to the Global Burden of Disease study, an estimated 57.3 million people were living with dementia worldwide in 2019, and this number is projected to rise to approximately 153 million by 2050, particularly in aging regions. ¹ In China, about 13.14 million individuals were estimated to have AD and related dementias in 2019, accounting for roughly one quarter of the global total. ² The growing prevalence of AD substantially affects patients and caregivers and places a major burden on healthcare systems and long-term care resources.^1,3

Although the pathogenesis of AD remains incompletely understood, several interacting mechanisms are thought to drive disease onset and progression. Among them, abnormal production and deposition of amyloid-β (Aβ) is widely considered an early and central event. Aggregated Aβ disrupts synaptic function and promotes downstream pathological changes through the formation of neurotoxic oligomers and plaques.^4,5 Hyperphosphorylation of tau and subsequent neurofibrillary tangle formation are also closely linked to neuronal dysfunction and clinical decline. ⁶ In addition, neuroinflammation contributes importantly to disease progression, as activated microglia and astrocytes amplify neuronal injury through sustained inflammatory signaling. ⁷

Conventional treatments, including acetylcholinesterase inhibitors and N-methyl-D-aspartate receptor modulators, provide only modest symptomatic benefit and do not alter disease progression.^8,9 By contrast, monoclonal antibodies targeting Aβ have emerged as a disease-modifying strategy and have become a major focus of AD drug development. ¹⁰ This shift is reflected in the approvals of aducanumab (ADU) in 2021, lecanemab (LEC) in 2023, and donanemab (DON) in 2024. These agents are all immunoglobulin G1 monoclonal antibodies directed against Aβ, but they differ in target specificity. ADU and LEC primarily bind aggregated soluble and insoluble forms of Aβ, whereas DON selectively targets insoluble N-terminally truncated, pyroglutamate-modified Aβ, which may contribute to differences in efficacy and safety profiles across agents.^{11 –13}

A major safety concern with anti-Aβ therapy is amyloid-related imaging abnormalities (ARIA), including amyloid-related imaging abnormalities with edema or effusion (ARIA-E) and amyloid-related imaging abnormalities with hemorrhage or microhemorrhage (ARIA-H).^{14 –16} Although many ARIA events, particularly ARIA-E, are asymptomatic or transient, they are widely regarded as imaging manifestations of treatment-related disruption of cerebrovascular homeostasis. Previous imaging and pathological studies suggest that ARIA-H lies within the same hemorrhagic continuum as intracerebral hemorrhage, whereas ARIA-E shares a related vascular pathological basis and may serve as an earlier marker of susceptibility to more severe cerebrovascular complications.^14,16,17 Clinical evidence also indicates variation in ARIA-E frequency across agents. For example, ADU was associated with relatively high ARIA-E rates in trials, LEC showed lower rates in Clarity AD, and DON showed intermediate ARIA frequencies.^13,18,19 In addition, the apolipoprotein E ε4 (APOE ε4) allele has consistently been associated with increased susceptibility to ARIA-E.^16,18 Together, the coexistence of therapeutic benefit and safety risk, along with limited post-approval experience, makes it difficult for clinical trials alone to fully characterize the adverse event spectrum and identify high-risk feature combinations in heterogeneous post-marketing populations. ²⁰

These challenges make post-marketing pharmacovigilance particularly important for anti-Aβ therapies. The Food and Drug Administration Adverse Event Reporting System (FAERS) and the WHO-VigiAccess public portal provide large-scale spontaneous reporting data that can complement trial evidence. Previous studies using FAERS and WHO-VigiAccess have consistently identified ARIA as the most prominent and stable safety signal for this drug class, while also suggesting differences across agents in the distribution of ARIA-E, ARIA-H, and related cerebrovascular events. ²¹ However, most analyses have remained descriptive, focusing on adverse event spectra and disproportionality signals rather than report-level risk characterization. At the same time, machine learning has shown potential value in pharmacovigilance research because it can handle high-dimensional, sparse, and imbalanced real-world data. ²² Nevertheless, systematic association analyses and predictive modeling specifically focused on anti-Aβ safety pathways remain limited. Against this background, the present study integrated FAERS and WHO-VigiAccess data to characterize and compare the adverse event profiles of LEC and DON, identify factors associated with report-level ARIA-E, and develop an interpretable machine learning model for risk stratification.

Methods

Data sources and study design

This retrospective pharmacovigilance study used spontaneous reporting data from the FAERS and the WHO-VigiAccess portal. FAERS data were obtained quarterly from the official Food and Drug Administration (FDA) website. WHO-VigiAccess was used as a descriptive reference for adverse event profile comparison, with data extracted on March 29, 2026. Data cleaning, integration, and preprocessing were performed using Python version 3.14.3. (Python Software Foundation, Wilmington, DE, USA). The reporting of this study conforms to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement. ²³

The overall analytical workflow is shown in Figure S1.

Deduplication was performed according to FDA recommendations for FAERS data processing. ²⁴ For reports sharing the same CASEID, the record with the most recent FDA receipt date was retained. If CASEID and receipt date were identical, the report with the highest PRIMARYID was retained. All adverse events were standardized and coded according to the Medical Dictionary for Regulatory Activities (MedDRA), version 29.0. ²⁵ Events were first identified at the Preferred Term (PT) level and then aggregated to the System Organ Class (SOC) level for comparative analyses.

In the predictive modeling phase, temporally split datasets were constructed. The development cohort included LEC reports from 2023Q1–2025Q2, and DON reports from 2024Q3–2025Q2, whereas the temporal external validation cohort included LEC and DON reports from 2025Q3–2025Q4. Reports for predictive modeling were required to fall within the predefined development or validation windows, involve the study drugs of interest, list the study drug as the primary suspect, and remain after duplicate handling. More specific eligibility criteria for regression and machine learning analyses are described in Sections 2.4 and 2.5.

Outcome and time-to-onset definition

Adverse events were coded using MedDRA PTs and grouped at the SOC level to support aggregation and comparison. ²⁵ For time-to-onset analyses, time to onset (TTO) was calculated as the interval between the first administration date (START_DT) and the adverse event date (EVENT_DT), when both dates were available, and was expressed in weeks.

The primary outcome of this study was report-level ARIA-E, defined by the presence of prespecified ARIA-E-related MedDRA PTs. This outcome was coded as a binary variable, with reports containing at least one ARIA-E-related term coded as 1 and reports without any prespecified ARIA-E-related term coded as 0 for analytic purposes. The same binary outcome definition was used in all subsequent regression and machine learning analyses.

