When to Restart Blood Thinners After a Bleed: Learning Patient-Specific Resumption Timing

Patients who take blood thinners (antiplatelet drugs such as aspirin and clopidogrel, or anticoagulants such as apixaban or warfarin) sometimes bleed, most often from the stomach or gut. Doctors then face a recurring question: when should the drug be restarted? Stopping it lowers the chance of bleeding again. But the reports in this file agree that stopping it fully and for a long time causes a rebound in clots, and the strokes and heart attacks that follow can kill more patients than the original bleed did (Management of acute upper gastrointestinal bleeding, Asia-Pacific working group consensus on non-variceal upper gastrointestinal bleeding: an update 2018). Machine-learning work on this trade-off answers a different question. It predicts who will bleed or clot under whatever treatment they happened to get. It does not predict what would happen if the drug were restarted on day 3 rather than day 10.

We propose to learn a resumption policy for each patient. Each hospital day after a bleed is treated as a decision point. At that point, a model estimates two quantities for each candidate restart day: the risk of bleeding again and the risk of a clot-related event. The model is trained with adjustment for the fact that sicker patients are restarted later. It reads both structured data and evidence pulled from clinical notes, such as endoscopy findings, transfusion needs, and a stable hemoglobin. We compare the resulting policy against four alternatives using off-policy evaluation, which estimates from historical records how a decision rule would have performed had it been followed. The four alternatives are guideline-style fixed restart days, a risk-threshold rule built on a standard predictor, and a tool-using language-model agent.

Core Claim

The decision clinicians actually face is the choice of a time to restart, and the existing literature does not model it. The closest work is AI-aided dynamic prediction of bleeding and ischemic risk after coronary stenting and subsequent DAPT. It updates gradient-boosted risk predictions at 3, 6, 12 and 18 months after a coronary stent. It beats the static DAPT score (ischemic AUROC 0.82 vs 0.79; bleeding 0.77 vs 0.72, Cerner HealthFacts). Its follow-up work, Contrastive learning with transformer for adverse endpoint prediction in patients on DAPT post-coronary stent implantation, reaches time-dependent concordance of 0.88 at one month for ischemia and 0.82 for bleeding on OneFlorida+.

Both papers make the same two simplifications. First, they predict outcomes under the treatment that was observed, so a patient who was sensibly kept off the drug looks low-risk for bleeding. That confuses the doctor's choice with the patient's biology. Second, they use only billing codes, and both name the missing clinical notes as their main limitation.

Our claim is that a model which takes the restart day as an input, and is fit with sequential confounding adjustment on features that include note-derived evidence, produces a resumption policy with lower estimated 90-day harm than three alternatives:

  • the best single fixed restart day;
  • a threshold rule on a well-tuned observational risk predictor;
  • a frontier language-model agent.

We pre-commit to reporting whether the per-patient policy beats the best fixed day at all. That comparison is the result that matters clinically.

Research Program

Cohort. We build the cohort in MIMIC-IV, chosen because it combines medication administrations, labs, vital signs, notes and linked out-of-hospital death dates. Patients enter at a hospital admission for gastrointestinal bleeding while taking an antiplatelet or anticoagulant drug. Admissions for planned procedures where the drug was held form a secondary cohort. Preprocessing follows An Extensive Data Processing Pipeline for MIMIC-IV.

Outcomes. These are recorded over 90 days:

  • re-bleeding (readmission or a transfusion-requiring event);
  • clot-related events (heart attack, ischemic stroke, venous clot, stent thrombosis);
  • death.

These are modeled as competing risks, meaning that the occurrence of one event changes or prevents observation of the others. Discrete-time cause-specific hazard losses come from Discrete Survival Knowledge Distillation for Competing Risks Analysis, which applied them to cancer and transplant outcomes but never to the paired risks of bleeding and clotting.

We state one limitation up front. MIMIC-IV observes non-fatal events only when the patient returns to the same hospital system. We therefore treat all-cause death as the primary endpoint and report the non-fatal events as secondary endpoints with known under-capture.

Note-derived evidence. Narrative evidence is extracted from notes as time-stamped events, following the LLM annotation method of MIMIC-IV-Ext-22MCTS, which produced 22.5M event–timestamp pairs with 0.951 correlation to expert timing. We restrict extraction to a closed vocabulary of resumption-relevant findings, for example "clean-based ulcer," "melena resolved," "high-risk stigmata," and "new chest pain." This turns the agent-versus-features question into a controlled comparison rather than a confound.

