Using Blanking-Period Recurrences to Test Delayed-Acting Rhythm Drugs After Ablation

Catheter ablation for atrial fibrillation, which burns scar lines to electrically isolate the veins that trigger the arrhythmia, is often followed by drugs meant to slowly undo the stretching and scarring of the heart chamber. Sacubitril/valsartan is one example. How these drugs are tested creates a trap. The first three months after ablation are a "blanking period": healing tissue produces short-lived arrhythmias that do not mean the procedure failed. The drugs themselves are expected to work gradually, over months. Trials, however, usually stop early and count recurrence as a single yes/no outcome at three months. Li et al. followed patients for exactly the blanking window and reported 9.26% recurrence with sacubitril/valsartan versus 23.91% with perindopril. A pooled analysis of sixteen trials found the same drug protects the lower chambers (the ventricles) but shows no significant effect on atrial fibrillation (meta-analysis). These two results are hard to reconcile, and neither analysis can say whether the benefit lasts.

We propose a trial analysis built around the time course that the biology predicts. Its central idea is simple. A recurrence during the blanking period should not be thrown away, and it should not be counted as failure either. It should be treated as a separate step in the patient's course, with its own treatment effect. The drug's effect on late recurrence is then estimated both from patients with no early episode and from patients who already had one. The model is a three-state model: ablation, then early (blanking) recurrence, then late recurrence. The boundary between phases is fixed in advance at the clinical blanking cutoff rather than estimated from the data. The model also accounts for the fact that recurrences are only detected at scheduled monitoring visits. The claim is that, with the 6–12 months of follow-up these trials can realistically afford, this design detects a delayed-onset stabilizing effect with more power than the analyses now in use, including the standard practice of simply ignoring blanking-period events. It also answers a question current analyses cannot: whether an early three-month win predicts late protection.

Core Claim

Upstream stabilizing drugs plausibly act on two separate things. The first is inflammation-driven early episodes, which Li et al. suggest the drug may already reduce. The second is substrate-driven late recurrence, the arrhythmias that arise from the structurally changed chamber tissue itself. These recur in 20–60% of patients (Liu et al.) and typically appear after 6 months (Yang et al.).

A single hazard ratio averages these two effects into one number. A hazard ratio is the ratio of the instantaneous event rates in the two arms. Averaging it is exactly the failure mode that Clement et al. documented in oncology: when a treatment's effect is delayed, the standard one-sample log-rank test drops below 30% power at n=80. Discarding the blanking period repairs the averaging but throws away information. Early recurrence is a known marker of fragile tissue, so a patient who has an early episode is at higher risk of the late recurrences the trial cares about.

The three-state model has three transitions: ablation to early recurrence, ablation to late recurrence, and early recurrence to late recurrence. Each gets its own treatment coefficient. Our primary quantity is the drug's effect on late recurrence, which the model estimates from both late-recurrence transitions and from the timing of all events. We expect this to reduce the variance of the late-effect estimate relative to a landmark analysis, which restarts the clock at the end of the blanking period and discards everything before it. We expect the gain to be largest when early recurrences are common, which is the regime Yang et al. observed (25.6%).

Research Program

The first component is methodological. We will specify the three-state model with piecewise-constant hazards, with phase boundaries at 3 months and optionally at 6 months. Event times will be interval-censored: a scheduled 24-hour Holter recording only shows that recurrence happened somewhere between two visits. Treating such times as exact is a quiet source of bias in short trials with sparse monitoring.

The primary test combines the drug's effects on the two late-recurrence transitions. We will pre-specify the whole analysis so the fixed blanking cutoff cannot be tuned after seeing the data. To generate simulated trials, we will reuse the approach of Zhang et al., who generate event times by inverting the cumulative hazard under piecewise, time-varying drug effects. That paper supplies our simulation machinery; its own application was antibody concentration cycles, not arrhythmia.

The second component is a simulation study calibrated to the numbers the file actually contains. The three-month recurrence rates come from Li et al.. The early-recurrence frequency comes from Yang et al.. Late-recurrence rates span the 20–60% range reported by Liu et al..

We will simulate six scenarios: no effect, an early effect only, a late effect only, both effects, crossing curves, and constant proportional hazards. Crossing curves would arise, for example, if the drug initially raises early events and then lowers late ones. Each scenario will be run at n=100 and n=300, with follow-up of 6 and 12 months and Holter monitoring every one or three months.

A secondary simulation tests the explanation for the conflict between the pooled analysis and the single trial. We will generate trials with follow-up spread from 8 weeks to 33.9 months, the range reported in the meta-analysis, and pool them with a single relative risk. The question is whether a true delayed atrial effect is diluted to non-significance when trials of such different lengths are combined this way.

