acceptodds
Under review as a conference paper at ICLR 2027

Diagnose Before Adapting: Intervention Headroom under Clinical Drift

Abstract

Retrieval systems can adapt their evidence memory, decision rule, or model parameters, yet these interventions are often evaluated separately. We study them jointly through Diagnose-Before-Adapt (DBA), a prequential intervention audit that measures conditional gains under delayed label release. Recent versus frozen memory measures staleness, cumulative memory measures retention, a support-constrained oracle bounds candidate-restricted decisions, and learned aggregation measures realized gain. Together, these probes characterize the interactions among adaptation components. We apply DBA to 117,328 consultations over 51 months and introduce Learned Neighbourhood Aggregation (LNA), a lightweight ranker over prior evidence with frozen text encoders. Fresh-memory headroom grows by more than five balanced-accuracy points per year, while retaining labelled history recovers 20.33–34.98 points. Against recalibrated cumulative memory, base LNA gains 2.83, 1.05, and 3.53 points on the three streams, with over lower reported GPU occupancy than monthly retraining. Gains are positive in 44 of 45 correlated seed–origin conditions and under label-blind overlap filtering. Public TCM-SD further replicates the aggregation design on a static task. Together, these results identify memory retention and learned evidence use as effective, resource-efficient adaptation components in the studied longitudinal streams.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Large Language Models, Retrieval-Augmented Generation, Retrieval-Based Adaptation, Continual Learning, Temporal Distribution Shift

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_cxuiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Diagnose-Before-Adapt (DBA), a retrospective framework for comparing memory updates, calibration, learned aggregation, and model adaptation under delayed label availability. Its proposed Learned Neighbourhood Aggregation (LNA) uses a lightweight gradient-boosted scorer over retrieval statistics from frozen encoders. Experiments use 117,328 consultations spanning 51 months from one clinical service, with additional physician-specific analyses and a static public benchmark. Cumulative memory provides the largest improvement, while LNA adds 1.05–3.53 balanced-accuracy points over recalibrated cumulative memory with lower reported GPU occupancy than parameter adaptation.

Critical strengths
  1. The temporal protocol is carefully specified. Predictions use only prior-month state, evaluation cases remain excluded from subsequent training and memory, and interventions are compared on identical monthly cases. The size-matched freshness control also separates recency from the amount of stored evidence.
  2. The paper tests several plausible explanations for the gains. Dense-retrieval and scorer-refitting ablations, alternative supervised aggregators, prior controls, and label-blind overlap filtering provide more convincing evidence than a comparison against frozen memory alone. Reporting CPU costs separately from GPU occupancy is also appropriate.
  3. The limitations are unusually explicit. The authors acknowledge overlapping streams, the static nature of the public benchmark, uncertainty across physicians, and the failure of the exploratory router to generalize to one stream. They also correctly state that the support oracle does not bound LNA, which can predict classes absent from the retrieved candidates.
Critical weaknesses
  1. The methodological contribution remains limited. Conditional headroom is a matched performance difference, the support ceiling follows directly from candidate inclusion, and the interaction proposition is a standard difference-in-differences identity. LNA combines established retrieval features with supervised boosting. This could still support a valuable empirical paper, but DBA’s additional diagnostic value is not yet demonstrated convincingly: it requires executing the interventions, and the exploratory routing rules reduce to constant actions within each evaluated stream. A direct comparison with equally lightweight classifiers trained on frozen query embeddings would also help establish the benefit of neighbourhood-based learning.
  2. The primary inference does not adequately address the temporal structure of the experiment. Base-LNA superiority uses month resampling, although successive months share accumulated memory and fitted models. Bonferroni correction addresses multiplicity, not serial dependence. The appendix’s dependence-aware superiority analysis concerns LNA+, while the base-LNA block analysis tests non-inferiority against different references. This matters because LNA+’s SingleMD comparison changes from significant under IID resampling to nonsignificant with longer blocks. The confirmatory base-LNA versus BL2b comparison needs the same sensitivity analysis.
  3. The breadth of evidence is narrower than the number of reported streams suggests. Full contains both principal physician streams, and all longitudinal results come from one service. TCM-SD validates static aggregation rather than adaptation under drift. Overlap filtering preserves relative gains, but absolute LNA performance falls substantially—for example, from 78.85 to 57.72 on Full—and filtering is applied at retrieval time without changing model fitting. These are useful controls, but they do not establish patient-disjoint generalization. The physician-level Hartung–Knapp interval also crosses zero, leaving uncertainty about how consistently the benefit extends beyond the dominant physicians.
Questions for authors
  1. Can you report moving-block or HAC confidence intervals for the primary base-LNA versus BL2b contrasts, including sensitivity to block length and the stated multiplicity correction?
  2. How does LNA compare with a linear classifier or small MLP trained directly on frozen query embeddings using the same historical labels, monthly update schedule, and tuning budget? The existing alternative-aggregator comparison keeps the neighbourhood representation fixed.
  3. Which record fields are available to each predictor at inference time? Please explicitly distinguish chief complaint and examination findings from prescriptions, syndrome names, and other information recorded after the clinical decision. Also clarify whether repeat patients can occur across training and evaluation consultations.
  4. Figure 17 shows approximately chance-level performance with 1,000–2,000 fitting neighbourhoods, followed by a sharp jump at 3,000 across all streams. What explains this discontinuity? Please clarify whether any fitting condition, fallback, or implementation threshold changes at that point.
10 comments
AuthorsOpenReview ↗

