In this large real-world cohort of patients with AD, we observed notable differences in all-cause mortality risk across several commonly prescribed SGAs. Using time-varying Cox models, we identified significant differences in mortality risk between different SGAs treatment, with aripiprazole and quetiapine showing lower associated mortality in several head-to-head comparisons. To explore potential treatment effect heterogeneity, we further applied causal tree models to identify subgroups with differential treatment responses. Across multiple comparisons, factors such as concurrent use of antidementia or T2DM medications appeared to influence the magnitude of treatment effects. These findings provide strong evidence on the necessary personalized risk–benefit evaluation when prescribing antipsychotics in AD populations.

Our survival analysis revealed that different SGAs are associated with distinct risks of all-cause mortality in patients with AD. Many prior studies have demonstrated that antipsychotic use, particularly among older adults with dementia, is linked to increased mortality [19,20,21]. This evidence led to the FDA’s black box warning in 2005 regarding the use of these medications in dementia-related psychosis [4]. However, these existing studies have focused on comparisons between overall antipsychotics and placebo or nonuser. In contrast, our study used an active-comparator design to directly compare single SGAs within a well-matched AD population, our findings are also supported by previous studies reporting similar safety patterns among SGAs. Kales et al. reported that, among patients with dementia in the United States Department of Veterans Affairs system, quetiapine was associated with the lowest mortality risk when compared with risperidone, and there is no significant difference in mortality between olanzapine and risperidone [22]. This pattern aligns with our findings. However, this study was based on a predominantly male veteran population, which limits its generalizability. Our data source had a more balanced gender distribution, potentially improving the external validity of our results. Similarly, Rossom et al. observed differential mortality risks when individual antipsychotics were compared with no treatment, with quetiapine showing the lowest hazard ratio, followed by olanzapine and risperidone [23]. While this study also suggests a safety hierarchy among SGAs, the comparison between treatment users and nonusers raise concerns of indication bias. For our study, under a real-world clinical decision-making setting, we care more about which specific antipsychotic poses the least risk when treatment is necessary. Real world evidence for head-to-head comparisons involving aripiprazole remains limited. Two network meta-analyses included aripiprazole but differed from our study in several prospectives. Yunusa et al. found no statistically significant differences in mortality among SGAs, but their conclusions were based on indirect comparisons, and the inclusion criteria were restricted to patients with dementia-related psychosis rather than an AD population [24]. Lü et al. did not examine aripiprazole’s impact on mortality, but rather emphasized its relative advantage in terms of acceptability, which limits the comparability of its findings to our results [25].

The consistently lower hazard ratios observed for aripiprazole in our study may be partly attributable to its favorable metabolic profile. While aripiprazole was consistently associated with lower mortality compared with olanzapine and quetiapine, its comparison with risperidone did not reach statistical significance. This may reflect limited statistical power due to the smaller matched sample size for this comparison, potentially resulting in a type II error. Nonetheless, the crude mortality rate still favored aripiprazole (97.2 versus 166.4 per 1000 person-years), aligning with the observed hazard ratio. Compared with other SGAs, aripiprazole exhibits a relatively favorable metabolic profile, with lower incidences of weight gain, glucose dysregulation, and dyslipidemia [26]. These properties may collectively reduce cardiovascular and metabolic complications, which are key contributors to mortality in patients with AD [27]. This explanation is further supported by our subgroup analysis, which found that the mortality benefit of aripiprazole was more notable among patients concurrently using T2DM medications. In addition, unlike olanzapine, risperidone, and quetiapine, aripiprazole is a partial agonist at D₂ and 5-HT₁A receptors, which allows it to modulate dopaminergic activity in a more balanced manner rather than fully blocking the receptor [28]. This mechanism has been associated with a lower risk of extrapyramidal symptoms (EPS), which are known to contribute to morbidity and mortality in older adults [28, 29].

Building on these overall mortality differences across SGAs, we further explored whether certain patient subgroups might experience differential benefits from specific medications. Using causal tree analysis with TMLE, we identified treatment effect heterogeneity (HTE) on the basis of baseline characteristics. In particular, among AD patients who were using T2DM medications, which is an indicator of preexisting metabolic dysregulation, we observed that the survival benefit of aripiprazole over quetiapine (CATE ≈ 0.139) and over risperidone (CATE ≈ 0.180) was more obvious than in the overall cohort. This finding may be explained by existing evidence that SGAs differ substantially in their metabolic impact. Citrome et al. reported that aripiprazole has been shown to carry a relatively low risk of inducing weight gain, insulin resistance, and new‑onset T2DM in adult populations [30]. Although Yeh et al. observed a slightly increased risk of long-term major adverse cardiovascular events with aripiprazole compared with risperidone in patients with schizophrenia and T2DM, the risk of sudden cardiac death was numerically lower for aripiprazole in both the first year and beyond 1 year [31]. While the patient populations and outcome definitions are different from ours, this trend aligns with the mortality benefit observed in our AD cohort. Taken together, these findings suggest that the mortality risk associated with SGAs in patients with AD may not be uniform across individuals and may depend on underlying clinical profiles such as comorbidity burden. The observed heterogeneity indicates that aripiprazole may represent a comparatively safer option for patients with preexisting metabolic dysregulation, as suggested by the use of T2DM or lipid-lowering medications. Although further validation is needed, these results support the potential for tailoring antipsychotic selection to individual clinical characteristics rather than relying on a one-size-fits-all treatment approach.