Disproportionality signal detection and severity-time analyses

Disproportionality analysis was performed at both the PT and SOC levels using frequency-based methods, including the reporting odds ratio (ROR) and proportional reporting ratio (PRR), together with Bayesian methods, including the information component (IC) and empirical Bayesian geometric mean (EBGM).^{26 –28} All metrics were calculated from standard 2 × 2 contingency tables, where a denoted reports involving the target drug and target adverse event, b denoted reports involving the target drug and other adverse events, c denoted reports involving other drugs and the target adverse event, and d denoted reports involving other drugs and other adverse events. Signal detection criteria were defined as follows: for ROR, the lower bound of the 95% confidence interval (CI) had to exceed 1 with a ⩾3; for PRR, the threshold was PRR ⩾2 with χ² ⩾4 and a ⩾3; for IC, IC025 had to exceed 0; and for EBGM, EBGM05 had to exceed 2.^27,28 A signal was considered robust only when all four criteria were satisfied and at least three reports were available. Full FAERS disproportionality metrics at the PT and SOC levels are provided in Tables S1 and S2, respectively. These tables include all evaluated PTs and SOCs rather than only terms satisfying the predefined robust-signal criteria. The contingency table structure and signal detection criteria are summarized in Table 2. Because WHO-VigiAccess does not provide case-level structured fields comparable to FAERS, its data were used only as a descriptive reference for reporting-pattern summaries and were not used for regression or machine learning analyses. To further characterize the temporal pattern and clinical seriousness of post-marketing adverse event reports, an exploratory severity-and-time dynamics analysis was conducted by jointly incorporating TTO and serious-outcome burden. TTO was defined as the interval from treatment initiation to adverse event onset and was expressed in weeks. These analyses were restricted to reports with valid non-missing TTO values. For Figures 1–4, the serious-outcome proportion was defined at the report level as the proportion of reports within a given PT-, SOC-, or age-specific subgroup that contained at least one serious outcome. The weekly TTO distribution for each subgroup was then combined with the corresponding report-level serious-outcome proportion to generate a composite descriptive severity-and-time dynamics profile. Accordingly, subgroups characterized by earlier onset together with a higher serious-outcome proportion contributed more prominently to the temporal severity curves (Table 1).

Table 2.

Contingency table structure and signal detection criteria for disproportionality analysis.

Methods	Statistic/criterion	Signal standard
Frequency-based	ROR	Lower 95% CI >1 and a ⩾3
Frequency-based	PRR	PRR ⩾2,χ⩾4 and a ⩾3
Bayesian	IC	IC025 >0
Bayesian	EBGM	EBGM05 >2

CI, confidence interval; EBGM, empirical Bayesian geometric mean; IC, information component; PRR, proportional reporting ratio; ROR, reporting odds ratio.

Figure 1.

PT-level serious-outcome spectrum and time-to-onset dynamics by age group for lecanemab (FAERS primary suspect reports). (a) Bubble matrix of PTs across age groups. Bubble size represents report count, and color indicates the proportion of serious outcomes (%). (b) Age-stratified time-to-onset dynamics of serious outcomes at the SOC level. Curves represent the weekly time-to-onset distribution weighted by the corresponding report-level serious-outcome proportion. The dashed vertical line indicates 90 days. Results for the ⩾90-year group should be interpreted cautiously because of the limited number of evaluable reports.

FAERS, FDA Adverse Event Reporting System; PTs, preferred terms; SOC, system organ class.

Figure 2.

PT-level serious-outcome spectrum and time-to-onset dynamics by age group for donanemab (FAERS primary suspect reports). (a) Bubble matrix of PTs across age groups. Bubble size represents report count, and color indicates the proportion of serious outcomes (%). (b) Age-stratified time-to-onset dynamics of serious outcomes at the SOC level. Curves represent the weekly time-to-onset distribution weighted by the corresponding report-level serious-outcome proportion. The dashed vertical line indicates 90 days. The ⩾90-year group is shown as insufficient data because of the very limited number of evaluable reports.

FAERS, FDA Adverse Event Reporting System; PTs, preferred terms; SOC, system organ class.

Figure 3.

Signal strength and proportion of serious outcomes across non-neurological SOC. Bubble plots show PT-level signals after exclusion of nervous system disorders. Panels are grouped by SOC and display results separately for LEC and DON. Bubble size represents the plot signal-strength value used for visualization, calculated as min [(ROR + EBGM)/2, 80], and color indicates the proportion of serious outcomes (%). The uncapped composite signal strength is provided in Table S6.

DON, donanemab; EBGM, empirical Bayesian geometric mean; LEC, lecanemab; PT, preferred term; ROR, reporting odds ratio; SOC, system organ classes.

Figure 4.

Signal strength and serious-outcome proportion across neurological PT-level signals in FAERS. Bubble plots show PT-level signals within nervous system disorders and display results separately for LEC and DON. Bubble size represents the plot signal-strength value used for visualization, calculated as min [(ROR + EBGM)/2, 80], and color indicates the proportion of serious outcomes (%). The uncapped composite signal strength is provided in Table S7.

DON, donanemab; EBGM, empirical Bayesian geometric mean; FAERS, FDA Adverse Event Reporting System; LEC, lecanemab; PT, preferred term; ROR, reporting odds ratio; SOC, system organ classes.

Table 1.

2 × 2 Contingency table for disproportionality analysis.

Item	Target AEs	Non-target AEs	Total
Target drug	a	b	a + b
Other drugs	c	d	c + d
Total	a + c	b + d	a + b + c + d

A complementary severity-focused stacked chart of PT-level adverse events was constructed to further characterize PT-specific serious-event burden and weighted clinical severity. Unlike the report-level serious-outcome proportion used in the severity-and-time dynamics analysis, this chart was based on serious-outcome entries linked to each PT, such that more than one serious-outcome category from the same report could contribute. Serious-event count was defined as the total number of serious-outcome entries linked to a given PT. Each entry was assigned a prespecified ordinal severity weight as follows: death (DE) = 5, life-threatening (LT) = 4, hospitalization or prolonged hospitalization (HO) = 3, disability (DS) = 3, congenital anomaly (CA) = 3, and required intervention to prevent permanent damage (RI) = 2. Weighted values were summed to obtain the severity-weight sum for each PT, and the severity score was calculated as the severity-weight sum divided by the serious-event count for that PT. This score, therefore, represents the average weighted severity among serious-outcome entries linked to a given PT. Thus, the severity-focused stacked chart reflects both the absolute burden of serious-outcome entries and the weighted severity intensity at the PT level, whereas the severity-and-time dynamics analysis reflects the proportion of reports containing at least one serious outcome in relation to onset timing.

