{"id":"e22bacb3-7167-43ee-b5ad-23cb586e5222","arxiv_id":"2608.13518","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An event-conditioned latent world model forecasts AF recurrence (AUROC 0.756) and scar extent (MAE 2.971 percentage points) after ablation using blanking-period events and ECG embeddings.","lead":"This paper proposes a machine-learning model that tracks a patient's recovery after atrial-fibrillation ablation using heart scans, procedure details, and irregular clinical events such as medication changes and cardioversions. The model predicts long-term recurrence better than standard sequence baselines in an internal dataset, and estimates scar extent without needing a follow-up MRI at prediction time.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 90-day event input window overlaps the 451-day outcome window: for 32 early-recurrence patients, post-outcome events leak into the predictors, so the AUROC gain may reflect leakage rather than learned dynamics.","rationale":"The reader's weakest assumption concerns the temporal alignment of the latent matching loss in Eq. (5): the forecast horizon t_T is not the follow-up MRI time, so matching \\hat z_T to z_post may distort the learned dynamics. That is a legitimate training-signal concern, but I judge the outcome-window overlap to be more load-bearing. The central claim is the AUROC/AUPRC improvement over a matched-input LSTM and the interpretation that state evolution matters. If the event inputs include observations that occur after the outcome for 35% of positive patients, the evaluation is contaminated: the model can learn to detect the consequences of early recurrence rather than to forecast long-term risk. This directly threatens the validity of the headline result and the mechanistic conclusion drawn from it. I recommend retaining the CONDITIONAL verdict, but with a sharper condition: the authors should re-run the evaluation under a blanking-period-adjusted endpoint or with event censorship at the recurrence date, and show that the performance gain persists. If it does not, the paper's central claim should be revised. The reader's latent-matching concern should also be checked, but it is secondary because it affects training signal quality rather than the fundamental validity of the prediction task.","tokens_in":7708,"tokens_out":6317,"duration_ms":65460,"concrete_test":"Recompute the Table 2 evaluation with a leakage-free protocol: redefine the endpoint as recurrence after the 90-day blanking period (the standard in AF ablation trials), or equivalently exclude the 32 patients with recurrence ≤90 days and censor each remaining patient's event sequence at the time of their earliest recurrence. Compare the full model's AUROC/AUPRC to the matched-input LSTM under the same protocol. If the margin shrinks substantially from 0.103/0.090 toward zero, the headline improvement is attributable to temporal leakage rather than to learned latent state evolution.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing concern is temporal leakage between the input window and the outcome window. Section 3 states that the model forecasts 451-day AF recurrence y using observations available by a query horizon, with clinical events censored at day 90. Fig. 3 then stratifies patients into early recurrence (≤90d, n=32), late recurrence (>90d, n=14), and no recurrence (n=45). For the 32 early-recurrence patients, the outcome occurs inside the 90-day input window. Any medication change, cardioversion, or repeat procedure recorded after the recurrence date but before day 90 is a post-outcome observation caused by the recurrence itself. The model therefore has access to direct evidence that the event of interest has already happened, and the reported AUROC 0.756 may reflect detection of that aftermath rather than forecasting of future recurrence. The matched-input LSTM receives the same leaked events, so the claimed 0.103 AUROC gain and the conclusion that 'state evolution, not input access alone, matters' (Section 4.2) are not established; the gain may only indicate that the world model better exploits post-outcome events. The limitations section discusses confounding and missingness but does not address this temporal overlap. This concern is more direct than the latent-matching misalignment in Eq. (5): even a perfectly aligned zpost would not fix label leakage from the event sequence.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an intervention-aware clinical world model that represents a post-ablation patient as a 3D latent anatomical state, updates that state through time-ordered blanking-period events (medication, cardioversion, repeat procedures), and uses a terminal horizon token to produce recurrence risk and scar-extent forecasts at different query times. The model is evaluated on DECAAF-II with internal 5-fold cross-validation on N=91 complete-record patients and on an auxiliary N=258 cohort without ablation geometry. The central empirical claims are an AUROC of 0.756 and AUPRC of 0.777 for 451-day recurrence prediction, a scar-extent MAE of 2.971 percentage points without follow-up MRI at inference, and a 0.103 AUROC gain over a matched-input LSTM that the authors attribute to the learned state evolution rather than to input access.","tokens_in":7988,"tokens_out":5745,"duration_ms":62562,"significance":"If the forecasting claims survive a leak-free re-evaluation, the paper would be a useful contribution to clinical risk updating: it models the post-ablation blanking period as an irregular