Sensitivity analyses demonstrated that the main findings were generally robust to variations in grace period definitions and follow-up periods. While most hazard ratios maintained similar directions and magnitudes across different grace periods (14, 60, and 90 days) and a shorter 1-year follow-up, statistical significance varied in a few comparisons. The comparison between risperidone and quetiapine demonstrated statistical significance only under the 30-day grace period and became marginally significant under the 90-day window. This suggests that the observed treatment effect for this comparison may be sensitive to exposure risk window definition. Moreover, when a shorter 14-day grace period was applied, several comparison pairs lost statistical significance, which suggests that the mortality risks associated with these antipsychotic medications may not emerge immediately after treatment ends. Instead, their effects on mortality may develop gradually over a longer duration. These findings showed the methodological importance of defining an appropriate grace period in pharmacoepidemiologic studies, as it can substantially impact the interpretation of treatment effects.

This study has several strengths. We used real-world electronic health record data from a large and diverse AD population. Importantly, the Truveta platform includes linked pharmacy dispensing data, which enhances the accuracy of medication exposure. We applied an active comparator, new user design and time varying Cox models, which improve upon traditional designs by reducing indication bias and better reflecting clinical practice. To further control for confounding, we implemented propensity score matching (PSM) to ensure balanced baseline characteristics between treatment groups. Given the high prevalence of medication nonadherence and discontinuation in real-world settings, modeling drug exposure as a time-varying variable allowed us to more accurately capture patients’ dynamic treatment patterns. Furthermore, we applied a causal machine learning framework, causal tree analysis with TMLE, to explore treatment effect heterogeneity. This method enabled data-driven identification of clinically meaningful subgroups that may benefit differently from specific SGAs, providing insights for future work in personalized prescribing.

However, several limitations should be noted. As with all EHR‑based studies, diagnostic accuracy may vary across providers and care settings. Although structured coding systems were used to identify AD and other comorbidity diagnoses, misclassification remains possible and could introduce bias. Our data did not include detailed clinical assessments of neuropsychiatric symptoms or AD severity, which could influence both treatment decisions and outcomes. Prescription dosage information was not incorporated into the analysis, as our primary objective was to compare overall medication effects rather than dose–response relationships. Further stratification by dosage could substantially reduce statistical power and limit interpretability. Despite careful study design and confounding adjustment, the possibility of residual confounding cannot be eliminated. In particular, prescribing patterns of SGAs in patients with AD are shaped by complex clinical considerations—such as behavioral symptom profiles, provider preferences, and prior treatment experiences—that may not be fully captured in structured data. In our cohort, quetiapine was prescribed substantially more often than olanzapine or risperidone despite similar market availability, suggesting nonrandom treatment allocation. This underlying treatment-selection bias may have influenced the observed associations and represents an important limitation to the causal interpretation of our findings. While drug exposure was defined using dispense records, actual medication adherence could not be confirmed. In addition, we focused on all-cause mortality rather than cause-specific death. Future research using more detailed mortality data may help explore the specific pathways through which SGAs influence mortality in this population. Survival times observed in our Kaplan–Meier curves may appear higher than those reported in earlier prospective studies [32]. This may reflect differences in follow-up structure, as Truveta may not capture patients who later transition to nonparticipating systems, leading to right censoring. To minimize the bias and ensure robust estimates, we limited follow-up to 2 years, during which our observed survival patterns were consistent with recent population-level estimates [33]. Moreover, we conducted multiple pairwise comparisons between four SGAs, which may raise concerns about inflated type I error. While our analyses focused on head-to-head comparisons with separately matched cohorts and without transitive inference, this limitation should still be considered when interpreting marginally significant results. Although the overall sample was large, some subgroups identified through causal tree analysis were relatively small, and future studies with more eligible participants are needed to validate these findings. Lastly, although we applied modern causal inference methods to explore treatment-effect heterogeneity, the retrospective and observational nature of our study limits the ability to draw definitive causal conclusions. The observed associations should therefore be interpreted as hypothesis-generating rather than confirmatory.