{"id":"94933af3-9e06-4001-98c7-2aeafd339333","arxiv_id":"2509.02898","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An RL-driven video acquisition policy for echocardiography keeps aortic stenosis classification at 80.6% balanced accuracy while acquiring only 47% of the videos on average.","lead":"This paper trains a reinforcement learning agent to choose which echocardiography videos to collect for each patient, then stops when it judges enough has been seen. On 2,572 patients the method kept 80.6% diagnostic accuracy while using about 47% of the videos, which could make point-of-care ultrasound faster and less expertise-dependent.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The efficiency claim lacks a random-subset baseline: since the classifier is trained with 50% random masking, random subsets of ~2 videos may already achieve 80.6%, making the RL agent's 47% video reduction uninformative.","rationale":"The reader's weakest assumption focuses on the four-clip construction as a deployment limitation. While that is a valid external validity concern, I identify a more load-bearing internal concern: the absence of a random or fixed subset baseline. The classifier is trained with random masking, making it robust to missing tokens; thus the ability to use fewer videos is already built into the classifier. Without showing that the RL agent selects subsets that outperform random subsets of equal size, the paper's claim that RL-driven personalized selection enables the 47% video reduction is not established. This is a critical missing control that directly affects the interpretation of the headline result. However, the paper is otherwise methodologically sound and the results are plausible; a conditional recommendation with a request for this baseline is appropriate. This is why I partially agree with the reader: we both question the practical significance, but I emphasize an internal validity issue that could be resolved with a simple experiment.","tokens_in":8078,"tokens_out":7166,"duration_ms":91724,"concrete_test":"Evaluate the frozen transformer classifier on the test set under three conditions: (1) the RL-selected subsets (as in the paper), (2) a random subset of two videos per patient (to match the average 1.88 selected), and (3) a fixed subset such as the first two videos (PLAX1, PSAX1). Report balanced accuracy, F1, and bMAE for each condition over the same 5 runs. If the random or fixed subset achieves bACC within the reported standard deviation of the RL result (e.g., 80.6±0.8), the RL agent's efficiency gain is not statistically distinguishable from random selection, and the central claim would need to be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that RL-driven active video acquisition reduces the number of videos to 47% while maintaining accuracy. However, the classifier used to evaluate this claim is explicitly trained with 50% random masking (Section 3.3), which makes it robust to missing videos by design. Therefore, it is expected that a classifier trained this way can maintain performance with only ~2 videos. The paper compares PRECISE-AS (with RL) against a full-acquisition baseline (w/o RL) and against other methods that use 100% of videos, but it never compares against a random or fixed subset of the same size. Without such a baseline, the observed 80.6% bACC with 1.88 videos on average could simply reflect the classifier's masking-based robustness, not the RL agent's ability to select informative videos. The RL agent's contribution to the efficiency gain is thus unverified. If random subsets of two videos achieve statistically equivalent accuracy, the 'personalized' selection is superfluous, and the efficiency result is an artifact of the classifier training scheme. This is a more direct internal validity issue than the four-clip deployment concern: it questions whether the claimed efficiency is attributable to the proposed method at all, even within the simplified experimental setup.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PRECISE-AS, an RL-driven active video acquisition framework for aortic stenosis (AS) severity classification from echocardiography. It encodes each of four pre-selected echo clips (PLAX1, PLAX2, PSAX1, PSAX2) with ProtoASNet, trains a Transformer classifier under 50% random token masking, and learns a DDQN policy that sequentially acquires or skips clips and terminates when confident. On a private dataset of 2,572 patients (252 test), the method reports 80.6% balanced accuracy while acquiring on average 47% of the videos (1.88 of 4 clips), matching the full-acquisition version (80.6%) and the ProtoASNet baseline (80.0%); the authors disclose that the improvement over ProtoASNet is not statistically significant (p > 0.05). The paper includes an ablation over the cost coefficient lambda and qualitative acquisition-pathway analysis.","tokens_in":8433,"tokens_out":4431,"duration_ms":50610,"significance":"If the efficiency claim survives scrutiny, the paper is a plausible contribution to active feature acquisition in high-dimensional video data, an under-explored area in medical imaging. The strengths are concrete: a large private clinical dataset with patient-level splits, held-out test set, repeated runs with standard deviations, honest reporting of the non-significant accuracy gap versus ProtoASNet, and public code. The central methodological