What this is

A pre-specified secondary analysis of the ANNEXA-I RCT (n=530) showing that andexanet’s reduction in anti-FXa activity and rise in endogenous thrombin potential at 1 hour independently predict both fewer hematoma expansions AND more arterial thromboembolic events — with both biomarker signals washed out by 12 hours. The therapeutic window for benefit-and-harm is the first hour.

1. Shadow Audit

a) The mediation is incomplete. The authors explicitly concede: “mediation analyses did not establish that the effects of andexanet on hematoma expansion were explained by effects on anti-FXa activity or ETP. These analyses were limited by lack of statistical power as reflected by the wide CIs.” That’s the load-bearing sentence — and it’s missing from the abstract’s confident framing. The headline (“biomarkers independently associated with both outcomes”) is a far cry from “the biomarkers mediate the trial effect.” This is an association paper wearing a mechanism paper’s clothes.

b) The “usual care” arm is heterogenous and 100% non-U.S. standard. “In the usual care group, 33 patients did not receive a prothrombin complex concentrate” — meaning the control arm is a mixed bag of PCC, 4F-PCC, andexanet-not-given, and probably some andexanet-ineligibles. The thrombotic excess attributed to andexanet may partly reflect that the “control” included both prothrombin concentrate AND nothing-at-all (where “nothing” has its own small pro-thrombotic fingerprint).

c) No functional outcome data are reported. The trial (and this substudy) never reports mRS at 30/90 days. Hematoma expansion is a surrogate. A 12.5 mL or 35% volume difference in the ICH that doesn’t translate to functional independence is a biomarker, not a benefit. The biomarker is the entire endpoint here.

d) Population excluded. Patients with GCS <7, planned surgical evacuation within 12h, thrombotic event within 2 weeks, or hematoma volume >60 mL were excluded. Translation: the sickest ICH patients — the ones who actually drive mortality — aren’t represented. The mean age is 79.1; 47.8% female. This is a frailer-than-average AF population, and they got FXa inhibitors, so they’re a selected survivor cohort even before trial entry.

e) Edoxaban is absent. The parent trial enrolled apixaban/rivaroxaban/edoxaban users, but the biomarker substudy restricts to apixaban + rivaroxaban only. Edoxaban’s dose-response to andexanet is materially different (high-dose regimen required), and the cleanest mechanistic story would have come from including it.

2. Inversion Engine

For the conclusion “andexanet’s harms outweigh its benefits” to be right, three things would need to flip:

a) The hematoma-expansion benefit would have to fail to translate to functional outcome. If you assume hematoma expansion is a valid surrogate (the authors do), the benefit is real. If you assume hematoma expansion in a mean-79-year-old AF cohort on FXa inhibitors is mostly a cosmetic CT finding that doesn’t change mRS at discharge — then andexanet’s entire rationale collapses. We don’t know, because mRS isn’t reported. The trial stopped at the surrogate.

b) The 1-hour ETP spike would have to be the cause of thrombosis rather than a marker of restoration-of-flow. The authors themselves note: “withdrawal of effective protection following reversal of anticoagulation” is a “mechanism that is supported by findings from the ANNEXA-4 study.” Translation: the ETP rise might not be prothrombotic per se; it might just be unmasking the underlying atrial disease that was being suppressed by the now-reversed anticoagulant. If that’s true, no amount of “more careful reversal” will reduce harm; you’ll just push hematoma expansion back.

c) PCC would have to actually work in FXa-associated ICH. The 2020 ACC pathway and most guidelines recommend PCC when andexanet is unavailable. The biomarker substudy’s quiet killing blow: “Our findings of only a minimal increase in ETP at 1 hour in the usual care arm and lack of association between the change between baseline and 12 hours in ETP and risk of hematoma expansion do not support their use for this indication.” If the field accepts that line, PCC — used in 89% of the usual-care arm — has just been demoted from “best alternative” to “plausible placebo.” A whole therapeutic category collapses.