W2 & Q1: Dependence-Aware Inference (Table 4a & 28) & SingleMD

We are indebted to your rigorous statistical and code-level audit, which directly strengthened our uncertainty quantification and uncovered a real implementation bug in our exploratory budget script. Keeping manuscript edits targeted (preserving all Figures 1–20 and Tables 1–3, adding two-panel Table 4 in Section 4.3, and placing full audit tables in Appendix K), we respond in four focused comments: Part 1 covers autocorrelation-robust inference (W2, Q1), Part 2 covers frozen-query classifiers and DBA's scope (W1, Q2), Part 3 covers the pre-decision field contract, overlap filtering, and four-clinician pooling (W3, Q3), and Part 4 covers the Figure 17 code audit (Q4).

1. Dependence-Aware Inference for Confirmatory Base LNA vs. BL2b (W2, Q1)

We fully agree that IID month resampling can understate variance when monthly evaluation series are autocorrelated, and that confirmatory Base LNA requires the same dependence-aware scrutiny previously reported only for LNA+ (Table 20). Using the exact monthly paired series, we added Table 4a directly in Section 4.3 (Page 10) alongside the complete 6-estimator Appendix K.4 (Table 28) reporting circular moving-block bootstrap (L = 3, 6, 12; Künsch, 1989), finite-sample Student-tt Newey–West HAC (lag = 12; Newey & West, 1987), and non-overlapping quarterly block 95% CIs plus raw and 63-test Bonferroni-adjusted pp-values:

Stream / Phase (Table 4a & Table 28)Δ\DeltaIID 95% CI (p63p_{63})Block-3 CIBlock-6 CIBlock-12 CINW-12 CI (prawp_{\text{raw}} / p63p_{63})Quarterly CI
Full (T=39T=39)+2.83[+2.06,+3.61] (0.0063)[+2.04,+3.63][+2.03,+3.58][+2.15,+3.45][+2.05,+3.61] (<0.0001<0.0001 / <0.0001<0.0001)[+1.86,+3.80]
IndepMD (T=38T=38)+3.53[+2.61,+4.51] (0.0063)[+2.48,+4.76][+2.47,+4.86][+2.55,+4.64][+2.34,+4.71] (<0.0001<0.0001 / <0.0001<0.0001)[+2.13,+4.93]
SingleMD policy (T=39T=39)+1.05[+0.07,+2.09] (1.0000)[-0.05,+2.18][-0.21,+2.35][-0.24,+2.48][-0.44,+2.53] (0.1621 / 1.0000)[-0.22,+2.32]
SingleMD pre-8k (T=19T=19)+0.08[-1.00,+1.10] (1.0000)[-1.08,+1.12][-1.16,+1.07][-0.82,+1.00][-0.73,+0.90] (0.8339 / 1.0000)[-1.80,+2.22]
SingleMD active (T=20T=20)+1.96[+0.34,+3.55] (1.0000)[+0.31,+3.59][+0.17,+3.62][+0.56,+3.37][+0.23,+3.70] (0.0285 / 1.0000)[-0.57,+4.34]

