{"id":"93ddb6e7-9866-40f4-a55c-38c40c9821a2","arxiv_id":"2607.25129","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"LATCN (multi-layer TCN + label-wise attention) improves F1 ~9% and recall ~28% over DR-CAML on the top-50 MIMIC-III ICD coding task.","lead":"A 2019 neural model pairs multi-layer temporal convolutions with per-label attention to assign ICD codes from hospital discharge notes. On the 50 most common MIMIC-III labels it raises F1 about 9% and recall about 28% over the then-SOTA DR-CAML, which matters for clinical decision-support tools that must not miss codes.","discovery_kind":"extension","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The headline gains (28% recall, 9% F1) are measured at a single, unstated decision threshold in one training run, while precision dropped 7% — consistent with an operating-point shift rather than a true capability gain over DR-CAML.","rationale":"The reader's weakest assumption (top-50-only evaluation limits transfer to real coding workloads) is valid and I agree the paper's claim is bounded by it — the authors themselves defer full-label evaluation to future work. But that concern limits the claim's *generality*, not its *truth on the stated benchmark*. The more load-bearing issue for the strongest claim as stated is whether the reported numbers reflect a genuine model improvement at all. The simultaneous recall-up/precision-down pattern, the never-stated binarization threshold, the much smaller AUC gap, the absence of any variance estimate across runs, and the apparently transcribed (not re-run) DR-CAML baseline together form a coherent alternative explanation: LATCN's advantage may be substantially an operating-point and evaluation-hygiene artifact. This is not an accusation of impropriety — the numbers may well hold up — but the paper provides none of the evidence (PR curves, threshold-tuned F1, seed variance, re-run baselines) that would rule the alternative out. Hence agreement with the reader is partial: same family (robustness of the reported gains), different mechanism. I do not adjust the verdict: CONDITIONAL remains right, since the benchmark comparison is plausible and clearly described, but acceptance as a demonstrated advance requires exactly the threshold-sweep/multi-seed/baseline-rerun check above. The work is honestly framed (explicit 2019 historical note), and the architecture is a reasonable recombination of cited components; the concern is purely about what the single-run, single-threshold numbers license.","tokens_in":7240,"tokens_out":2149,"duration_ms":21064,"concrete_test":"Reproduce both LATCN and DR-CAML on the identical MIMIC-III top-50 split/preprocessing, train LATCN with ≥5 random seeds, and for each model sweep the sigmoid decision threshold from 0.05 to 0.95, plotting precision–recall curves and reporting max-F1 and recall at matched precision. If DR-CAML reaches micro-recall ≈0.70 at some threshold with micro-F1 within one seed-standard-deviation of LATCN's 0.681, the headline recall/F1 claim is largely an operating-point artifact; if LATCN dominates the full PR curve beyond seed noise, the claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a metric comparison: micro-F1 0.681 vs 0.633 and micro-recall 0.710 vs 0.584 against DR-CAML. Three compounding weaknesses make this less secure than the reader's scope concern. (1) Threshold confound: the model outputs per-label sigmoid probabilities (Eq. 4), so F1/recall depend on a binarization threshold the paper never states. LATCN simultaneously gains 12.6 recall points and loses 3.6 micro-precision points (0.655 vs 0.691) — exactly the signature of a less-conservative operating point, not necessarily a better ranker. The authors themselves attribute the precision drop to \"generalization,\" but never show PR curves, max-F1 over thresholds, or AUC-PR, any of which would disentangle ranking quality from calibration. Notably, the threshold-insensitive metric shows a much smaller gap: micro-AUC 0.931 vs 0.916. (2) Statistical robustness: results are a single 200-epoch run with grid-searched hyperparameters selected on dev; no seed variance, no error bars, no significance test. Seed-to-seed variation of 1–2 F1 points is routine on this benchmark, which would materially erode the 4.8-point micro-F1 gap. (3) Baseline provenance: the DR-CAML row carries a citation to Mullenbach et al. (2018), suggesting numbers transcribed from the original paper rather than re-run under identical preprocessing; any mismatch in tokenization, split, or truncation silently inflates or deflates the comparison. If the recall advantage collapses once thresholds are matched, the paper's main stated contribution — \"remarkable increase in recall... the more important metric\" — collapses with it.