event sequence, couples imaging and procedural geometry with ECG embeddings, and provides explicit multi-horizon queries. The paper has notable strengths: the follow-up MRI latent is used only as training supervision, the input-edit analysis is explicitly labeled as associational rather than causal, the limitations section is candid about cohort size and internal validation, and ablations separate the contributions of events, ECG, ablation map, and latent matching. However, the current evaluation has a temporal overlap between the input window and the outcome window, and this concern is directly supported by the manuscript's own Fig. 3. The latent-matching objective also has a temporal-alignment ambiguity with respect to the forecast horizon. Both issues bear on the central claim that the model forecasts recurrence rather than detecting its early aftermath, so the significance of the result is currently conditional on a revised evaluation.","major_comments":[{"comment":"The evaluation protocol has a temporal overlap between input and outcome windows. Section 3 states that clinical events are censored at day 90, and Fig. 3 defines early recurrence as recurrence by day 90 (n=32). For these 32 patients, blanking-period records after the recurrence date (medication changes, cardioversion, repeat procedures) are consequences of the outcome, not predictors available before it. The reported AUROC 0.756 and the 0.103 improvement over the matched-input LSTM in Table 2 may therefore reflect detection of the aftermath of an already-occurred recurrence rather than forecasting of future recurrence. The Section 4.2 conclusion that 'state evolution, not input access alone, matters' is not established by the current numbers. Please re-run the evaluation under a leak-free protocol: either exclude patients whose recurrence precedes the query horizon, or define recurrence as occurring after the 90-day blanking period, and report the resulting AUROC and AUPRC for the model and for the matched baselines.","section":"Section 3, Fig. 3, Table 2"},{"comment":"The latent-matching objective is not aligned with the variable forecast horizon. Eq. (5) minimizes the squared error between the predicted terminal latent z_T and the follow-up latent z_post, while Section 3 states that T is the forecast horizon, not the follow-up scan time. If the follow-up scan is acquired at a fixed clinical time and the model is queried at horizons ranging from 7 to 210 days, the loss supervises the state at time t_T with a latent from a different time point. This is load-bearing because Table 4 shows that removing the latent matching term reduces AUROC by 0.132 and raises the standard deviation to 0.200. Please state the distribution of follow-up scan times, condition the matching term on the actual scan time, or verify that the reported results are robust to this misalignment.","section":"Section 3, Eq. (5)"}],"minor_comments":[{"comment":"The ablation 'w/o terminal token' is not fully specified: because Eq. (1) already includes the horizon time (t_T - t_abl)/180 in every event token, it is unclear whether the ablation removes only the appended terminal token or also removes horizon information from the event tokens. Please clarify what is actually removed.","section":"Table 4"},{"comment":"The elapsed-time scaling tau = 30 days and the ECG pooling window of 7 days are introduced without a sensitivity analysis. Please add a sentence justifying these constants or report a sensitivity check, since the model's dynamic behavior depends on the time scaling.","section":"Section 3, Eq. (3)"},{"comment":"The phrase 'OOF Brier score 0.201' is not defined. Please clarify that it is an out-of-fold Brier score and report the folds and seeds over which it is averaged.","section":"Section 4.2"},{"comment":"Table 2 would be easier to interpret if the query horizon used for the main recurrence prediction were stated explicitly (e.g., day 90), because the anytime analysis in Fig. 3 shows that the model's separation depends on the horizon.","section":"Table 2"},{"comment":"The text describes the input-edit analysis as testing 'the model's ability to compare alternative blanking-period scenarios.' Since the edits are explicitly associational probes, consider rewording this to avoid any implication that the model computes treatment counterfactuals.","section":"Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is the temporal overlap between the 90-day event window and the recurrence outcome, which is not mentioned in the limitations section. The latent-matching misalignment in Eq. (5) is secondary but should also be addressed. If the leak-free re-analysis does not preserve the reported gains over the matched LSTM, the paper's central forecasting claim would not be supported; if the gains persist, a revised version could be acceptable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real attempt to couple 3D LGE-MRI, ablation geometry, irregular blanking-period events, ECG embeddings, and horizon-token queries. That combination is new, the ablations are informative, and the internal evaluation is honestly reported. But I do not trust the headline AUROC as a forecasting number until the event-window / outcome-window overlap is fixed.\n\nWhat is actually new: the event-conditioned latent state with horizon tokens, and the training-only follow-up latent matching objective. Table 4 is the most useful part of the paper: removing events or latent matching costs roughly 0.13 AUROC, and the stability gain from latent matching (SD drops from 0.200 to 0.051) is plausible. The scar MAE of 2.971 percentage points is competitive with the oracle-style post-MRI baseline, though the margin is within fold noise. The authors also deserve credit for explicitly labeling the input-edit experiments as associational, not causal.