2. What These Estimators Establish (Full & IndepMD vs. SingleMD Phases)

  1. Full and IndependentMD are robust to serial dependence and 63-test multiplicity: Across every moving-block length (3, 6, 12), quarterly aggregation, and NW-12 HAC, unadjusted 95% CIs remain strictly above +1.86 pts on Full and +2.13 pts on IndepMD. Under NW-12, both streams survive 63-test Bonferroni correction at p63<10−4p_{63}<10^{-4} (IID bootstrap p63=0.0063p_{63}=0.0063 reflects the 63/10,00063/10{,}000 resampling resolution floor).
  2. SingleMD 39-month policy vs. pre-/post-8k phase decomposition: On SingleMD, the 39-month deployable policy (+1.05 pts) is marginally positive under unadjusted IID resampling ([+0.07,+2.09], praw=0.0432p_{\text{raw}}=0.0432) but crosses zero under all block and HAC estimators (NW-12 [-0.44,+2.53], praw=0.1621p_{\text{raw}}=0.1621, p63=1.0000p_{63}=1.0000). Decomposing fit_log descriptively—while noting explicitly that 8k ranker activation co-varies with calendar time and historical archive size rather than being randomized—the 19 pre-8k fallback months average +0.08 pts (NW-12 [-0.73,+0.90]), whereas the 20 post-8k active months average +1.96 pts (unadjusted NW-12 [+0.23,+3.70], praw=0.0285p_{\text{raw}}=0.0285, but p63=1.0000p_{63}=1.0000 and quarterly [-0.57,+4.34]). Thus, active-phase single-physician gains are directionally positive on unadjusted monthly HAC, but remain suggestive rather than multiplicity-robust (Section 4.3 & Appendix K.4).
AuthorsOpenReview ↗

W1 & Q2: Direct Frozen-Query Classifiers, Aggregators, & DBA Scope

Your point in W1 and Q2—asking how matched parametric classifiers trained directly on frozen query embeddings perform, alongside bounding DBA's audit versus adaptive routing—identifies a key experimental control that we want to address with complete candor.

1. Candid Status of Direct Classifiers on Raw Frozen Query Embeddings (W1, Q2)

You asked how monthly Logistic or MLP heads trained directly on frozen query features (sparse IDF + 512-D bge-small-zh query embeddings, without neighbour retrieval) perform under the exact same <t monthly budget (3,000 or 12,000 cases).

We want to be completely transparent: we agree this is the cleanest direct parametric control for isolating retrieval-mediated classification from direct embedding classification, and indirect comparisons cannot substitute for it. Because raw consultation text and per-visit query embeddings (all_records.jsonl) reside exclusively on the restricted hospital server and are not exported in our local artefact bundle, direct frozen-query Logistic/MLP runs on the hospital streams remain unexecuted (we state this gap explicitly in Section 4.3 and Appendix E, and include a <t prequential runner script for host execution).