3. Second-Order Catalyst

Within 30 days: Pharmacy committees at academic stroke centers add anti-FXa and ETP level-of-care-orders for any patient receiving andexanet. The ACC Expert Consensus pathway (Reference 5) gets a footnote — andexanet-induced ETP surge is now a labeled adverse event with monitoring recommendations.

Within 90 days: Neurocritical care fellowship curricula reframe andexanet as “1-hour-window drug with measurable biomarker signature” rather than “reversal agent.” Order sets move anti-FXa and ETP draw-time from “before-and-after” to “Q1h for 4h” — a small but real practice change.

Within 6 months: AHA/ASA’s ICH guideline committee (next cycle) and Neurocritical Care Society position both face the surrogate-versus-functional question. ANNEXA-I’s biggest gap (no mRS) gets called out. Industry-funded follow-ups (ANNEXA-I Extension, ANNEXA-I-Plus) start registering.

Within 12 months: The field splits. One camp argues the 1h ETP overshoot is the price of admission and develops “andexanet + brief anticoagulation bridge” protocols (start a DOAC at hour 12 — the exact time the biomarker normalizes). The other camp demands a head-to-head RCT against PCC powered for functional outcome, not hematoma expansion.

Last to move: Community hospitals and rural telestroke networks, who never had andexanet anyway and never had rapid ETP assays. They stay on PCC and get more evidence that their practice is mechanistically empty but pragmatically defensible.

4. Asymmetric Leverage

The single biggest claim with the most leverage: “ETP at 1 hour independently predicts hematoma expansion AND thromboembolic events in opposite directions.” If true, ETP becomes the single biomarker that maps the entire risk-benefit calculus of andexanet. You could, in principle, titrate the dose — push ETP up just enough to stop the bleed without crossing the thrombosis threshold. That’s a $500M drug-design target and a personalized-medicine play.

The single smallest claim that costs the most if wrong: “Andexanet’s benefits are explained at least in part by rapid reversal of anti-FXa activity and increase in ETP.” If the mediation analysis (which the authors admit is underpowered) turns out to be negative on replication, the entire mechanistic narrative — and with it the FDA label rationale and the 2024 NEJM publication’s mechanistic framing — collapses. You’re left with “the drug works, we don’t know why, and it causes clots.” That’s a regulatory migraine.

The asymmetric-payoff test the field should run: A randomized comparison of high-dose vs. low-dose andexanet in FXa-ICH, with ETP at 1h as the dose-titration target. If lower-dose achieves the same hematoma-expansion benefit with a flatter ETP curve and fewer thromboembolic events, the high-dose regimen is dead. AstraZeneca-funded or not, this is the natural next study and it’s almost certainly being designed right now.

5. Paradigm Destroyer

a) The mechanistic credibility of PCC for FXa-inhibitor reversal. The Discussion’s quiet poison pill — “Our findings of only a minimal increase in ETP at 1 hour in the usual care arm… do not support their use for this indication” — is the death of PCC’s claim to efficacy in this population. PCC never had RCT-level evidence for hematoma expansion as an endpoint; it had surrogate biomarker data and physiological plausibility. Both just got weaker.

b) The “andexanet is just a reversal agent” framing. The biomarker data confirm the inconvenient truth from the parent trial: andexanet is a reversal agent that overshoots. That overshoot IS the thrombotic risk. It is not a side effect to be mitigated; it is the pharmacology.

c) The implicit assumption that hematoma expansion is a clean surrogate for functional outcome. The authors never re-test that assumption because they don’t have mRS. The field has been living on that assumption for a decade. ANNEXA-I was supposed to put it to bed; instead it exposed the gap.

What it does NOT kill: the role of PCC when andexanet is unavailable; the role of 4-factor PCC in warfarin reversal (where INR-driven dosing is the target, not ETP); the broader case for andexanet over PCC in FXa-ICH (which still favors andexanet on the primary endpoint of hematoma expansion — the harms question is different).

