Anticoagulation Treatment Response Among Phenotypic ESUS Clusters: Evidence From CASPR

Reconstructed on October 14, 2026 for the missed October 7, 2026 publication slot. This is a replacement article, not a recovered original. Source selection: the paper appeared in the archived October 7 briefing; its selection is reconstructed from that briefing, not recovered from an original article draft.

Clinical Question (PICO)

Among adults with acute stroke classified as embolic stroke of undetermined source (ESUS) in a retrospective multicenter registry, was observed anticoagulation (AC), with or without antiplatelet therapy, associated with different time to recurrent stroke, major bleeding or death than antiplatelet therapy (AP) alone—and did phenotype clusters identify treatment-effect heterogeneity?[1] Treatment was chosen by clinicians, not randomized.[1] The paper evaluates both prognostic cluster differences and modeled conditional treatment effects; these are distinct questions.[1]

Bottom Line

Among 1,869 patients, propensity-adjusted causal survival forest analysis estimated an AC-versus-AP restricted mean survival time (RMST) difference of −25.8 days over 800 days (95% CI −67.9 to 16.4; P=0.23), not a statistically significant population-level difference.[1] Three clusters differed in event-free survival, but cluster assignment did not significantly predict treatment response.[1] Higher LVEF and cancer history were associated with less modeled benefit from AC in best linear projection (BLP) analyses; these observational, exploratory signals do not establish who should receive either treatment.[1]

Design

Study type: Retrospective, multicenter observational registry analysis; no random assignment.[1]
N: 1,869 complete-case patients with ESUS from 2,685 CASPR participants screened for this analysis.[1]
- Antiplatelet without anticoagulation: 1,694 (90.6%).[1]
- Anticoagulation with or without antiplatelet: 175 (9.4%).[1]
Setting: Electronic records from 27 US academic hospitals; screening period 2015–2024.[1]
Follow-up: Median 487 days (IQR 129–1,054); inclusion required at least 30 days of follow-up unless a primary outcome occurred.[1]
Exposure assignment: Treatment recorded on day 7 after stroke or at hospital discharge, whichever came first; the treating clinician selected therapy.[1] Indication, rationale and duration were not captured, and AC could have been given for a non-stroke indication.[1]
Analysis: Partitioning Around Medoids (PAM) phenotyping and propensity-score-adjusted causal survival forest, estimating RMST differences over an 800-day horizon.[1] An “honest” split used 50% of data for tree splits and 50% for terminal-node effect estimates; no statistical imputation was performed.[1]
Primary outcome: Time to first recurrent ischaemic or haemorrhagic stroke, major bleeding, or all-cause death.[1] The composite was assessed over observed follow-up; the modelled treatment contrast used the 800-day RMST horizon.[1]

Population

Inclusion Criteria

Adults aged 18 years or older with acute stroke retrospectively screened from CASPR records at the 27 hospitals and classified as ESUS under the study’s inclusive definition.[1]
The analysis included selected patients with severe left atrial enlargement, patent foramen ovale, heart failure with reduced ejection fraction (EF ≥20%), cancer, later-detected paroxysmal AF, or LV dysfunction; these features were not all exclusions under the inclusive ESUS definition.[1]

Exclusion Criteria

Patients not meeting the analytic eligibility definition were not included; a complete itemized exclusion list is not given in this analysis.[1] Follow-up under 30 days excluded a patient unless a primary outcome had already occurred.[1] The source refers readers to the parent CASPR methods for detailed definitions.[1]

Baseline Characteristics

Mean age 64 years (SD 15); 912/1,869 (49%) female.[1] Baseline NIHSS median 4 (IQR 2–10); pre-admission mRS was 0 in 1,336 (71%).[1]
Hypertension was present in 1,310 (70%), hyperlipidaemia in 877 (47%), prior stroke in 311 (17%) and cancer in 210 (11%).[1] Overall mean LVEF was 59.6% (SD 10.2); severe left atrial enlargement was reported in 163 (8.7%).[1]