\n\nThe soft spot I keep coming back to is temporal leakage. Outcomes are binary recurrence within 451 days, and 32 of 91 patients have early recurrence at or before day 90, which is inside the 90-day event-input window. For those patients, any medication change, cardioversion, or repeat procedure recorded after the recurrence but before day 90 is a post-outcome observation caused by the recurrence itself. The model sees it. The matched-input LSTM also sees the same leaked events, so the 0.103 AUROC gain and the claim that \"state evolution, not input access alone, matters\" are not established. The gain may only mean the world model is better at exploiting post-outcome events. The limitations section discusses confounding and missingness but does not mention this overlap. That is a load-bearing flaw, not a minor caveat.\n\nThe latent-matching timing issue is real but secondary. Equation (5) matches the predicted terminal latent at forecast horizon T to the follow-up MRI latent z_post, while the paper states T is not the follow-up scan time. If the follow-up MRI was acquired at a different time than the queried horizon, the training target is misaligned. This should be addressable with a time-conditioned target or a clear statement about scan timing, but it is not reconciled.\n\nCode and data are not available, so the result cannot be independently re-run. On balance, the architecture and ablation study are worth a serious referee. The main result needs re-analysis with an outcome definition that excludes blanking-period recurrences, or with events censored at the recurrence date. If the gain survives that re-analysis, it is a solid contribution; as reported, I would not cite the headline number.","headline":"A thoughtfully assembled world-model paper with informative ablations, but the headline 0.756 AUROC may partly reflect label leakage from events recorded after early recurrences; the main claim needs re-analysis.","tokens_in":8569,"tokens_out":2604,"would_cite":false,"duration_ms":28695,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that evolving a 3D latent anatomical state through irregular post-procedure events forecasts atrial fibrillation recurrence at AUROC 0.756, outperforming matched-input sequence baselines.","keywords":["clinical world model","atrial fibrillation","catheter ablation","latent state dynamics","anytime risk prediction","LGE-MRI","event-conditioned forecasting","recurrence prediction"],"falsifier":"Re-run the internal 5-fold cross-validation with the latent matching loss applied only to patients whose follow-up MRI was acquired within a few days of the queried horizon, and compare AUROC with the reported 0.756; a large drop would indicate the gain depends on temporally misaligned supervision.","tokens_in":7454,"feed_emoji":"🫀","tokens_out":3482,"duration_ms":34027,"temperature":0.7,"pith_summary":"The paper argues that post-operative recovery is not a one-step mapping from a baseline scan to a final outcome, but an irregular trajectory of clinical events that should continuously update risk. It builds a model that encodes the pre-ablation MRI into a 3D latent anatomical state, then evolves that state through medication changes, cardioversions, repeat procedures, and ECG context, and finally predicts 451-day atrial fibrillation recurrence. On the DECAAF-II cohort, the model reports AUROC 0.756 and AUPRC 0.777 for recurrence prediction, an improvement of 0.103 AUROC over a matched-input LSTM baseline. It also predicts follow-up scar extent to a mean absolute error of 2.971 percentage points without receiving follow-up MRI intensities at inference. If these results hold, the model would support dynamic, anytime risk updating during the 90-day blanking period rather than relying on a single baseline prediction.","feed_headline":"Latent world model lifts post-ablation AF risk to 0.756 AUROC","feed_subtitle":"Event-conditioned state evolution beats matched-input LSTM by 0.103 AUROC on DECAAF-II recurrence forecasting.","key_machinery":"The central object is the event-conditioned latent clinical world model: a 3D spatial latent state $z_0 = E_\\theta(x_0)$ initialized from the pre-ablation MRI, updated by a residual 3D CNN $f_\\phi$ and a context-dependent drift MLP $g_\\phi$ at each irregular event token. A horizon token sets the query time $t_T$, and the training objective includes a latent matching loss $\\lambda_z \\| \\hat{z}_T - z_{\\text{post}} \\|_2^2$ that pushes the predicted terminal latent toward the follow-up MRI latent, providing training-only structural supervision. This machinery lets the model produce anytime risk estimates and retrospective input edits while never requiring the follow-up MRI at inference.","core_discovery":"The central discovery is that an event-conditioned latent state, rather than simply more input features, drives the improvement in post-ablation recurrence forecasting. The model represents each patient's atrial anatomy as a 3D spatial latent vector obtained from a frozen VAE, and updates it stepwise through tokens encoding ablation geometry, static covariates, elapsed time, and peri-event ECGFounder embeddings. The terminal token sets the query horizon, allowing the same trained model to output risk at different