To situate what our completed experiments do and do not show in the meantime:

  • Holding <t retrieved neighbourhoods fixed (Table 4b, Section 4.3): On identical dual-retriever relational matrices (B=3,000B=3,000), LNA–HGB (77.74% / 65.19% / 76.83%) consistently beats a linear Logistic ranker (76.11% / 62.74% / 74.12%), a Two-layer MLP (76.48% / 63.61% / 74.49%), and permutation-invariant DeepSets (69.88% / 59.27% / 69.31%), isolating the benefit of gradient-boosted threshold splits given the retrieved neighbourhood representation.
  • Direct end-to-end encoders & parameter-updated LLMs: On public 148-way TCM-SD (Table 12), Base LNA (82.72% Acc, 56.25% Macro-F1) and LNA+ (83.19%, 58.50%) outperform end-to-end fine-tuned RoBERTa (82.26%, 47.55%) and domain-pretrained ZY-BERT (82.19%, 51.01%). On the longitudinal hospital streams (Tables 1, 2, 15), Base LNA stays within -1.27 to +1.49 pts of parameter-updated Qwen3.5-4B (CDA-L3 LoRA and BL5 monthly full-history retraining, passing 2-pt Bonferroni non-inferiority in 3/6 pairs in Table 15) at 0.0020–0.0111 vs. 2.28–2.64 GPU-hours.

2. Bounding DBA as a Counterfactual Audit Rather Than a Dynamic Controller (W1)

We also agree with both of your points in W1 regarding DBA's diagnostic role: DBA requires executing the compared arms offline, and in Table 19 and Appendix K.3 (Table 27) our exploratory LOSO threshold router selects a constant action (12/12 quarterly probes) on each stream (LNA+ on Full and IndepMD, LoRA on SingleMD), incurring 1.83 pts of static regret on IndepMD and not executing a late-horizon switch on SingleMD (Q8–Q12). In revised Section 4.2 and Appendix G, we therefore bound DBA strictly to offline counterfactual prequential attribution—measuring size-matched freshness gaps (Figure 3), retention (Table 2), multi-candidate coverage (Table 26), and memory–aggregation substitution (Proposition 2, Table 17) on a shared timeline—rather than claiming a zero-execution predictor or adaptive online controller.

AuthorsOpenReview ↗

W3 & Q3: Field Contract, Text-Overlap Controls, & 4-MD Pooling

Turning to W3 and Q3, we found your questions on exact inference-time schema fields, patient-level repeat visits under lexical filtering, and cross-physician uncertainty across the four eligible clinicians especially helpful for tightening our external-validity claims.

1. Exact Pre-Decision Field Contract Across Code Modules (W3, Q3)

To remove any ambiguity about whether post-diagnosis information leaks into encoders or prompts, new Appendix K.1 (Table 25) (referenced in Section 3.1) maps every schema field to its exact pipeline stage and function:

Record Field (Appendix Table 25)Clinical StageOnline Query, Retrieval, & Prompt (S_{t-1})Post-Lock Role (After y_t Locks)
chief_complaintPre-decision intakeSparse IDF (record_features) + dense BGE (presentation_of) + LLM promptStored in <t+1 keys & exemplars
tongue_pulsePre-decision examSparse IDF (record_features) + dense BGE (presentation_of) + LLM promptStored in <t+1 keys & exemplars
diagnosis_westernPre-decision (optional)Sparse IDF (record_features) + LLM prompt (not in dense BGE)Stored in <t+1 lexical index
symptoms / age / sexPre-decision (optional)Formatted in LLM prompt (cda/fuse.py) if present (not in LNA)Retained in consultation record
prescription (herbs)Post-decision orderStrictly withheld (never in encoders, features, or prompts)Model-free herb-drift audit (Fig. 2b)
syndrome / zangfuPost-decision noteStrictly withheld (never in encoders, features, or prompts)Novel-term drift audit (Fig. 2d)
gold_meridian (yty_t)Post-decision targetStrictly withheld until month-tt predictions lock (score_month)Scores month tt; non-eval <t+1 updates StS_t