Interventions

Observed AC strategy: Anticoagulation, with or without concomitant antiplatelet therapy (n=175); drug, dose and treatment duration are not specified in this analysis.[1]
Observed AP strategy: One or more antiplatelet agents without anticoagulation (n=1,694).[1] The study classified treatment at day 7 or discharge, whichever occurred first; treatment switching during follow-up was not captured.[1] These are clinician-selected exposure groups, not trial arms.[1]

Outcomes

The source reports the following cohort event counts and model estimates; the table retains the denominators, confidence intervals, P values and 800-day RMST horizon as stated.[1]

Outcome or model estimate	Result	95% CI	P value / note
Composite: recurrent stroke, major bleeding or death	417/1,869 (22.3%)	Not reported	During observed follow-up
Recurrent ischaemic stroke	207/1,869 (11.1%)	Not reported	Component count reported separately; overlap/first-event counting not clarified
All-cause mortality	206/1,869 (11.0%)	Not reported	Component count reported separately; overlap/first-event counting not clarified
Major bleeding	54/1,869 (2.9%)	Not reported	No treatment-arm rate supplied here; component counting not clarified
Overall AC-versus-AP RMST difference	−25.8 days	−67.9 to 16.4	P=0.23; 800-day horizon
Age as BLP predictor of modeled treatment effect	β=−45.60	−258.91 to 167.70	P=0.675; not significant
Baseline NIHSS as BLP predictor	β=−54.00	−323.31 to 215.32	P=0.694; not significant
Patent foramen ovale as BLP predictor	β=11.10	−87.42 to 109.63	P=0.825; not significant
LVEF as BLP predictor of modeled treatment effect	β=−327.34	−580.23 to −74.45	P=0.011; continuous LVEF association
LV dysfunction as BLP predictor	β=99.42	−25.21 to 224.05	P=0.118; not significant
LV wall-motion abnormality as BLP predictor	β=66.92	−64.06 to 197.89	P=0.317; not significant
Cancer history as BLP predictor of modeled treatment effect	β=−226.00	−440.78 to −11.21	P=0.039
Hypertension as BLP predictor	β=6.59	−79.08 to 92.27	P=0.880; not significant
Coronary artery disease as BLP predictor	β=−16.17	−177.00 to 144.65	P=0.844; not significant
Cluster 2 vs 1; cluster 3 vs 1 BLP	β=38.65; β=40.75	−62.65 to 139.96; −60.07 to 141.56	P=0.455; P=0.428; neither significant

Model diagnostics reported that 100% of terminal nodes exceeded the minimum size threshold (median leaf size 234; minimum 49); estimated treatment propensities ranged from 0.06 to 0.19, and trees averaged 6–8 leaves. These checks describe model structure and observed overlap, not proof that the unmeasured-confounding assumption holds.[1]

SHAP ranked LVEF as the strongest signal, cancer history as a strong signal and age as a moderate signal.[1] BLP associations were not statistically significant for age, NIHSS, patent foramen ovale, LV dysfunction, LV wall-motion abnormality, hypertension or coronary artery disease.[1]

The overall estimate is not evidence of equivalence: its interval includes both negative and positive RMST differences.[1] Table 3 states that positive BLP coefficients indicate greater modeled benefit from AC; the negative LVEF and cancer coefficients therefore indicate a relative shift away from AC benefit in this model, not a patient-level treatment effect or validated threshold.[1] The authors report post hoc LVEF strata (<35% and 35%–49%) but do not establish an optimal cut point; those stratum estimates are not reproduced in the article text available here.[1]

PAM generated three clusters across 21 baseline features.[1] Cluster 1 was the younger/lower-comorbidity group; cluster 2 was older and predominantly female with intermediate comorbidity; cluster 3 was predominantly male with the greatest vascular-risk burden.[1] Table 2 reports median ages of 55 (IQR 43–65), 73 (65–80) and 61 (54–71), respectively.[1] Event-free survival differed across clusters (log-rank P<0.0001), with cluster 2 having the worst prognosis; this is prognostic separation, not evidence of treatment-response separation.[1]