Model. At each day $t$ the model receives the history $H_t$ and a candidate restart day $a$. It outputs cause-specific discrete hazards for each outcome under $a$. Training uses inverse-probability-of-treatment weights from a model of the clinicians' daily restart decisions. A doubly robust correction, which stays consistent if either the outcome model or the treatment model is correct, is then applied when estimating policy value. This is the discrete-time version of the estimators that Dynamic Treatment Effects under Functional Longitudinal Studies derives in continuous time. That paper's estimators were tested only in simulation, so it supplies our method rather than motivating the problem.

Policy. The policy restarts the drug at the candidate day with the lowest estimated composite harm:

$$\hat a(H_t)=\arg\min_{a\in\mathcal{A}}\ \hat P(\text{death}_{90}\mid H_t,a)+\lambda\,\hat P(\text{non-fatal event}_{90}\mid H_t,a)$$

Here $\mathcal{A}$ is the set of restart days 1–14 that historical practice actually supports, and $\lambda$ is a clinician-elicited weight for non-fatal events. We report results across a range of $\lambda$.

Explanations. To make day-to-day changes in the recommendation inspectable, we apply SWING from Delta-XAI (AITRICS/Delta-XAI). SWING attributes the change in a prediction between two time steps to the inputs that changed, integrating along the patient's real history rather than an artificial all-zero patient. A clinician can then see, for instance, that the recommended restart moved earlier because the hemoglobin stabilized.

Baselines

Fixed restart days. The naive reference is a fixed-day rule: restart on day $k$ for every patient, evaluated for each $k$ in $\mathcal{A}$. We pick the best $k$ by doubly robust value, and the same estimator scores every policy. This is the target-trial-emulation answer that guidelines approximate. It is the bar that personalization has to clear.

Observed clinician policy. We evaluate the restart decisions clinicians actually made, which tells us whether any learned rule improves on current practice.

Risk-threshold rule (strongest existing method). This baseline restarts the drug once a predicted bleeding risk falls below a threshold. The predictor is gradient boosting in the style of AI-aided dynamic prediction of bleeding and ischemic risk after coronary stenting and subsequent DAPT. That paper reports AUROC 0.82 for ischemia and 0.77 for bleeding on Cerner HealthFacts. Those numbers come from a different cohort and task, so they serve as context rather than a target. We found no reported result on Lacuna for any method on post-bleed resumption policy value in MIMIC-IV.

To keep the comparison fair, the baseline receives the identical feature set, including the note-derived events. It gets the same hyperparameter search budget as our model, and its threshold is tuned on the validation split for policy value. Our method beats it only if the improvement in estimated 90-day mortality-weighted harm has a bootstrap confidence interval that excludes zero.

Language-model agent. This baseline is a tool-using agent in the style of ClinSeekAgent, run on the same frozen database snapshot. It sees the record day by day and outputs a restart day. That paper's agent reached 62.6 F1 versus 47.5 with curated inputs on ClinSeek-Bench, which is a different task.

Ablations. Two ablations isolate our key ingredients. The first removes confounding adjustment: an unweighted model that still takes the restart day as input. The second removes the note-derived evidence.

Why not just…?

Why not just add note features to a gradient-boosted risk model? This is the most serious threat, and the design tests it directly by giving the risk-threshold baseline the same note features. A risk model answers "is this patient likely to bleed?" under observed care. Patients restarted late are the ones clinicians judged fragile, so the risk model mixes a protective treatment effect into its estimates. If confounding by indication turns out to be weak in this cohort, the threshold rule will match our policy. In that case we will report that the simpler tool suffices.

Why not just ask a frontier LLM agent? Neural-MedBench found that 51% of model errors were reasoning failures. It also found strong anchoring, where a model fails to revise its first hypothesis when new longitudinal data arrive. ClinSeekAgent reports that agents get distracted on narrow decision tasks. We test anchoring directly by measuring how often the agent's recommendation moves after disconfirming evidence appears, compared with how often our model's recommendation moves.

Why not just a target-trial emulation of fixed restart days? The fixed-day baseline is exactly this. If its best day is within noise of our policy, personalization adds nothing. That would itself be useful to know, because it would justify simple guidance.

Decisive Checks

Primary result. The primary result is the doubly robust estimate of 90-day harm for each policy, with patient-level bootstrap intervals, on a temporally held-out test split (later admission years).

Assumption checks. Three checks guard against the main failure modes of the causal estimates:

  • Overlap: we report the distribution of treatment-model weights and truncate candidate days that have little support.
  • Unmeasured confounding: we report E-values for the main contrast, which give how strong a hidden confounder would need to be to explain the effect away.
  • Estimator calibration: we run a negative-control outcome that the restart day should not affect.

Note evidence and anchoring. We stratify the comparison with and without notes by whether endoscopy findings appear only in narrative text. For the agent comparison, we compute the revision rate after hemoglobin stabilization or endoscopic hemostasis. That rate shows whether the anchoring failure documented on neuroimaging cases carries over to antithrombotic decisions.