idea—learning a personalized acquisition policy over frozen video-token embeddings—is reasonable. However, the load-bearing efficiency claim (47% video reduction at equal accuracy) is not yet internally validated, because no random-subset baseline is reported despite the classifier being explicitly trained for 50% missing tokens. The significance of the paper therefore depends on the outcome of that missing control experiment.","major_comments":[{"comment":"The central efficiency claim is not supported without a random-subset baseline. The classifier is deliberately trained with 50% random token masking (§3.3), so it is robust to missing approximately two of four videos by construction. If PRECISE-AS (with RL) acquires 1.88 videos on average, a trivial random or fixed subset of two videos (e.g., [PLAX1, PSAX1]) may achieve the same 80.6% bACC simply because the classifier was trained for the missing-token regime. Table 1 compares against 100%-video baselines only; Table 2 sweeps λ but never includes a same-cost random policy. The observed 47% figure could therefore be an artifact of the masking rate rather than of the learned selection policy. Please add random-k and fixed-subset curves across k=1..4, and report the performance gap between RL and random selection at matched average acquisition counts.","section":"§3.3, §4.3, Table 1"},{"comment":"The deployment claim is more limited than the abstract suggests. The MDP operates over exactly four pre-selected clips per patient ordered [PLAX1, PLAX2, PSAX1, PSAX2], with unit cost per clip. The 47% is therefore a reduction from 4 predetermined clips to 1.88 on average, not a reduction in the number of images a sonographer would capture in a real POCUS exam. View selection, clip count, variable clip quality, and operator time are not modeled. At a minimum, the title/abstract should say \"among four pre-selected echo clips,\" and the discussion should acknowledge this simplification as a boundary on the efficiency claim.","section":"§4.1"},{"comment":"The 50% masking rate and the cost coefficient λ are free parameters that jointly determine the efficiency result. λ=0.001 is selected post hoc; no principled selection criterion or sensitivity analysis is given, and the claim of \"using only 47%\" is one point on the λ sweep. Similarly, the masking rate is fixed at 50% without evidence that this rate matches the RL agent's typical acquisition policy. Please report performance over a range of masking rates, and define how λ is chosen in practice (e.g., by a validation-based accuracy/cost trade-off) before claiming the 47% operating point as the headline result.","section":"§3.3, Table 2"}],"minor_comments":[{"comment":"The notation m_i · f(x_i) with m_i a boolean scalar is unconventional; as written this is scalar multiplication, not a dot product. Replace with m_i f(x_i) or clarify.","section":"Eq. (1)"},{"comment":"The update s_{t+1}[i] = f(x_i) is dimensionally inconsistent if s is an N×D state; it should read s_{t+1}[i, :] = f(x_i) or otherwise state that the i-th row is replaced.","section":"§3.2, Eq. (2)"},{"comment":"DDQN hyperparameters are under-specified: no replay buffer size, ε-greedy schedule, target network update frequency, learning rate, or batch size. This limits reproducibility despite the code release.","section":"§4.2"},{"comment":"Because two PLAX and two PSAX videos are \"randomly selected\" per patient, the four-clip set is not deterministic. Please state whether all patients had exactly two usable clips per view, and report variance over the random clip selection; otherwise the 47% average count may itself be a random quantity.","section":"§4.1"},{"comment":"The left panel is described as F1 score versus average number of acquired videos, but the surrounding text sometimes refers to accuracy. Clarify the metric and the meaning of the shaded regions (standard deviation over runs, presumably).","section":"Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest and technically plausible, but the central efficiency claim is currently underdetermined. The missing random-subset baseline is the single most important experiment; if random two-video subsets achieve 80.6% bACC, the RL contribution is largely uninformative and the paper should be reframed accordingly. I recommend major revision rather than rejection because the issue is fixable within scope by adding the control experiments and tempering the deployment claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper applies RL-based active feature acquisition to echo video selection for AS, and that is new. The pipeline is clean: a frozen ProtoASNet encoder, a transformer classifier, and a DDQN that picks up to four videos per study. They train the classifier with 50% random token masking, which is sensible for robustness. On 252 held-out patients, the RL-selected ~1.9 videos match the full 4-video accuracy at 80.6% bACC, and the authors honestly note the improvement over ProtoASNet is not significant. The ablation over the cost coefficient is useful, and they release code.\n\nThe soft spot is real: there is no random-subset or fixed-subset baseline. Because the classifier is explicitly trained with 50% masking, it should handle any two videos reasonably well. The 47% figure could simply be that robustness, not the RL agent's skill at picking informative videos. The paper compares against full-acquisition baselines, but not against a random two of four or even the first two PLAX/PSAX. Adding that baseline is essential; without it, the headline efficiency claim is unsupported.