2. Repeat Visits, Text-Overlap Drop (78.85% -> 57.72%), & Patient-ID Boundary (W3)

As noted in revised Sections 3.1 and 4.3, de-identified records are keyed by consultation ID without a cross-visit patient_id, so exact repeat-patient rates cannot be audited. In Table 3, withholding all strictly earlier normalized-exact or high-similarity character-trigram matches affects 46.08%–61.21% of queries and substantially lowers absolute accuracy as close clinical templates are removed (e.g., Full drops from 78.85% to 57.92% / 57.72%, SingleMD to 43.57% / 43.29%, IndepMD to 56.14% / 55.92%), while paired LNA gains over BL2b widen to +4.09 / +4.36, +2.16 / +2.11, and +4.84 / +4.78 pts. Crucially, we state explicitly in Section 4.3 that retrieval-time lexical filtering bounds template-copying sensitivity during inference, but does not remove repeat patients from <t scorer fitting and cannot establish patient-disjoint generalization.

3. Four-Clinician Heterogeneity & Zero-Crossing Hartung–Knapp Intervals (W3)

Across the K=4K=4 mutually exclusive clinicians meeting outcome-blind volume thresholds (summarized directly in Section 4.3 and Appendix I, Tables 21–22), we report all uncertainty estimators side-by-side without selective emphasis:

Clinician (Table 21)CasesEval / Active MoPolicy Δ\Delta (95% CI)Active Δ\DeltaMeta-Estimator (Table 22, K=4K=4)Pooled (95% Interval)
8056e996 (IndepMD)79,53538 / 38+3.53 ([+2.57, +4.49])+3.53Normal random-effects CI+1.73 ([+0.15, +3.31])
f3ec4945 (SingleMD)26,09939 / 20+1.05 ([+0.02, +2.07])+1.96Hartung–Knapp (t3t_3) CI+1.73 ([-0.71, +4.17])
69d22f35 (pre-8k)6,06535 / 0+1.77 ([+0.14, +3.39])—Modified Hartung–Knapp CI+1.73 ([-0.84, +4.30])
a90628ca (pre-8k)2,09211 / 0-0.27 ([-2.93, +2.40])—Between-MD prediction interval+1.73 ([-5.16, +8.62])

Leave-one-physician-out point estimates stay positive (+1.11 to +2.14 pts), but only 2 of 4 clinicians reach the 8,000-case ranker activation threshold, and both Hartung–Knapp ([-0.71, +4.17]) and prediction ([-5.16, +8.62]) intervals cross zero. We highlight this explicitly in Section 4.3 (Page 9) so cross-clinician generalizability is not overstated.

Reviewer_Vs5eOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Diagnose-before-adapt (DBA), a prequential evaluation framework for quantifying the effects of memory freshness, memory retention, retrieval support, and learned aggregation under clinical temporal drift. The authors introduce Learned neighborhood aggregation (LNA), which uses a lightweight supervised ranker over retrieval-based features while keeping text encoders frozen. Experiments on 117,328 clinical consultations spanning 51 months show that cumulative memory improves balanced accuracy by 20.33–34.98 points over frozen memory, while LNA provides additional gains of 1.05–3.53 points over recalibrated cumulative memory with substantially lower GPU occupancy than parameter adaptation.

Critical strengths
  • The paper examines multiple adaptation components rather than focusing exclusively on model parameter updates.
  • The prequential evaluation protocol, matched intervention comparisons, and explicit temporal constraints strengthen the reliability of the results.
  • The paper includes component ablations, sensitivity analyses, overlap filtering, and resource-efficiency comparisons.
Critical weaknesses
  • DBA primarily organizes existing intervention comparisons into a unified evaluation framework, while LNA builds on established supervised ranking techniques. The paper could more clearly articulate its methodological contributions beyond this integration.
  • The framework measures intervention headroom retrospectively, but its ability to guide prospective adaptation decisions remains insufficiently established. The exploratory routing results in Appendix G provide limited evidence of improvement over fixed adaptation policies.
  • The longitudinal experiments originate from a single clinical service, and the three evaluation streams are not fully independent. Although TCM-SD provides additional validation of LNA, it does not evaluate temporal adaptation.
  • The support-constrained oracle suggests 21.17–26.14 points of potential improvement over fixed aggregation, whereas LNA achieves much smaller realized gains. The practical interpretation of this diagnostic gap could be further clarified.
  • The discussion of calibration drift and related work relies heavily on outdated studies, with limited coverage of recent advances. The authors should discuss whether these limitations persist in more recent research and better position their contributions within the current literature.