Multivariable regression analysis for factors associated with ARIA-E

To examine report-level factors associated with ARIA-E reporting, multivariable logistic regression was performed using case-level FAERS reports that met the predefined eligibility criteria. Eligible reports had undergone duplicate handling, listed LEC or DON as the primary suspect drug, contained complete information on age, sex, and body weight, and had classifiable ARIA-E status according to prespecified MedDRA PT criteria. Reports containing at least one ARIA-E-related PT were coded as ARIA-E positive, whereas reports without any prespecified ARIA-E-related PT were coded as non-ARIA-E for analytic purposes. Reports with missing age, sex, or body weight, or with unclassifiable ARIA-E status, were excluded. After screening and preprocessing, 1158 reports were included in the final complete-case regression analysis. The case-level dataset is provided in Table S3.

ARIA-E reporting status was modeled as a binary outcome. LEC was used as the reference drug. Covariates included drug type, age as a continuous variable, sex, and body weight. Sex was coded as male versus female, with female as the reference category. Body weight was scaled per 10-kg increment and centered at the sample mean before construction of the interaction term to improve interpretability and model stability.

Because population pharmacokinetic analyses of DON have shown that body weight influences drug exposure, whereas exposure metrics have not been established as direct predictors of ARIA-E, body weight was treated here as an exploratory report-level factor rather than a causal exposure–risk variable. ²⁹ An interaction term between drug type and body weight was incorporated to assess whether the association between body weight and ARIA-E reporting differed between DON and LEC. To aid interpretation, marginal predicted probabilities of ARIA-E reporting were estimated from the adjusted model and plotted across the continuous body weight range using a reference profile of a female report at the median age of 76 years. Regression results were reported as odds ratios (ORs) with 95% CIs, and a two-sided p value <0.05 was considered statistically significant.

To further examine the drug-by-body weight interaction, weight-stratified analyses were performed using quartiles derived from the merged regression dataset. Body weight was categorized as ⩽54, >54–65, >65–78, and >78 kg, with rounded labels used for presentation. Within each weight stratum, ARIA-E counts and reporting proportions were summarized separately for LEC and DON, and stratum-specific logistic regression models were fitted with adjustment for age and sex. A DON-only stratified analysis using the same weight categories was also performed to assess within-drug differences in ARIA-E reporting across body weight strata. Given the spontaneous reporting nature of FAERS, all ORs were interpreted as reporting associations rather than as causal risk estimates or incidence measures.

Machine learning-based ARIA-E risk prediction and model evaluation

Machine learning-based risk stratification was conducted using a curated case-level FAERS dataset restricted to LEC and DON reports. A temporal split design was applied. The development cohort included LEC reports from 2023Q1 to 2025Q2, and DON reports from 2024Q3 to 2025Q2, whereas the temporal external validation cohort included LEC and DON reports from 2025Q3 to 2025Q4. After deduplication and case-level curation, only reports with classifiable report-level ARIA-E status were retained. Duplicate records, reports outside the prespecified time windows, reports in which neither LEC nor DON was listed as the primary suspect drug, and reports with indeterminate ARIA-E classification were excluded.

Age and body weight were not required to be complete for inclusion. Instead, missingness indicators were retained in the model, sex recorded as unknown was preserved as a separate category, and body weight values outside the prespecified plausible range of 40–110 kg were treated as missing during dataset preparation. For model fitting, missing age and body-weight values were imputed using multiple imputation by chained equations after generation of the corresponding missingness indicators. ³⁰ The imputed numeric values were retained as model inputs and should be interpreted together with the age- and body-weight-missing indicators.

The final prediction model was implemented using extreme gradient boosting (XGBoost). ³¹ Before model fitting, categorical variables were expanded by one-hot encoding, and the feature matrix of the external validation cohort was aligned to that of the development cohort to ensure structural consistency. Model development within the training data used nested stratified cross-validation to reduce overfitting. ³² The outer loop used 10-fold stratified cross-validation to estimate internal performance, and the inner loop used 5-fold stratified cross-validation for hyperparameter optimization. Within each inner-loop search, 20 random hyperparameter combinations were evaluated. Model selection prioritized area under the receiver operating characteristic curve (AUROC), followed by area under the precision–recall curve (AUPRC) and then the Brier score. In each outer fold, the model selected from the inner loop was refit on the outer training partition and evaluated on the held-out outer validation fold, and pooled out-of-fold predicted probabilities were used to summarize internal performance. To account for class imbalance, scale_pos_weight was recalculated according to the ratio of non-ARIA-E to ARIA-E reports within each outer training partition. The classifier used a binary logistic objective. After nested cross-validation, the final hyperparameter set was selected in the full development cohort, and the model was refit using n_estimators = 400, max_depth = 2, learning_rate = 0.02, min_child_weight = 7, gamma = 0, subsample = 0.6, colsample_bytree = 0.6, reg_alpha = 0.1, and reg_lambda = 10.0.

After internal model development, the final XGBoost model was retrained on the full development cohort and evaluated in the temporally external validation cohort. Model performance was summarized using AUROC, AUPRC, and the Brier score. Calibration performance was assessed using calibration curves, and potential clinical utility was explored using decision curve analysis (DCA) by comparing the model with treat-all and treat-none strategies. ³³ Performance was also explored in prespecified subgroups defined by sex, drug type, and age group. Model interpretability was assessed for the final fitted XGBoost model using SHapley Additive exPlanations (SHAP), which quantified the relative contribution of each feature to the model predictions. ³⁴

Evaluation metrics, sensitivity analysis, and software

Model performance was primarily evaluated using the AUROC and the AUPRC. Calibration was assessed using calibration curves, and potential clinical utility was explored using DCA by comparing the model with the treat-all and treat-none strategies. ³³ Internal performance was summarized from pooled out-of-fold predictions, and external performance was assessed in the temporally separate validation cohort from 2025Q3 to 2025Q4.