\n\nAlso, the problem is intentionally simplified: each study is reduced to four predetermined clips [PLAX1, PLAX2, PSAX1, PSAX2], so the 47% is a reduction from four clips to 1.9, not a reduction in real-world acquisition. That is a fair simplification for a first pass, but it limits the point-of-care claims. The dataset is private, so external validation is on the authors.\n\nOverall, this is a sensible first application of AFA to video echo, and the missing baseline is fixable. Worth a serious referee, but the authors should be pushed to add random and fixed-subset comparators before the efficiency claim is published.","headline":"The 47% video-reduction claim lacks a random-subset baseline; since the classifier is trained with 50% token masking, random selection of two videos may already match full acquisition, so the RL agent's contribution is unverified.","tokens_in":8889,"tokens_out":2705,"would_cite":false,"duration_ms":31533,"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":"RL-driven active video acquisition matches full echocardiography accuracy while using 47% of clips","keywords":["aortic stenosis","echocardiography","point-of-care ultrasound","active feature acquisition","reinforcement learning","video classification","Double DQN","clinical decision support"],"falsifier":"Give the trained policy a real-time acquisition setting where candidate clips are not already present: an operator uses the policy's chosen view label to steer the probe, and the metric is total scan time or number of clips acquired until the policy terminates. If matching the 47% clip reduction requires acquiring more than 1.88 clips to obtain the selected ones (for example, because the first PLAX clip is unusable and must be retaken), the claimed efficiency gain fails to transfer. Alternatively, a prospective multicenter study could check whether the RL-selected subset yields 80.6% balanced","tokens_in":8068,"feed_emoji":"🫀","tokens_out":6579,"duration_ms":72422,"temperature":0.7,"pith_summary":"The paper tries to show that a reinforcement-learning agent can decide, patient by patient, which echocardiography video clips to keep and when to stop acquiring more, without hurting diagnostic accuracy for aortic stenosis. On 2,572 patient studies, the system reaches 80.6% balanced accuracy for three severity classes while using 47% of the clips a full study would use—about 1.88 of the four clips considered—and this matches both its own full-acquisition classifier and the state-of-the-art ProtoASNet baseline with no statistically significant difference. The authors argue this is the first active-video-acquisition framework for echocardiography and that it could make point-of-care ultrasound faster and more accessible by replacing fixed protocols with personalized stopping decisions. A reader should care because the result is a concrete demonstration that sequential, cost-aware image selection can be applied to high-dimensional medical video, not just tabular data or single static images.","feed_headline":"RL agent cuts echo clips to 47% at same diagnostic accuracy","feed_subtitle":"A policy trained on 2,572 patients uses about 1.88 clips per study instead of four while scoring 80.6% balanced accuracy.","key_machinery":"The load-bearing object is the active video acquisition MDP: state s is an N×D matrix holding the feature embeddings of selected videos with zero vectors for unselected positions; actions are acquire video i or terminate; and the only nonzero reward, R(sT)=1[correct]−λΣci, is given at termination. Double DQN learns the policy over this MDP using features from the frozen ProtoASNet encoder, while a Transformer with positional encodings and a CLS token performs study-level classification and supplies the correctness term. The design makes the problem a personalized stopping problem: when the expected diagnostic value of another clip no longer exceeds its cost, the agent terminates.","core_discovery":"On its own terms, the paper's central claim is that echocardiography video acquisition for aortic stenosis severity assessment can be modeled as a Markov decision process and solved with reinforcement learning to produce per-patient acquisition policies. Starting from a zeroed state over four fixed, ordered clips (two parasternal long-axis, two parasternal short-axis), the agent either selects one more clip or terminates; the terminal state is classified by a Transformer that treats selected clips as tokens and a missing clip as a masked token. The reward is +1 for a correct severity label minus a cost coefficient times the number of clips acquired. Trained with Double DQN on frozen features","pith_inferences":["The 47% figure counts savings over four pre-selected clips, not over live scanning time; if a sonographer must acquire and process more than the chosen clips to find good representatives of each view, the realized savings will be smaller than 47%.","One testable extension is to let the policy request a view label rather than a specific clip index and combine it with a real-time view detector, turning the system into a probe-guidance assistant that tells the operator which standard view to capture next.","The recipe—frozen feature encoder, masked-token