The third component is a pre-registered analysis plan for a real trial. The plan extends the sacubitril/valsartan-versus-perindopril design of Li et al. to 12 months, with scheduled Holter monitoring. It adds one mediation check. The three-month change in left atrial diameter and NT-proBNP, a blood marker of heart strain, will be entered as a time-dependent covariate after the blanking period. We then ask how much of the late-transition effect it absorbs.

The motivation is the gap between structural and clinical outcomes that Hu et al. documented. In their pooled cardiac resynchronization studies, remodeling response improved (RR 1.33), but mortality did not change significantly (RR 0.69, p=0.24). We expect the surrogate to absorb only part of the effect, and we treat that outcome as informative rather than as a failure.

Baselines

The naive reference is the analysis actually used by Li et al.: a chi-square comparison of the proportion of patients with any recurrence at 3 months. It reported 9.26% versus 23.91% (p<0.05). We run it on identical simulated trials, and at 12 months for longer follow-up. It shows how often a win in the blanking window coexists with no late effect.

The conventional time-to-event baseline is a log-rank test with a single-hazard-ratio Cox model on all recurrences. It is the default reviewers expect.

The strongest clinical competitor is a landmark Cox analysis that ignores blanking-period events and restarts the clock at 3 months. It is what electrophysiologists would actually do. No reported result was found on Lacuna for this analysis on any dataset in the file, so its performance will come only from our simulations.

The strongest statistical competitor is the max-Combo test of Clement et al.. It combines several weighted log-rank tests and is robust to early, delayed, and crossing effects. In their delayed-effect simulations at n=80, it held power above 80%, while the standard test fell below 30%. We will adapt it from the single-arm setting to two arms. We will also include restricted mean survival time, the average recurrence-free time up to a fixed horizon, taken from the same paper.

For a fair comparison, every method sees the same simulated trials and the same interval-censored monitoring schedule. Each method gets the same pre-specification: no data-driven choice of weights or cutoffs is allowed, for any method, including ours. Type I error is calibrated to 5% for every method before power is compared.

Beating the competitors means clearly higher power than landmark Cox and at least matching max-Combo in the delayed and combined-effect scenarios, while keeping nominal type I error. It also means lower bias and mean squared error for the late-recurrence effect. Max-Combo gives a test rather than an effect estimate, so that part of the comparison is against landmark Cox and RMST.

Two ablations isolate the key ingredient. The first uses the same multi-state model but drops the early-recurrence-to-late-recurrence transition. If the gain disappears, early events carry the information. The second estimates the phase boundary from the data instead of fixing it. This tests whether clinical prior knowledge of the blanking window, rather than model flexibility, drives the result.

Why not just…?

Why not just use the standard landmark analysis after the blanking period? This is the most serious threat. If early recurrence carries little information about late recurrence, landmark Cox will match our model with less complexity. The first ablation and the comparison of the two methods across early-recurrence frequencies decide this directly. If the gain is small at a realistic 25% early-recurrence rate, we will recommend landmark Cox with interval-censoring correction and report our model as unnecessary.

Why not just use max-Combo or restricted mean survival time, which are already robust to delayed effects? Clement et al. show max-Combo recovers power, and it may match us on detection alone. It returns a yes/no answer, though. It says nothing about whether the drug acts on inflammatory episodes or on the chamber tissue itself, and it cannot test whether early wins predict late ones, which is the clinical question at stake. If max-Combo ties on power, the model's remaining justification is the separate early and late estimates. We will state that justification plainly rather than claim a power advantage.

Why not just rely on 3–6 month remodeling surrogates, or a deep survival model? Surrogates improved without any significant mortality benefit in Hu et al., so they cannot replace event follow-up. Our mediation check measures how much of the late effect they actually carry.

Deep survival models do outperform linear Cox models on rich imaging data. SSCAR reached a C-index of 0.74 versus 0.58 externally (Popescu et al.); the C-index measures how well a model ranks patients by risk. DeepSurv reached 0.867 versus 0.816 on METABRIC (Katzman et al.). These are prediction tasks on hundreds to thousands of patients. A 100-patient randomized trial needs a treatment-effect estimate that is valid under randomization, not better risk ranking, so we do not consider these models competitors here.

Decisive Checks

The proposal succeeds if three things hold across the calibrated simulations. At n=100 with 12-month follow-up, the three-state test beats landmark Cox in power in the delayed and combined-effect scenarios, while holding 5% type I error under the null and under the early-only scenario. Removing the early-to-late transition erases that gain. And pooling trials of varied length with a single relative risk reproduces a null atrial effect even when a true delayed effect exists.

The proposal fails, and we will say so, if landmark Cox or max-Combo matches it at realistic early-recurrence rates. In that case, the contribution shrinks to two recommendations: correct for interval censoring, and report blanking-window and late effects separately.