MVP — Minimum Viable Proof

The cheapest single trial that would change this paper’s evidentiary weight:

A multicenter, biomarker-driven, dose-finding RCT of andexanet in FXa-ICH with ETP at 1 hour as the dose-titration target. Arms: (1) standard high-dose andexanet, (2) ETP-titrated andexanet (target ETP rise 400–600 nmol/L·min, below the thrombosis inflection), (3) usual care / 4F-PCC. Primary endpoint: hematoma expansion at 12h. Secondary: 30-day composite of ischemic stroke / MI / systemic embolism AND mRS at 90 days.

n ≈ 600 (200 per arm) gives 80% power to detect a 30% relative reduction in thrombotic events between standard-dose and ETP-titrated arms. Cost: ~$25M. Timeline: 3 years. Would resolve: does the ETP overshoot drive the harm? Can the benefit be preserved at lower doses? Is the entire field of andexanet dose-selection currently set too high?

A single biomarker dataset from this trial would either validate or refute the entire mechanistic claim in this substudy — and either keep andexanet at the top of the formulary or push it into a niche.

Best Combination

The single move with the highest asymmetric payoff based on the 5 frameworks:

Add an ETP-guided re-anticoagulation protocol to every center using andexanet for FXa-ICH. Begin a prophylactic-dose DOAC at hour 12 — the exact moment the biomarker substudy shows ETP no longer separates andexanet from usual care. This is a free intervention (DOACs are generic), it directly addresses the withdrawal-of-protection mechanism the authors themselves invoke (“withdrawal of effective protection following reversal of anticoagulation”), it requires no new regulatory approval, and it leverages the one robust temporal finding in this paper (the 12-hour washout) rather than the underpowered mediation claim.

Cost to test: a single-center quality-improvement cohort, n=50. Asymmetric upside: if the protocol reduces 30-day thromboembolic events by even 30%, it changes AHA/ASA guidance within a year.

Overvalue Warning

a) The mediation claim is genuinely underpowered. The authors admit this in the Discussion. Do not read “biomarkers independently associated with outcomes” as “biomarkers explain the treatment effect.” Those are different statistical claims and the paper only supports the first.

b) Surrogate endpoints. Hematoma expansion at 12h is a CT-volume surrogate. The biomarker-outcome associations are surrogate-on-surrogate. Two layers of surrogate inference compound; the field has repeatedly been burned by this (think: MISTIE, CLEAR-III).

c) The thrombotic-event count is small. “Among the 426 patients with both baseline and 1-hour anti-FXa activity results, 27 patients experienced an arterial thromboembolic event… 20 patients… 22 patients.” A 27-event endpoint in a 426-patient cohort with multiple covariates is a fragile finding. The OR 1.08 (1.00–1.16) has a lower CI that brushes 1.00 — p=0.047 is one decimal place from “not significant.”

d) Industry involvement. Four authors are AstraZeneca employees (Bamberg, Falkenberg, Ladenvall, Narayan) or hold AstraZeneca stock (Penland). The ANNEXA-I trial itself was AstraZeneca-sponsored. The discussion of ETP-as-thrombosis-mechanism benefits the drug narrative (it positions the harm as “expected pharmacology”) and the framing of PCC-as-ineffective removes the cheap alternative. Read the Discussion with that filter on.

e) Single time point dominance. The 1-hour window is the entire story. In a real-world stroke unit, “1 hour post-andexanet” is when the patient is in the scanner, not when the lab is drawing blood. The biomarker story may be biologically true but operationally inconvenient.

Bias to watch for: novelty bias on a drug whose mechanism of harm was already known from ANNEXA-4 (Reference 6/7). This isn’t a new finding; it’s a more granular measurement of a finding the field already had. The asymmetry here is mostly temporal precision, not biological surprise.