times by changing only that token. Reported internal cross-validation on 91 complete-record patients yields AUROC 0.756 and AUPRC 0.777, and ablations show that removing the latent matching loss or the event sequence produces the largest performance drops, indicating that both the structural supervision and the event-driven state evolution are load-bearing.","pith_inferences":["A natural extension, not in the paper, is to condition the latent matching loss on the actual follow-up scan time; the current training target $z_{\\text{post}}$ may come from a scan acquired at a different time than the queried horizon, and testing this alignment directly could reveal whether the reported gains are robust.","The horizon-token formulation effectively turns the model into an anytime risk updater; one could extend it to continuous-time queries by interpolating the terminal token's time embedding, a change the authors do not explore.","The input-editing sensitivity results are framed as associational probes, but they hint at a potential planning tool: if the latent dynamics are trustworthy, one could search over event sequences to identify interventions that most reduce predicted recurrence risk, though causal validation would be required."],"forward_implications":["Clinicians could query recurrence risk at any chosen horizon using only records available by that time, enabling dynamic updates as blanking-period events accrue.","The model's ability to forecast scar extent without follow-up MRI at inference could reduce the need for routine follow-up imaging in risk stratification.","The learned latent state supports retrospective 'what-if' editing of blanking-period records, allowing clinicians to compare alternative event scenarios for a patient.","The same event-conditioned latent architecture could transfer to other interventions where irregular post-procedure events plausibly affect long-term outcomes."],"supporting_citations":[{"why":"Supplies the DECAAF-II dataset with paired pre/post LGE-MRI, ablation points, blanking-period events, and recurrence outcomes that the model is evaluated on.","marker":"[14]"},{"why":"Provides the ECGFounder embeddings used as peri-event physiological context in each event token.","marker":"[12]"},{"why":"Introduces SOFA, an action-conditioned scar simulation baseline that this work extends by modeling the post-procedural blanking period as an event sequence.","marker":"[7]"},{"why":"Defines the clinical 90-day blanking period after atrial fibrillation ablation, which motivates the time window over which events are collected.","marker":"[5]"},{"why":"Provides the world-model conceptual foundation of learning latent state transitions for forecasting, which the paper adapts to a clinical setting.","marker":"[9]"}],"fun_headline_variants":["Event-conditioned latent state boosts post-ablation AF risk prediction","World model uses clinical events to forecast post-op cardiac recurrence","Intervention-aware latent model hits 0.756 AUROC for AF recurrence","Event-driven state evolution improves post-ablation risk forecasting","Same latent model answers AF risk at any time after ablation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training loss forces the predicted latent at the query horizon to match the follow-up MRI latent, but the model never conditions on the actual follow-up scan time, so the target may come from a different time point than the queried horizon, which could distort the learned dynamics.","fun_headline_variants_meta":{"raw":{"variants":["Event-conditioned latent state boosts post-ablation AF risk prediction","World model uses clinical events to forecast post-op cardiac recurrence","Intervention-aware latent model hits 0.756 AUROC for AF recurrence","Event-driven state evolution improves post-ablation risk forecasting","Same latent model answers AF risk at any time after ablation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000589,"raw_usage":{"total_tokens":2758,"prompt_tokens":930,"completion_tokens":1828,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":1743}},"tokens_in":546,"tokens_out":1828,"duration_ms":11662,"temperature":1.0,"reasoning_tokens":1743,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:13:19.048370+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the internal 5-fold cross-validation with the latent matching loss applied only to patients whose follow-up MRI was acquired within a few days of the queried horizon, and compare AUROC with the reported 0.756; a large drop would indicate the gain depends on temporally misaligned supervision.","supporting_citations":[{"cited_title":"Jama327(23), 2296–2305 (2022) 10 Y","cited_arxiv_id":null,"evidence_quote":"Supplies the DECAAF-II dataset with paired pre/post LGE-MRI, ablation points, blanking-period events, and recurrence outcomes that the model is evaluated on."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Introduces SOFA, an action-conditioned scar simulation baseline that this work extends by modeling the post-procedural blanking period as an event sequence."},{"cited_title":"Ep Europace20(1), e1–e160 (2018)","cited_arxiv_id":null,"evidence_quote":"Defines the clinical 90-day blanking period after atrial fibrillation ablation, which motivates the time window over which events are collected."},{"cited_title":"Nature pp","cited_arxiv_id":null,"evidence_quote":"Provides the world-model conceptual foundation of learning latent state transitions for forecasting, which the paper adapts to a clinical setting."}],"review_version":1}