Sensitivity analyses focused on internal validation strategy, predictor structure, and threshold interpretation. Specifically, results obtained from the original random 10-fold cross-validation were compared with those from the outer 10-fold loop of the nested cross-validation to examine the robustness of model discrimination, calibration, and net benefit under different internal validation frameworks. A feature-correlation analysis was performed for the main predictors included in the final model to assess interdependence among candidate variables. In addition, exploratory threshold analyses were performed for the final XGBoost model. Data-driven cutoffs, including a Youden-optimal threshold and sensitivity-oriented and specificity-oriented thresholds, were derived from pooled out-of-fold predictions in the development cohort and then evaluated in the temporally external validation cohort. To further examine potential decision-analytic implications under different background event-frequency settings, scenario-based net benefit analyses were also explored at assumed prevalences of 5% and 10%. These threshold- and prevalence-based analyses were intended to support report-level risk prioritization and should not be interpreted as direct clinical decision rules or patient-level incidence estimates.

All data processing, model development, internal validation, external validation, and interpretability analyses were performed using Python version 3.14.3 (Python Software Foundation, Wilmington, DE, USA).

Results

Research sample and report characteristics

The post-marketing safety profiles of LEC and DON were characterized using spontaneous reporting data from FAERS and WHO-VigiAccess (Tables 3 and 4). In FAERS, LEC was associated with 6173 PT-level adverse event records and 2961 deduplicated primary suspect reports, of which 1130 (38.2%) contained at least one serious-outcome code and 117 (4.0%) involved death. DON was associated with 3526 PT-level adverse event records and 1542 deduplicated primary suspect reports, of which 768 (49.8%) contained at least one serious-outcome code and 33 (2.1%) involved death. In WHO-VigiAccess, 2525 LEC-related ADR reports and 1450 DON-related ADR reports were identified, providing a descriptive external reference for the FAERS findings.

Table 3.

Baseline characteristics of adverse event reports for LEC and DON in the WHO-VigiAccess database.

Characteristics	LEC	DON
Number	2525	1450
Gender
Female	1286 (50.9%)	679 (46.8%)
Male	941 (37.3%)	407 (28.1%)
Unknown	298 (11.8%)	364 (25.1%)
Age (years)
18–45 years	3 (0.1%)	0 (0.0%)
45–64 years	251 (9.9%)	61 (4.2%)
65–74 years	696 (27.6%)	250 (17.2%)
⩾75 years	869 (34.4%)	369 (25.4%)
Unknown	706 (28.0%)	770 (53.1%)
Region
Americas	2221 (88.0%)	1330 (91.7%)
Asia	214 (8.5%)	12 (0.8%)
Europe	84 (3.3%)	86 (5.9%)
Oceania	6 (0.2%)	22 (1.5%)
Africa	0 (0.0%)	0 (0.0%)
Reporting year
2026	317 (12.6%)	477 (32.9%)
2025	1102 (43.6%)	968 (66.8%)
2024	970 (38.4%)	5 (0.3%)
2023	130 (5.1%)	0 (0.0%)
2022	5 (0.2%)	0 (0.0%)
2021	1 (0.1%)	0 (0.0%)

DON, donanemab; LEC, lecanemab.

Table 4.

Baseline characteristics of adverse event reports for LEC and DON in the FAERS database.

Characteristics	LEC	DON
Number	2961	1542
Gender
Female	1581 (53.4%)	725 (47.0%)
Male	1098 (37.1%)	442 (28.7%)
Missing	282 (9.5%)	375 (24.3%)
Age (years)
90+	10 (0.3%)	3 (0.2%)
75–89	1184 (40.0%)	438 (28.4%)
60–74	930 (31.4%)	302 (19.6%)
45–59	105 (3.5%)	12 (0.8%)
18–44	5 (0.2%)	0 (0.0%)
0–17	0 (0.0%)	1 (0.1%)
Missing	727 (24.6%)	786 (51.0%)
Country/region
United States	2430 (82.1%)	1409 (91.4%)
Japan	423 (14.3%)	99 (6.4%)
China	35 (1.2%)	25 (1.6%)
Others	73 (2.5%)	9 (0.6%)
Reporting year
2023	242 (8.2%)	0 (0.0%)
2024	1244 (42.0%)	68 (4.4%)
2025	1475 (49.8%)	1474 (95.6%)

DON, donanemab; LEC, lecanemab.

Time-window and trial-referenced sensitivity analyses showed window- and source-dependent ARIA-E reporting proportions (Figure S2 and Table S15). These comparisons were descriptive and were not interpreted as true under-reporting or over-reporting estimates.

Across both databases, adverse event reports were concentrated in older adults, particularly those aged 65 years and above, and were reported predominantly from the Americas. LEC reports showed a slightly higher proportion of females, whereas DON reports included a relatively higher proportion of males and reports with unknown sex. DON also had a substantially higher proportion of missing age data in both databases.

Hierarchical structure of adverse reaction signals and temporal dynamics of onset and severe outcomes

Comparative analysis of FAERS primary suspect reports showed that adverse event signals for both LEC and DON were dominated by nervous system disorders, although the overall spectrum was broader for LEC than for DON. Among LEC reports (n = 2961), the main age groups were 75–89 years (n = 1184) and 60–74 years (n = 930), whereas DON reports (n = 1542) had a substantially higher proportion of missing age data. In both drugs, headache and ARIA-related events were reported across multiple age groups, while cerebral hemorrhage and related cerebrovascular events were more concentrated in older patients. DON retained the same ARIA-centered pattern but showed a more compact overall profile, with recurrent signals including infusion-related reaction, flushing, dyspnea, nausea, dizziness, and confusional state, as illustrated in Figures 1(a) and 2(a). Reports in the oldest age group were limited for both drugs and should therefore be interpreted cautiously. The full PT-level and SOC-level disproportionality metrics underlying the FAERS signal characterization are provided in Tables S1 and S2. At the SOC level, nervous system disorders showed the most consistent positive disproportionality pattern for both drugs, while vascular disorders also met the robust-signal criteria for DON.