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper proposes LATCN, a multi-label text classification model for automatic ICD-9 coding that stacks a multi-layer temporal convolutional network (with exponentially dilated residual blocks) under the label-wise attention mechanism of Mullenbach et al. (2018). The motivation is that dilated convolutions give a large receptive field suited to long discharge summaries, while per-label attention preserves label-specific document views. On the MIMIC-III top-50 label benchmark (discharge summaries truncated to 2,500 tokens), the authors report micro-F1 0.681 and micro-recall 0.710 versus 0.633 and 0.584 for the previous state of the art DR-CAML (Table 2), and argue the recall gain (~28% relative) is the clinically important improvement. The manuscript carries an explicit note that the work was done in 2019 and is posted as a record of that work rather than a current account of the field.","tokens_in":7605,"tokens_out":2465,"duration_ms":38492,"significance":"If the recall gain survives matched-threshold evaluation, the result is a meaningful contribution to the 2019-era medical coding literature: LATCN is a simple, parameter-modest architecture (4 residual blocks, 100 channels, k=4) whose exponentially dilated receptive field is well motivated for long discharge summaries (median 1,388 words), and the label-wise attention preserves the interpretability property of CAML. The paper's explicit framing as an archival record of 2019 work is honest and appropriately scopes its novelty claims. However, the work is evaluated only on the 50 most frequent labels with single-run results, so its evidentiary weight for current practice is limited, and it neither ships code nor provides the threshold analysis needed to certify the headline numbers.","major_comments":[{"comment":"Table 2 / Eq. (4): the model outputs per-label sigmoid probabilities, but the binarization threshold used to compute the reported precision, recall, and F1 is never stated anywhere in the manuscript. This is load-bearing for the central claim: relative to DR-CAML, LATCN gains 12.6 micro-recall points (0.710 vs 0.584) while losing 3.6 micro-precision points (0.655 vs 0.691). That joint movement is exactly the signature of a less conservative operating point rather than necessarily a better ranker, and the paper's own explanation ('generalization', §4) does not rule the threshold confound out. Notably, the threshold-insensitive comparison is much narrower: micro-AUC 0.931 vs 0.916. The authors should state the threshold, and ideally report PR curves or max-F1 over thresholds for LATCN and DR-CAML under matched operating points, so the recall claim can be separated from calibration.","section":"§4, Table 2"},{"comment":"§3.3 / §4: all results come from a single 200-epoch training run with hyperparameters grid-searched on the development set. No seed variance, error bars, or significance tests are reported, yet the abstract and §4 describe the gains as 'significantly better' and 'remarkable'. On this benchmark, seed-to-seed variation of 1–2 micro-F1 points is routine, which would materially erode the 4.8-point micro-F1 gap over DR-CAML. At minimum the word 'significantly' should be removed or supported; ideally the model should be retrained over several seeds with mean ± std reported.","section":"§3.3, §4"},{"comment":"Table 2: the CAML, DR-CAML, LEAM, CNN, Bi-GRU, and LR rows are attributed to prior papers, which strongly suggests the numbers were transcribed from Mullenbach et al. (2018) and related work rather than re-run under identical preprocessing. The manuscript does not state whether the train/dev/test splits, tokenization, and truncation (2,500 tokens, §3.1) match those used to produce the baseline numbers. Any mismatch silently inflates or deflates the comparison. The provenance of each baseline row, and the exact split used, must be stated explicitly.","section":"Table 2"},{"comment":"§3.1 and title/abstract: evaluation is restricted to the 50 most frequent ICD-9 labels, with the full-label-set evaluation explicitly deferred to future work (§6). The ~50-code setting covers only the head of the distribution, while real coding workloads are dominated by long-tail codes, and it is known from subsequent literature that relative model rankings on MIMIC-III top-50 do not always transfer to the full label set. The title ('Improve Medical Coding') and the clinical decision-support framing in §3.2 and §4 are broader than the evidence. The claims should be scoped to the top-50 setting, or full-label results added.","section":"§3.1, title/abstract"}],"minor_comments":[{"comment":"Eq. (2): the softmax is written self-referentially (α_ℓ = exp(α_ℓ) / Σ_n exp(α_ℓn)) and contains a typesetting artifact '(sof tmax)'. Presumably the intended form is α_ℓn = exp(a_ℓn) / Σ_m exp(a_ℓm) where a_ℓn is the (n,ℓ) entry of A from Eq. (1). Please fix the notation.","section":"§2, Eq. (2)"},{"comment":"Table 2: P@5/R@5 and several precision/recall cells are left blank (dashes) for LR, CNN, Bi-GRU, and LEAM. Either