Transformer classifier, and RL over acquisition actions—should transfer to other multi-view medical imaging or sequential diagnostic tests, where the same accuracy-versus-cost tradeoff exists.","Since the agent learns to prefer first PLAX then PSAX, a simpler non-RL protocol of fixed view ordering might achieve part of the savings; comparing RL against that fixed-order baseline would isolate the value of learning personalized stopping."],"forward_implications":["At the λ=0.001 operating point, the policy uses 47% of the four-clip acquisition (1.88 clips per study) at 80.6% balanced accuracy; the same classifier without RL needs all four clips to reach that accuracy.","The cost coefficient λ gives a tunable efficiency-accuracy frontier: λ=0.01 yields 41% of clips at 79.9% balanced accuracy, and λ=0.25 yields 31% at 77.1%, so a deployment could pick an operating point before scanning.","The learned acquisition order—typically a PLAX clip first, then a PSAX clip—is interpretable and consistent with complementary views of the aortic valve, suggesting the policy can expose underlying acquisition rules.","Because the policy matches the full-acquisition accuracy within statistical noise (p>0.05), the framework's practical value is efficiency at equal accuracy rather than improved accuracy."],"supporting_citations":[{"why":"Supplies the frozen ProtoASNet video feature extractor and is the state-of-the-art AS classification baseline that PRECISE-AS must match.","marker":"[21]"},{"why":"Establishes the reinforcement-learning approach to active modality selection in medical diagnosis and the sparse terminal reward scheme adapted by this paper.","marker":"[4]"},{"why":"Double DQN is the value-based RL algorithm used to learn the acquisition and termination policy.","marker":"[20]"},{"why":"The Transformer encoder with a CLS token is the architecture used to aggregate selected video tokens into a study-level classification.","marker":"[22]"},{"why":"View-detection algorithm automatically identifies PLAX and PSAX cine clips from echo studies, forming the pool of candidate videos the agent selects from.","marker":"[13]"},{"why":"Transformer-based spatio-temporal AS classifier used as a video-level baseline for comparison on the same private dataset.","marker":"[1]"}],"fun_headline_variants":["RL picks echo clips: 47% fewer videos, same accuracy","Personalized RL trims echo videos to 47% without losing accuracy","Reinforcement learning cuts echo scans by half for AS diagnosis","Per-patient clip selection: AI needs only 47% of echo videos","Smarter echo: RL matches full acquisition at nearly half the videos"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The efficiency result rests on modeling a study as exactly four predetermined clips—two PLAX and two PSAX chosen in advance, each with unit cost—so the 47% reduction is a saving over four clips, not over the number of video clips a sonographer would acquire in a live point-of-care exam.","fun_headline_variants_meta":{"raw":{"variants":["RL picks echo clips: 47% fewer videos, same accuracy","Personalized RL trims echo videos to 47% without losing accuracy","Reinforcement learning cuts echo scans by half for AS diagnosis","Per-patient clip selection: AI needs only 47% of echo videos","Smarter echo: RL matches full acquisition at nearly half the videos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1135,"prompt_tokens":744,"completion_tokens":391,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":299}},"tokens_in":488,"tokens_out":391,"duration_ms":4914,"temperature":1.0,"reasoning_tokens":299,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:16:40.095948+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give the trained policy a real-time acquisition setting where candidate clips are not already present: an operator uses the policy's chosen view label to steer the probe, and the metric is total scan time or number of clips acquired until the policy terminates. If matching the 47% clip reduction requires acquiring more than 1.88 clips to obtain the selected ones (for example, because the first PLAX clip is unusable and must be retaken), the claimed efficiency gain fails to transfer. Alternatively, a prospective multicenter study could check whether the RL-selected subset yields 80.6% balanced","supporting_citations":[{"cited_title":"In: Medical Image Computing and Computer Assisted In- tervention – MICCAI 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen ProtoASNet video feature extractor and is the state-of-the-art AS classification baseline that PRECISE-AS must match."},{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2022","cited_arxiv_id":null,"evidence_quote":"Establishes the reinforcement-learning approach to active modality selection in medical diagnosis and the sparse terminal reward scheme adapted by this paper."},{"cited_title":"IEEE Transactions on Medical Imaging39(6), 1868–1883 (2020)","cited_arxiv_id":null,"evidence_quote":"View-detection algorithm automatically identifies PLAX and PSAX cine clips from echo studies, forming the pool of candidate videos the agent selects from."},{"cited_title":"IEEE Transactions on Medical Imaging 43(1), 366–376 (2024)","cited_arxiv_id":null,"evidence_quote":"Transformer-based spatio-temporal AS classifier used as a video-level baseline for comparison on the same private dataset."}],"review_version":1}