Against this age-stratified background, time-to-onset analyses showed that valid time-to-onset information was available for 1719 LEC reports and 486 DON reports, mainly in the 60–74- and 75–89-year groups. In LEC, the 60–74-year group included 635 evaluable reports. Severe outcomes in this group peaked early after treatment initiation and declined thereafter, becoming markedly attenuated by around 90 days; the leading SOCs were nervous system disorders (n = 416), general disorders and administration site conditions (n = 192), and gastrointestinal disorders (n = 70). The 75–89-year group included 804 evaluable reports and showed a higher early peak and a heavier early severe-outcome burden, driven mainly by nervous system disorders (n = 464), general disorders and administration site conditions (n = 230), and injury, poisoning, and procedural complications (n = 170), although the overall pattern still attenuated over time. The ⩾90-year LEC subgroup contained only nine evaluable reports and was, therefore, interpreted cautiously. DON showed a similar but more compact temporal pattern. Among DON reports, 135 cases in the 60–74-year group and 220 in the 75–89-year group had valid time-to-onset data. In the 60–74-year group, severe outcomes clustered within the first several weeks and declined markedly by around 6–8 weeks, with the main contributors being nervous system disorders (n = 72), injury, poisoning, and procedural complications (n = 41), and general disorders and administration site conditions (n = 37). In the 75–89-year group, early clustering remained evident, again driven mainly by nervous system disorders (n = 109), injury, poisoning, and procedural complications (n = 68), and general disorders and administration site conditions (n = 60), but the curves attenuated more quickly than in LEC and were largely flattened by around 10–12 weeks. The ⩾90-year DON subgroup contained only two evaluable reports and was not suitable for stable interpretation, as shown in Figures 1(b) and 2(b). Age-stratified report counts, valid time-to-onset records, and leading PT and SOC contributors underlying Figures 1 and 2 are summarized in Tables S4 and S5, including younger and missing age groups not displayed in the main figures for completeness.

Consistent with these temporal patterns, clinically relevant burden extended beyond the nervous system. After exclusion of nervous system disorders, the remaining PT-level signals were distributed mainly across general disorders and administration site conditions, injury, poisoning and procedural complications, investigations, psychiatric disorders, vascular disorders, immune disorders, and skin disorders. LEC retained a broader non-neurological profile, whereas DON showed a more compact pattern centered on infusion-related reaction, flushing, confusional state, blood pressure-related abnormalities, and hypersensitivity or anaphylactic reactions, as shown in Figure 3. Detailed PT-level signal strength and serious-outcome proportions underlying Figure 3 are provided in Table S6.

When the analysis was restricted to neurological PTs, both drugs remained strongly clustered around ARIA-E, ARIA-H, and related imaging abnormality terms, together with cerebral hemorrhage, seizure, dizziness, brain edema, and cerebrovascular accident. However, the neurological profile of LEC was broader and more heterogeneous, extending beyond core ARIA-related terms to a wider range of cerebrovascular, consciousness-related, and seizure-related events. DON shared the same central ARIA-related pattern but showed a narrower spread of additional neurological PTs, indicating a more concentrated neurological burden, as shown in Figure 4. Detailed PT-level signal strength and serious-outcome proportions underlying Figure 4 are provided in Table S7.

Event burden and severity intensity were not always concordant. In the neurological domain, ARIA-related PTs and cerebral hemorrhage accounted for the main serious-event burden, whereas higher severity scores were concentrated in a smaller group of clinically severe PTs. In LEC, notably high-severity neurological PTs included status epilepticus (serious count = 14; severity score = 3.64), loss of consciousness (7; 3.57), and cerebral hemorrhage (59; 3.41). In DON, corresponding examples included balance disorder (5; 3.60), cerebral hemorrhage (35; 3.46), and loss of consciousness (5; 3.40), although some estimates were based on small numbers. In the non-neurological domain, after excluding death, notable high-severity PTs in LEC included myocardial infarction (13; 4.00) and pulmonary embolism (17; 3.59), whereas in DON they included atrial fibrillation (9; 3.78) and blurred vision (13; 3.54). These findings indicate that the PTs with the greatest clinical seriousness were not necessarily those with the highest reporting burden, and that prioritization should consider severity intensity in addition to signal volume and serious-event burden, as illustrated in Figure 5. Detailed PT-level serious-event counts, severity-weight sums, and severity scores underlying Figure 5 are provided in Table S8.

Figure 5.

Severity-focused stacked chart of PT-level serious-event burden and weighted severity in FAERS. Stacked horizontal bars show the distribution of serious-outcome entries linked to each PT, presented separately for nervous system PTs and non-nervous system PTs and for LEC and DON. Bar length represents the total serious-event count for each PT. Colored segments indicate the composition of serious-outcome categories, including DE, LT, HO or prolonged HO, DS, CA, and required intervention to prevent permanent damage (RI). Text labels indicate the total serious-event count and the average severity score for each PT. Severity scores were calculated from prespecified ordinal weights (DE = 5, LT = 4, HO/DS/CA = 3, RI = 2), divided by the serious-event count for each PT.

CA, congenital anomaly; DE, death; DON, donanemab; DS, disability; EBGM, empirical Bayesian geometric mean; FAERS, FDA Adverse Event Reporting System; HO, hospitalization; LEC, lecanemab; LT, life-threatening; PT, preferred term; ROR, reporting odds ratio; SOC, system organ classes.

Overall, ARIA-related events together with selected cerebrovascular, seizure-related, infusion-related, and systemic reaction signals combined strong disproportionality, early onset, substantial serious-event burden, and, in some cases, high-severity intensity, supporting their prioritization in subsequent explanatory and predictive analyses.

Results of multivariable logistic regression

Among the 1158 eligible complete-case reports, 978 involved LEC and 180 involved DON. ARIA-E was identified in 210 reports overall, including 179 LEC reports and 31 DON reports, corresponding to reporting proportions of 18.3% and 17.2%, respectively. As shown in Figure 6(a), the multivariable logistic regression model, which included age, sex, body weight, and the drug-by-body weight interaction, did not show a statistically significant overall difference in ARIA-E reporting between DON and LEC at the mean-centered body weight level (OR, 0.93; 95% CI, 0.59–1.45; p = 0.733). In the LEC reference group, body weight per 10-kg increase was not independently associated with ARIA-E reporting at the conventional significance level (OR, 1.10; 95% CI, 0.99–1.22; p = 0.081). Age and sex showed nominal associations with ARIA-E reporting, with older age associated with lower reporting odds (OR per year, 0.97; 95% CI, 0.95–0.99; p = 0.002) and male sex associated with lower reporting odds than female sex (OR, 0.69; 95% CI, 0.49–0.99; p = 0.043).

Figure 6.