fill these from the cited sources or note why they are unavailable; the partial rows make cross-model comparison harder to read.","section":"Table 2"},{"comment":"§2: the receptive field formula 2×(2^{ℓ+1}−1)×k appears to assume two convolutions per residual block and stride 1; please make the indexing convention explicit (whether ℓ starts at 0) so the formula can be checked against the 4-block architecture in Figure 1.","section":"§2"},{"comment":"§3.1/§3.3: train/development/test split sizes and the source of the split (e.g., the standard MIMIC-III top-50 split of Mullenbach et al.) are not given; nor are the word2vec embedding dimension d_e or whether embeddings are fine-tuned. These are needed for reproduction.","section":"§3.1, §3.3"},{"comment":"No code release or reproducibility artifact is mentioned. Given the small model and standard dataset, releasing training code and the selected configuration would substantially strengthen the paper's value as a record of the work.","section":"§3.3"},{"comment":"Minor: 'state-or-the-art' (§1) should be 'state-of-the-art'; the Schwenk et al. reference is missing a year in the citation; Figure 1 labels channel width and dilation but the figure could also indicate the 2,500-token input length for self-containment.","section":"§1, References, Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The authors' own prefatory note frames this as a 2019 manuscript posted \"as a record of the work rather than as a current account of the field.\" The editor should decide whether the venue's bar for an archival 2019 result is the same as for a current contribution; judged against 2019 SOTA the result was competitive, but the evaluation practices (single run, no threshold reporting, transcribed baselines) were common then and are below current expectations. The comparison rows in Table 2 appear to be transcribed from Mullenbach et al. (2018) without re-implementation; if the venue has a reproducibility policy, that should be explicitly flagged."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a competent recombination of Bai’s TCN with Mullenbach-style label-wise attention, and Table 2 does show a clear lift over DR-CAML on MIMIC-III top-50—especially micro-recall. It is not a new framework.\n\nWhat is actually new is the empirical pairing. Multi-layer dilated TCN is a sensible way to get long-range context on discharge notes (median ~1.4k tokens, truncated at 2.5k), and feeding that into per-label attention is a natural next step after CAML. Architecture, residual blocks, and the small hyperparameter grid are described clearly enough to reimplement. The clinical argument that recall matters more than precision for a coder-assist tool is fair and stated without hype. Macro/micro AUC also move in the right direction (0.908/0.931 vs 0.884/0.916), so this is not pure threshold gaming.\n\nSoft spots, in proportion. Evaluation is top-50 only; full-code-set is left as future work, which is the real clinical workload. No error bars, no seeds, one 200-epoch run—routine 1–2 point F1 noise would eat a chunk of the 4.8-point micro-F1 gap. The paper never states the binarization threshold for F1/P/R; the simultaneous recall jump and precision drop (~7%) is exactly what you get from a less conservative cut, and they never show PR curves or max-F1-over-threshold. AUC gap is much smaller than the recall claim, so the “28% recall” headline oversells ranking quality. DR-CAML numbers look transcribed rather than re-run under identical preprocessing. No code or data release. None of that sinks the directional result; it just means the magnitude is softer than the abstract suggests.\n\nWho it is for: people building or benchmarking automated ICD coding on MIMIC-style notes who want a strong 2019-era convolutional baseline with attention. Not for anyone needing full-label or production claims.\n\nI would send it to referees. Expect them to demand threshold analysis, variance, and at least a full-label or long-tail experiment. Worth a careful read if you work this problem; not a must-cite otherwise.","headline":"Clean 2019 engineering combo of TCN + label attention; real-looking top-50 gains, but the big recall number is under-supported and partly an operating-point story.","tokens_in":8646,"tokens_out":585,"would_cite":false,"duration_ms":20045,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A multi-layer temporal convolution network with per-code attention raises medical coding F1 by 9% and recall by 28% on common ICD labels.","keywords":["medical coding","ICD classification","temporal convolutional networks","label-wise attention","multi-label text classification","MIMIC-III","clinical NLP","discharge summaries"],"falsifier":"Train and evaluate the identical LATCN architecture on the full MIMIC-III ICD label set under the same splits; if the F1 and recall lifts over the prior best model vanish or reverse on the long-tail