The survival-forest estimate found no significant overall AC advantage, and cluster assignment did not predict differential treatment effects.[1] SHAP identified candidate modifiers, then BLP testing supported LVEF and cancer history; the LVEF strata were post hoc.[1] The report does not provide comparative event or major-bleeding rates by treatment arm in the article text, so absence of a reported between-arm safety difference must not be mistaken for demonstrated safety.[1]

Figures

Flow diagram of CASPR screening and analytic inclusion. — *Figure 1.* Cohort inclusion flow diagram, with study abbreviations explained in the source caption.[1]

Three PAM patient clusters visualized by consensus t-SNE. — *Figure 2.* Consensus t-SNE display of the three PAM clusters, selected from 100 runs; the clustering used 21 baseline variables and bootstrap stability assessment.[1]

Kaplan–Meier event-free survival curves by cluster. — *Figure 3.* Kaplan–Meier survival curves comparing the composite-event experience across the three patient clusters.[1]

Distribution of model-estimated conditional treatment effects by cluster. — *Figure 4.* Distribution of model-estimated conditional average treatment effects across the phenotypic clusters.[1]

SHAP feature importance for continuous and categorical treatment-effect modifiers. — *Figure 5.* Negative SHAP values align with a favorable AP response (lower RMST to event under AC); positive values align with a favorable AC response (greater RMST to event under AC).[1]

Criticisms

Internal: Treatment was not randomized, and the AC group was small (175/1,869; 9.4%).[1] Residual confounding by indication remains plausible, including AC for venous thromboembolism; the authors explicitly state that causal implications are limited.[1]
Internal: Treatment rationale, duration and switching were not captured.[1] The recorded day-7/discharge exposure may not represent subsequent therapy; outcome ascertainment could also miss events outside participating systems.[1]
Methodological: The causal interpretation depends on unconfoundedness, overlap and independent censoring.[1] The authors note that omnibus calibration tests were unavailable for this survival-forest implementation; the BLP signals are exploratory and no optimal LVEF threshold was established.[1]
External validity: Complete-case analysis and 27 academic hospitals may not represent other populations.[1] No independent external validation of the treatment-effect model is reported.[1]
External commentary: No separate editorial or letter is assessed in this source-grounded draft; no outside critique is presented as retrieved evidence.

Funding

The study received investigator-initiated awards from Medtronic and Philips; the paper does not state the funders’ role.[1] James E. Siegler reported consulting for AstraZeneca and Bayer; the remaining authors reported no competing interests.[1]

The paper

Authors. Elena Badillo Goicoechea et al.[1]
Title. Anticoagulation Treatment Response Among Phenotypic ESUS Clusters: Evidence From CASPR.[1]
Journal. Annals of Clinical and Translational Neurology. 2026.[1]
DOI. 10.1002/acn3.70542
PMCID / PMID. PMC13642850 / 42839308.[1]
Registration. CASPR parent study: NCT06398366.[1]
Source license. Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 (CC BY-NC-ND 4.0); the embedded figures are unmodified originals.[1] See the license terms.[1]
Primary source mapping: [1] https://www.ebi.ac.uk/europepmc/webservices/rest/PMC13642850/fullTextXML — Anticoagulation Treatment Response Among Phenotypic ESUS Clusters: Evidence From CASPR.

Deep Dive — click to expand

What this is

This study asks two questions that are easy to conflate: which clinical phenotypes carry different baseline prognosis, and whether those phenotypes respond differently to AC versus AP.[1] Clustering separated event risk, but did not significantly separate modeled treatment response.[1] Two individual variables—LVEF and cancer history—were associated with conditional-effect estimates, but treatment was clinician-selected in an observational registry.[1] The result is a methodological and hypothesis-generating analysis, not a treatment-selection trial.[1]