Multivariable logistic regression and drug-by-body weight interaction for ARIA-E reporting. (a) Adjusted ORs with 95% CIs from the multivariable logistic regression model, including drug type, body weight (per 10 kg), age, sex, and the drug-by-body weight interaction, with LEC as the reference drug. (b) Marginal predicted probabilities of ARIA-E reporting across the continuous body weight range, estimated from the adjusted model for a female report at the median age of 76 years. Shaded areas indicate 95% CIs, and the dashed vertical line marks the mean-centered body weight reference (67.7 kg). Analyses were restricted to FAERS primary suspect reports with classifiable ARIA-E status and complete data on age, sex, and body weight.

ARIA-E, amyloid-related imaging abnormalities with edema or effusion; CIs, confidence intervals; FAERS, Food and Drug Administration Adverse Event Reporting System; LEC, lecanemab; ORs, odds ratios.

A significant interaction between body weight and drug type was observed (interaction OR per 10-kg increase, 0.64; 95% CI, 0.48–0.86; p = 0.003). Marginal predicted probabilities of ARIA-E reporting were plotted across the continuous body weight range for a female report at the median age of 76 years (Figure 6(b)). The prediction curves crossed at approximately 67.7 kg. At lower body weight levels, DON showed a higher predicted probability of ARIA-E reporting than LEC, whereas at higher body weight levels the predicted probability for DON became lower than that for LEC.

Weight-stratified analyses are shown in Figure 7(a) and (b). In the all-case stratified analysis, DON was compared with LEC within each body weight quartile, with LEC serving as the reference drug in each stratum and with adjustment for age and sex. No statistically significant difference in ARIA-E reporting between DON and LEC was observed in Q1 ⩽54 kg, Q2 >54–65 kg, or Q3 >65–78 kg. In Q4 >78 kg, DON was associated with lower odds of ARIA-E reporting than LEC (OR, 0.23; 95% CI, 0.07–0.79; p = 0.019). In the DON-only stratified analysis, comparisons were restricted to DON reports, with Q1 ⩽54 kg serving as the reference category. No statistically significant differences were observed for Q2 >54–65 kg or Q3 >65–78 kg, whereas Q4 >78 kg showed lower odds of ARIA-E reporting than Q1 ⩽54 kg (OR, 0.18; 95% CI, 0.04–0.89; p = 0.036).

Figure 7.

Weight-stratified analysis of ARIA-E reporting in the regression cohort. (a) Adjusted DON-versus-LEC comparisons of ARIA-E reporting across body-weight quartiles. ORs were estimated within each quartile using logistic regression adjusted for age and sex, with LEC as the reference drug. DON was associated with a lower adjusted odds of ARIA-E reporting only in the highest body-weight quartile (>78 kg). (b) Adjusted within-DON comparisons across body-weight quartiles. ORs were estimated among DON reports using the lowest body-weight quartile (⩽54 kg) as the reference group and adjusting for age and sex. The highest body-weight quartile showed a lower adjusted odds of ARIA-E reporting compared with the lowest quartile. (c) Descriptive ARIA-E reporting proportions across body-weight quartiles in the full regression cohort. Pairwise comparisons were performed against the lowest quartile, and no statistically significant differences were observed. (d) Descriptive ARIA-E reporting proportions across body-weight quartiles among DON reports only. Compared with the lowest body-weight quartile, the highest quartile showed a lower unadjusted ARIA-E reporting proportion. Error bars and intervals in panels (a) and (b) represent 95% CIs. p Values are two-sided.

ARIA-E, amyloid-related imaging abnormalities with edema/effusion; CIs, confidence intervals; DON, donanemab; LEC, lecanemab; ORs, odds ratios.

Descriptive reporting proportions across body weight quartiles are shown in Figure 7(c) and (d). In the full regression dataset, ARIA-E reporting proportions were 16.6% (48/289) in Q1, 19.7% (57/290) in Q2, 17.6% (51/289) in Q3, and 18.6% (54/290). Compared with Q1, the unadjusted comparisons were not statistically significant for Q2 (p = 0.342), Q3 (p = 0.740), or Q4 (p = 0.525). Within DON reports, the corresponding proportions were 25.0% (8/32), 19.6% (9/46), 22.4% (11/49), and 5.7% (3/53), respectively, with the lowest proportion observed in Q4 >78 kg. Overall, the stratified analyses indicated body weight-related heterogeneity in DON-associated ARIA-E reporting, although several strata contained limited event counts.

Results of integrated machine learning modeling

Machine learning-based risk stratification was performed using a curated case-level FAERS dataset restricted to LEC and DON reports. The final XGBoost model was developed in 2760 reports from the prespecified development window, including 399 ARIA-E-positive reports (14.5%), and evaluated in a temporally independent cohort of 1743 reports from 2025Q3 to 2025Q4, including 284 ARIA-E-positive reports (16.3%). The corresponding case-level development and temporal external validation datasets are provided in Tables S9 and S10, respectively.

In the primary internal validation based on the outer 10-fold loop of nested cross-validation, pooled performance reached an AUROC of 0.677, an AUPRC of 0.240, and a Brier score of 0.215. In the temporally external validation cohort, performance decreased to an AUROC of 0.630, an AUPRC of 0.234, and a Brier score of 0.241, indicating attenuation under temporal distribution shift but retention of modest discriminative ability. Detailed performance summaries and prediction-level source data underlying Figures 8 and 9 are provided in Table S11.

Figure 8.

Internal validation performance of the XGBoost ARIA-E risk-stratification model using nested 10-fold CV. (a) Receiver operating characteristic curves from the outer folds of nested CV, with the pooled curve shown in black. The pooled AUROC was 0.677, indicating modest discrimination for report-level ARIA-E classification. (b) Precision–recall curves from the outer folds, with the pooled curve shown in black. The pooled AUPRC was 0.240, reflecting limited positive-class enrichment under class imbalance. (c) Calibration curves comparing mean predicted probabilities with observed ARIA-E event rates across probability bins. The pooled calibration curve is shown in black. (d) DCA across threshold probabilities. The pooled model curve is compared with treat-all and treat-none strategies. Net benefit was limited and was not sustained across the full range of thresholds.

ARIA-E, amyloid-related imaging abnormalities with edema/effusion; AUPRC, area under the precision–recall curve; AUROC, area under the receiver operating characteristic curve; CV, cross-validation; DCA, decision curve analysis.

Figure 9.