[Minor]

  • In Definition 1, the notation should be revised to make the dependence of Hs(I∣z)H_s(I\mid z) on zz and I(z)I(z) explicit.
  • The notation BstB_{st} and BsztB_{szt} should also be made consistent.
  • The orange markers mentioned in the caption of Figure 3 appear to be missing. The authors should carefully review the manuscript for similar inconsistencies.
Questions for authors

See above

3 comments
AuthorsOpenReview ↗

W1 & Minor: Counterfactual Decomposition, Substitution, & Notation

Your review raised sharp conceptual questions about what DBA and LNA contribute beyond system integration, alongside generous remarks on our strict prequential no-leakage protocol, multi-component perspective, ablations, and text-overlap controls—feedback we greatly value. Keeping manuscript edits targeted (preserving all Figures 1–20 and Tables 1–3, adding Table 4 in Section 4.3, and placing audit tables in Appendix K), we split our reply into three comments: Part 1 addresses methodological contribution and notation (W1, Minor), Part 2 addresses the oracle gap (W4), and Part 3 addresses prospective/routing scope, dataset boundaries, and 2025–2026 literature (W2, W3, W5).

1. Methodological Contribution Beyond System Integration (W1)

While LNA deliberately uses simple, deployable primitives (frozen encoders + CPU gradient-boosted ranking), our core methodological contribution is counterfactual identification under delayed-label clinical drift rather than ad-hoc pipeline assembly:

  • Deconfounding Temporal Staleness from Archive Growth (Figure 3, Table 6): In longitudinal clinical streams where visit volume grows over time, comparing a recent window against a frozen warm-up index confounds recency with sample size. DBA's cardinality-matched rolling-window probe subsamples each 12-month window to the exact warm-up size, yielding size-matched freshness-gap slopes of +5.60 pts/yr on Full (manifesting as absolute decay from 47.29% to 36.74%) and +5.38 pts/yr on SingleMD (manifesting as forgone improvement as fresh rises +5.08 pts/yr while frozen stagnates at -0.30 pts/yr).
  • Formalizing Path-Dependent Component Substitution (Proposition 2, Table 17): Single-axis comparisons do not reveal whether an intervention's gain changes when other components are updated. Proposition 2 proves that along the two adaptation paths (M0,A0)→(M1,A0)→(M1,A1)(M_0,A_0)\to(M_1,A_0)\to(M_1,A_1) and (M0,A0)→(M0,A1)→(M1,A1)(M_0,A_0)\to(M_0,A_1)\to(M_1,A_1), each lever's marginal headroom shifts by the difference-in-differences interaction:

I=V11−V10−V01+V00=(−9.86,  −8.44,  −7.81) ptsI = V_{11} - V_{10} - V_{01} + V_{00} = (-9.86,\; -8.44,\; -7.81)\text{ pts}

Across Full, SingleMD, and IndepMD (Table 17), LNA+ gains +13.49, +10.86, and +12.48 pts over hand voting when memory is frozen (V01−V00V_{01}-V_{00}), but +3.64, +2.42, and +4.67 pts once cumulative memory is retained (V11−V10V_{11}-V_{10}). Thus, post-retrieval aggregation headroom is 7.81–9.86 pts smaller under cumulative than frozen memory, showing formally and empirically that aggregation gains depend on memory state rather than transferring directly across memory regimes.