1. Shadow Audit

The vivid cluster picture can distract from the treatment contrast: 1,694 patients received AP-only and 175 received AC with or without AP.[1] The cohort’s clusters differed in event-free survival (log-rank P<0.0001), yet BLP cluster comparisons were nonsignificant (cluster 2 vs 1 P=0.455; cluster 3 vs 1 P=0.428).[1] A phenotype that forecasts worse outcomes does not necessarily identify a larger relative benefit from one treatment.[1]

The LVEF and cancer findings deserve a separate caution.[1] SHAP nominated candidate variables and BLP then tested them; the paper does not describe multiplicity correction for this sequence, and the authors acknowledge limited AC exposure and imprecise estimates in lower-EF strata.[1] Cancer type, stage and treatment were incompletely characterized, while unrecorded thromboembolic indications could influence AC selection.[1] Statistical significance therefore does not convert these coefficients into causal subgroup effects.[1]

2. Inversion Engine

The overall modeled AC-minus-AP RMST difference was −25.8 days, with a 95% CI from −67.9 to 16.4 and P=0.23 over 800 days.[1] The interval crosses zero, so the study neither establishes AC benefit nor proves equivalence.[1] A different result could emerge in a genuinely responsive subgroup, but confidence in that claim would require replication under better treatment balance and stronger control of indication, treatment changes and outcome capture.[1] The present analysis does not identify a validated LVEF cutoff.[1]

3. Second-Order Catalyst

The near-term consequence should be better-designed evidence, not an automatic antithrombotic switch.[1] A prospective study could prespecify LVEF and cancer hypotheses, define treatment indications and exposure duration, capture switching and competing risks, and use an independently validated outcome model.[1] Randomized confirmation or a carefully specified target-trial emulation could test whether the BLP signals persist; this paper alone does not establish the needed treatment rule.[1]

4. Asymmetric Leverage

The registry offers a large overall cohort (N=1,869) but only 175 AC-exposed patients, and the article reports 417 composite events without treatment-arm component rates in the text.[1] That imbalance limits how confidently readers can interpret candidate treatment heterogeneity, especially for uncommon characteristics or lower-EF strata.[1] The larger leverage here is conceptual: separating prognosis from prediction of treatment response prevents a high-risk cluster from being mistaken for an anticoagulation-responsive cluster.[1]

5. Paradigm Destroyer

The paper’s strongest corrective is that unsupervised phenotyping is not itself precision treatment.[1] Its three clusters were prognostically distinct, but cluster assignment did not significantly predict AC-versus-AP effect estimates.[1] That challenges the reflex “find a cluster, then personalize therapy.” It does not kill population-level evidence, establish a different standard of care, or justify individual treatment decisions from these registry coefficients.[1]

MVP — Minimum Viable Proof

In this observational cohort, the estimated overall AC-versus-AP difference was −25.8 RMST days (95% CI −67.9 to 16.4; P=0.23), and phenotype clusters did not show significant treatment-effect separation.[1] LVEF and cancer history are candidate modifiers for validation, not a proven basis for prescribing.[1]

Best Combination

The authors describe the neutral overall estimate as consistent with prior randomized ESUS trials and discuss previously reported subgroup signals.[1] Those comparisons are the authors’ discussion only in this reconstruction; the underlying trials were not independently retrieved, so no outside trial estimates are used here.[1] The practical evidence package still needs externally validated effect-modifier hypotheses and a design that captures treatment indication, changes in therapy, competing events and treatment-specific safety outcomes.[1]

Overvalue Warning

Do not read the BLP coefficients as treatment effects for an individual patient, or infer a clinical LVEF cutoff: the analysis explicitly did not establish one.[1] Do not infer that AC is safe or harmful from the absence of treatment-arm bleeding rates in the report, or treat nonsignificant results as equivalence.[1] Observational confounding, a 9.4% AC exposure share, complete-case analysis and unvalidated model assumptions remain central.[1]

Educational use notice. The Time Is Brain Review is dedicated to medical education. Figures from published articles are embedded for teaching purposes, with attribution to the original publication. We do not claim copyright over any embedded figure; all rights remain with the original publisher and authors under the article's published license. If you are a rights holder and would like a figure removed, please contact the editorial team. See our editorial notice for the full policy.