Temporal external validation of the XGBoost ARIA-E risk-stratification model in 2025Q3–2025Q4 reports. (a) Receiver operating characteristic curve in the temporal external validation cohort. The AUROC was 0.630, indicating attenuated but above-chance discrimination when the model was applied to later reports. (b) Precision–recall curve in the temporal external validation cohort. The AUPRC was 0.234, suggesting limited positive-class enrichment under temporal validation. (c) Calibration curve comparing mean predicted probabilities with observed ARIA-E event rates across probability bins. The external calibration curve was below the ideal calibration line across most bins, indicating that predicted probabilities tended to overestimate observed ARIA-E reporting risk. (d) DCA in the temporal external validation cohort. The model was compared with treat-all and treat-none strategies across threshold probabilities. Net benefit was limited to a narrow lower-threshold range and was not sustained at higher thresholds.

ARIA-E, amyloid-related imaging abnormalities with edema/effusion; AUPRC, area under the precision–recall curve; AUROC, area under the receiver operating characteristic curve; DCA, decision curve analysis.

Calibration and decision-analytic results were also moderate. In both internal nested cross-validation and temporal external validation, calibration plots suggested that predicted probabilities tended to exceed observed event rates in the higher-risk range. DCA showed only a limited net benefit advantage over the treat-all and treat-none strategies, mainly within a relatively narrow low-threshold range, whereas this advantage diminished at higher thresholds. These findings suggest that the model may be more suitable for low-threshold risk prioritization or screening than for high-confidence rule-in decision-making.

As a sensitivity analysis of the internal validation strategy, the original random 10-fold cross-validation yielded slightly higher internal discrimination estimates, with an AUROC of 0.682 and an AUPRC of 0.246, while the overall performance pattern remained similar to that observed in nested cross-validation. The corresponding expanded diagnostic plots suggested only partial separation between ARIA-E-positive and ARIA-E-negative reports, suboptimal calibration in the higher predicted-probability range, and preserved but modest ranking utility, as reflected by cumulative gain and lift curves above the random baseline (Figure S3). Exploratory subgroup analyses within this internal framework also suggested some heterogeneity across sex, drug, and age strata, although these findings should be interpreted cautiously. A supplementary feature-correlation analysis showed that most pairwise correlations among the main predictors were weak to low-to-moderate. The strongest correlations were mainly observed among engineered drug-by-weight interaction terms and related drug-class indicators, as well as between partially overlapping cerebrovascular proxy variables. These findings did not suggest marked global collinearity across the final predictor set (Figure S4 and Table S12).

Exploratory threshold analyses further indicated that model performance was sensitive to cutoff selection. When development-derived thresholds were applied to the temporally external validation cohort, lower thresholds improved sensitivity at the expense of specificity, whereas higher thresholds showed the opposite pattern. In scenario-based net benefit analyses assuming prevalences of 5% and 10%, positive net benefit was observed only within very low-threshold ranges. These findings support cautious interpretation of the model as a report-level prioritization tool rather than a source of clinically actionable cutoff values (Tables S13 and S14). Taken together, the final XGBoost model showed partial temporal transportability, but its discrimination, calibration, and potential clinical utility remained moderate.

Model interpretability

To improve model interpretability, SHAP summary plots were generated separately for the test and training sets (Figure 10(a) and (b)). In the test set (Figure 10(a)), reporter_type_Physician was the most influential feature, followed by reporting_country_JP, sex_M, weight_missing, drug_weight_LEC, age_years, and drug_weight_donanemab. Physician reporting consistently shifted model output toward higher predicted ARIA-E risk, whereas reporting_country_JP generally shifted predictions toward lower risk. Male sex also tended to shift predictions toward lower risk. Body-weight-related variables ranked prominently, and lower body weight tended to be associated with higher predicted risk. The high ranking of the drug-specific weight interaction terms further indicated that the contribution of body weight to prediction differed between LEC and DON.

Figure 10.

SHAP summary plots for the final XGBoost ARIA-E risk-stratification model. (a) SHAP summary plot in the test set. (b) SHAP summary plot in the training set. Features are ranked by their mean absolute SHAP values, with each point representing one report. Positive SHAP values indicate contributions toward a higher predicted probability of ARIA-E reporting, whereas negative SHAP values indicate contributions toward a lower predicted probability. Feature values are shown by color, with higher values in red and lower values in blue. Across the test and training sets, physician reporter type consistently contributed toward higher predicted ARIA-E risk, whereas reporting from Japan and male sex generally contributed toward lower predicted risk. Lower age and lower body weight tended to contribute toward higher predicted risk, and drug-specific body-weight terms ranked among the more influential predictors. These SHAP patterns describe feature contributions to model output and should not be interpreted as causal effects.

ARIA-E, amyloid-related imaging abnormalities with edema/effusion; SHAP, Shapley additive explanations.

In the training set (Figure 10(b)), the overall pattern was broadly similar. reporter_type_Physician again ranked first, followed by weight_missing, sex_M, age_years, drug_weight_LEC, and drug_weight_DON, while reporting_country_JP remained among the leading contributors. The directional patterns of the main features were generally consistent with those observed in the test set. oral_antiplatelet and several report-derived or missingness-related variables also contributed to model output, although their effects were smaller or less stably ranked than those of the leading features.

Overall, the SHAP patterns in the test and training sets were broadly concordant and support the internal consistency of the final model. These SHAP values should be interpreted as contributions to prediction rather than as causal effects.

Discussion

This study provided an integrated evaluation of adverse events associated with anti-Aβ monoclonal antibodies using two large spontaneous reporting databases. Regression and machine learning approaches were further applied to examine factors associated with ARIA-E and to explore report-level risk stratification. Across databases, the overall adverse event profiles were broadly consistent. Neurological events predominated for both drugs, and ARIA-E together with ARIA-H remained central within the reported safety spectrum. These findings reinforce ARIA as the principal safety issue requiring close attention during the real-world use of anti-Aβ therapies. ²¹ ARIA-E proportions reported in pivotal trials are not directly comparable with FAERS reporting proportions because the two sources differ in denominator structure, event ascertainment, and reporting mechanisms. Accordingly, the observed differences were interpreted as apparent reporting-proportion discrepancies rather than as direct evidence of true under-reporting or over-reporting.