  • Strict <t Leave-One-Out Relational Formulation (Section 3.4, Table 4b): Rather than pooling embeddings directly (DeepSets) or reweighting priors (Saerens et al., 2002), LNA maps dual-retriever support, rank, similarity, cross-space agreement, and historical prior π<t(c)\pi_{<t}(c) into six 14-D candidate rows fitted on leave-one-out historical neighbourhoods N<t(−x)(x)N^{(-x)}_{<t}(x) (retaining π<t(c)\pi_{<t}(c) when candidate support is zero so all six classes remain scoreable). On identical <t matrices (Table 4b), LNA–HGB outperforms Logistic (+1.63..+2.71 pts), Two-layer MLP (+1.26..+2.34 pts), and DeepSets (+5.92..+7.86 pts).

2. Fixes for Definition 1 Notation & Figure 3 Caption (Minor Comments)

We also appreciate your keen eye on both notation and caption details:

  1. Definition 1 & Proposition 1 (Sections 3.2–3.3, Page 4): We replaced BsztB_{szt} and BstB_{st} uniformly with Bs,t(z)B_{s,t}(z) for the balanced accuracy of deployed state zz on stream ss in month t∈Tst\in T_s, and wrote Hs(I∣z)=Vs(I(z))−Vs(z)=∣Ts∣−1∑t∈Ts(Bs,t(I(z))−Bs,t(z))H_s(I\mid z)=V_s(I(z))-V_s(z)=|T_s|^{-1}\sum_{t\in T_s}(B_{s,t}(I(z))-B_{s,t}(z)) so its dependence on both zz and I(z)I(z) is explicit.
  2. Figure 3 Caption (Section 4.1, Page 6): Removed the stray orange-marker clause and explicitly described the solid blue recent-memory curves, dashed grey frozen-memory curves, shaded paired-gap ribbon, and panel-header summaries.
AuthorsOpenReview ↗

W4: Multi-Candidate Recall o_t vs. Prospective Single-Label Models

We are grateful for your question in W4 asking why the support-constrained oracle oto_t reaches 95.96% on Full while realized single-label accuracy (Base LNA, LoRA, and monthly retraining) sits around 77%–79%—it prompted an essential clarification separating multi-candidate set inclusion from single-label prediction.

1. Why oto_t (95.96%) and Single-Label Learners (~77%–79%) Measure Distinct Objects

We agree that our original phrasing could be misread as treating oto_t as an achievable single-label accuracy ceiling. In revised Section 4.2 (Page 8) and new Appendix K.2 (Table 26), we explicitly separate both the mathematical object and the evaluation horizon:

Reference / Learner (Appendix Table 26)FullSingleMDIndepMDGPU-hHorizon & Object Measured
Fixed voting rule (on oracle probes)74.7953.5871.53013 third-month probes: single-label top-1 prediction
Candidate-inclusion bound oto_t95.9679.7393.67—13 third-month probes: multi-candidate set recall (4–5 of 6 classes/query)
BL2b Recalibrated cumulative memory76.0356.2072.860Full 39/39/38 mo: non-parametric single-label baseline
CDA-L3 Drift-triggered 4B-LLM LoRA78.6857.4077.662.28–2.54Full 39/39/38 mo: parametric 4B-LLM adapter
BL5 Monthly 4B-LLM full-history retrain77.3758.4576.012.56–2.64Full 39/39/38 mo: full-history 4B-LLM retraining
Base LNA (ours, CPU HGB)78.8557.2576.390.0020–0.0111Full 39/39/38 mo: -1.27..+1.49 pts vs. 4B-LLM (3/6 NI@2pt, Table 15)

Specifically, under dual retrieval (k=32k=32 sparse + k=32k=32 dense), the retrieved pool Rt(x)R_t(x) spans 4 to 5 distinct candidate meridians out of 6 on a typical consultation. By Proposition 1, ot(x,y)=yo_t(x,y)=y if y∈Rt(x)y\in R_t(x) (and ⊥\bot otherwise), crediting a hit (I[y∈Rt(x)]\mathbb{I}[y\in R_t(x)]) whenever the gold meridian appears anywhere in that 4–5 class candidate union—equivalent to top-4/top-5 set recall rather than a single-label Bayes optimal classifier max⁡cPr⁡(Y=c∣xintake)\max_c \Pr(Y=c\mid x_{\text{intake}}).