codes, the central performance claim does not hold for real coding workloads.","tokens_in":8386,"feed_emoji":"🏥","tokens_out":872,"duration_ms":46496,"temperature":0.7,"pith_summary":"Medical coding asks a system to assign many diagnosis and procedure codes from a single long hospital note, where evidence for different codes can sit far apart. The authors treat this as multi-label text classification and show that a stack of temporal convolutional layers, which expand their view of the document exponentially, followed by a separate attention pass for each code, captures both global context and code-specific focus better than earlier local convolutional models. On the fifty most frequent ICD codes in MIMIC-III discharge summaries, their LATCN model improves average F1 by about nine percent and recall by about twenty-eight percent over the prior best system. They argue that the recall gain matters more in clinical decision support, because missing a true code is costlier than offering an extra candidate for a human to filter. The result positions long-range dilated convolution plus label-wise attention as a practical route to more useful automated coding aids.","feed_headline":"TCN plus label attention lifts medical-coding recall 28%","feed_subtitle":"Global context over long discharge notes beats prior models on common ICD codes for decision support.","key_machinery":"LATCN: residual temporal convolutional blocks with exponentially growing dilations that build a long-range document matrix, followed by a learned per-label attention matrix that produces one dense vector per ICD code for a final sigmoid prediction.","core_discovery":"On the MIMIC-III top-50 ICD label set, a multi-layer temporal convolutional network followed by label-wise attention (LATCN) significantly outperforms the previous state-of-the-art attentional CNN model, raising average F1 by roughly 9% and micro-recall from 0.584 to 0.710. The authors treat the large recall lift as the clinically more important result for decision-support use.","pith_inferences":["The same TCN-plus-per-label-attention pattern may transfer to other long-document multi-label tasks outside medicine, such as multi-topic legal or scientific tagging.","Because precision falls while recall rises, the practical deployment value is as a high-recall first-pass filter for humans rather than a fully automatic coder.","Adding contextual embeddings or hierarchy-aware losses, both noted as future work, is a concrete next experiment for closing the remaining gap on rare codes."],"forward_implications":["High-recall automated coding can surface more true diagnoses and procedures for human review, lowering the chance a code is missed entirely.","Recall-first design becomes a defensible default for clinical NLP decision-support tools.","Multi-layer dilated convolutions are a viable encoder for long clinical notes in place of local CNNs or recurrent nets.","Label-specific attention remains useful even when the encoder already sees global context.","Measuring the same architecture on the full ICD label set and with hierarchical code structure is the direct next test the paper itself flags."],"fun_headline_variants":["Label-wise TCN attention lifts medical-coding recall 28%","LATCN raises ICD coding F1 9% and recall to 0.710","Multi-layer TCN plus label attention beats prior coding models","Temporal conv nets with per-label focus improve MIMIC-III coding","Deep label-attentive TCN gains 28% recall on clinical codes"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The gains are measured only on the fifty most common codes, with notes truncated at 2,500 tokens and only discharge summaries used; full-label-set results are left as future work.","fun_headline_variants_meta":{"raw":{"variants":["Label-wise TCN attention lifts medical-coding recall 28%","LATCN raises ICD coding F1 9% and recall to 0.710","Multi-layer TCN plus label attention beats prior coding models","Temporal conv nets with per-label focus improve MIMIC-III coding","Deep label-attentive TCN gains 28% recall on clinical codes"]},"model":"grok-4.5","effort":"low","cost_usd":0.004098,"raw_usage":{"total_tokens":1208,"prompt_tokens":733,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":40984000,"prompt_tokens_details":{"text_tokens":733,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":391,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":733,"tokens_out":84,"duration_ms":8187,"temperature":1.0,"reasoning_tokens":391,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T00:32:52.549898+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train and evaluate the identical LATCN architecture on the full MIMIC-III ICD label set under the same splits; if the F1 and recall lifts over the prior best model vanish or reverse on the long-tail codes, the central performance claim does not hold for real coding workloads.","supporting_citations":[],"review_version":1}