From a clinical perspective, the time-dynamic findings suggest that the early post-treatment period represents a higher-risk window for both LEC and DON. Severe adverse outcomes for both anti-Aβ antibodies were concentrated during the initial treatment phase and then attenuated over approximately 90 days, a pattern that is broadly consistent with previous trial-based and real-world observations, although earlier studies mainly focused on individual adverse events or overall reporting patterns rather than severe-outcome dynamics across drugs and age groups.^13,19,21 At the same time, the dominant patterns of serious outcomes differed by age. In patients aged 60–74 years, the serious-outcome burden extended beyond neurological events to include vascular and injury-related categories, whereas in patients aged 75 years and older, severe neurological outcomes were more prominent. These findings suggest that the early treatment phase may warrant closer monitoring across age groups, but that follow-up priorities may need to be age-sensitive rather than entirely uniform.

Severe adverse outcomes associated with LEC and DON remained centered on ARIA-related and other neurologic events. In the present analysis, ARIA-E contributed the largest serious-event burden, followed by ARIA-H and cerebral hemorrhage, while seizure, dizziness, and cerebrovascular accident also made substantial contributions. This pattern was more pronounced for LEC, which showed a broader severe neurologic spectrum, whereas DON remained more concentrated around ARIA-related and hemorrhagic events. These findings are broadly consistent with prior imaging, pathological, and clinical observations suggesting that ARIA-E, ARIA-H, and treatment-emergent hemorrhagic complications are better understood as related manifestations of treatment-associated disturbance of cerebral vascular homeostasis during amyloid clearance, rather than as entirely separate entities.^13,17,35 Severe non-neurologic outcomes were also not negligible. Infusion-related reaction accounted for the largest non-neurologic serious-event burden for both drugs, while death, confusional state, fall, nausea, vomiting, pyrexia, dyspnea, and selected cardiopulmonary events also contributed to the overall severity profile. Overall, these results indicate that the more serious post-marketing burden of anti-Aβ therapy remains driven mainly by ARIA-related and other neurologic events, but also extends to a broader set of clinically relevant systemic complications.

Body weight appeared to modify ARIA-E reporting in a treatment-specific manner. In the present FAERS analysis, no clear body weight gradient was observed for LEC, whereas lower body weight was associated with less favorable ARIA-E reporting patterns in DON, as reflected by the significant interaction term, stratified analyses, and the crossover pattern in the marginal prediction curves (Figure 6(b)). This pattern is biologically plausible because DON is administered using a fixed-dose regimen after titration, whereas LEC is administered according to body weight. Population PK/PD analyses have shown that body weight influences DON exposure through its effects on clearance and volume of distribution.^29,36 However, these findings should not be interpreted as establishing body weight as a causal predictor of DON-related ARIA-E. Rather, body weight may be worth considering when interpreting DON-related reporting patterns or designing future post-marketing studies, particularly under fixed-dose regimens. Age and sex were statistically significant in the adjusted FAERS model, but they were not the primary focus of the present analysis and were included as prespecified covariates to reduce potential confounding. More broadly, current ARIA literature places greater emphasis on anti-Aβ antibody exposure, APOE ε4 carrier status, pre-existing microhemorrhages, CAA-related imaging features, antithrombotic exposure, and treatment timing than on age or sex alone. Accordingly, age and sex are better regarded here as adjustment variables than as primary mechanistic findings, and body weight should not be viewed as a substitute for established ARIA monitoring frameworks centered on MRI surveillance and other recognized risk factors. Further studies integrating body weight, genotype, MRI findings, dosing history, and PK exposure data will be needed to determine whether the observed reporting pattern reflects a true exposure–response relationship.

The model developed in this study is better regarded as a tool for report-level risk stratification than as a diagnostic method or a source of precise individual probability estimates.^37,38 DCA suggested limited and threshold-dependent potential utility when compared with strategies of no additional monitoring or universal monitoring. ³³ Exploratory threshold analyses further showed that the practical implications of the model were highly cutoff-dependent, with positive net benefit observed only within very low threshold ranges under the assumed prevalence scenarios. Taken together, these findings suggest that the model may support low-threshold prioritization of closer surveillance for reports more likely to be associated with ARIA-E, but its role should remain supportive rather than definitive and should not be interpreted as providing clinically actionable MRI surveillance cutoffs. Its practical value will depend on further external evaluation and integration with clinical judgment and routine monitoring workflows.

Limitations

This study has several limitations. First, both FAERS and WHO-VigiAccess are spontaneous reporting systems and therefore lack reliable exposure denominators, making it impossible to estimate true incidence or directly compare reporting proportions with trial-based event rates. These databases are also subject to under-reporting, selective reporting, duplicate reporting, variable report quality, and incomplete clinical information. Second, several clinically important variables were unavailable or incompletely captured, including baseline neuroimaging findings, microbleed burden, concomitant vascular risk profiles, and genetic factors such as APOE ε4 status. As a result, residual confounding could not be fully excluded, and mechanistic interpretation remained limited. Third, ARIA-E status in the present study was defined at the report level according to prespecified MedDRA PTs. Reports without ARIA-E-related terms were classified as non-ARIA-E for analytic purposes, but this should not be interpreted as definitive clinical exclusion of ARIA-E, because some cases may have gone undetected or unreported. Fourth, some variables used in the machine learning analyses were approximated from available FAERS fields rather than directly measured. In particular, antithrombotic or anticoagulant use, antihypertensive treatment, cerebrovascular disease history, and related clinical background variables were derived indirectly from concomitant medications, indication terms, or event coding. These approximations may have introduced misclassification and reduced the precision of model-based inference. Fifth, although temporally split development and validation datasets were used in the predictive modeling phase, the model was not validated in a fully independent prospective cohort, which limits the strength of its generalizability claims. Finally, because WHO-VigiAccess does not provide case-level structured data comparable to FAERS, it could be used only as a descriptive external reference rather than as a platform for regression or predictive modeling. Despite these limitations, the present study provides an updated and structured characterization of post-marketing safety patterns associated with anti-Aβ therapies and identifies several clinically relevant signals that warrant continued monitoring in real-world practice.

Conclusion

This study updated the post-marketing safety characterization of anti-Aβ monoclonal antibodies using real-world spontaneous reporting data and developed an XGBoost-based model for ARIA-E risk stratification. The model showed modest discrimination, partial temporal transportability, and limited threshold-dependent clinical utility. It may therefore serve as a supportive tool for risk prioritization and monitoring, particularly during the early treatment period, but should not be regarded as a diagnostic substitute or a source of precise individualized risk estimates.