2. What Our Evidence Supports—and What Remains Open

Separating these two quantities sharpens both our positive claim and our boundary:

  • Role of high oto_t in DBA: High candidate inclusion (93.67%–95.96% on the two large streams) rules out complete retrieval misses as the primary failure mode on Full and IndepMD (unlike SingleMD rare Shaoyang at 18.99% coverage in Table 23), motivating post-retrieval discrimination rather than claiming that 96% top-1 accuracy is attainable.
  • Direct empirical basis for LNA's efficiency trade-off: LNA's practical value is established not by proximity to oto_t, but by direct head-to-head prequential comparison against BL2b (+2.83 / +1.05 / +3.53 pts) and parameter-updated Qwen3.5-4B baselines (CDA-L3 and BL5, differing by -1.27 to +1.49 pts across all six comparisons at >200x lower GPU occupancy; two-sided differences are non-significant in Table 2, while 2-pt Bonferroni non-inferiority holds in 3/6 pairs in Table 15).
  • Explicit non-claim on the residual gap: As now stated explicitly in Section 4.2 and Appendix K.2, our experiments do not establish an absolute single-label performance ceiling, and cannot yet decompose how much of the remaining gap between ~77%–79% and o_t stems from intrinsic pre-prescription intake ambiguity (overlapping chief complaints before herbal prescribing) versus frozen-encoder representation, model family, or training-procedure limits.
AuthorsOpenReview ↗

W2, W3, & W5: Prospective Scope, LOSO Routing, & 2025-26 Citations

To conclude our reply, we welcome the opportunity to clarify DBA's prospective vs. retrospective scope and LOSO routing behavior (W2), single-service dataset boundaries (W3), and recent 2025–2026 clinical-drift and continual-learning literature (W5).

1. Retrospective Audit vs. Prospective Scope & Constant-Action LOSO Routing (W2)

You raised an important point: if DBA primarily diagnoses past trajectories, how much prospective decision value does it offer? In revised Section 4.2 (Page 8), Appendix G (Table 19), and Appendix K.3 (Table 27), we draw a strict boundary between early diagnostic signals and dynamic routing:

  • Early-window diagnostic consistency (without claiming a locked trial): Already in post-warm-up Year 1 (Months 1–12), DBA's size-matched freshness probe exposes substantial staleness (+24.04 pts on Full, +5.45 pts on SingleMD, widening to +36.79 and +16.68 by Year 4; Table 6), and warm-up validation assigns 0.0 fusion weight to the frozen 4B-LLM. At the same time, we state candidly that we did not run a separate locked prospective arm-selection trial on the remaining 27 months.
  • Exploratory LOSO router behavior (constant action per stream): Using thresholds (τC,τH)(\tau_C,\tau_H) tuned on the other two stream objects (Appendix Table 27), our held-out quarterly router behaves as follows:
Held-Out Stream (Appendix Table 27)(τC,τH)(\tau_C,\tau_H)LOSO RouterAlways MemoryAlways LNA+Always LoRAHindsightStatic RegretQuarterly Action Trace (Table 19)
HOSP-Stream-Full(80%, 15)79.3975.4179.3977.9880.210.00Selects LNA+ in 12/12 quarters (9/12 hindsight)
HOSP-Stream-SingleMD(90%, 15)58.6357.4457.2458.6359.930.00Selects LoRA in 12/12 quarters (6/12 hindsight)
HOSP-Stream-IndependentMD(80%, 15)76.8172.3676.8178.6478.811.83Selects LNA+ in 12/12 quarters (1/12 hindsight)

Table 5 reports temporal trends on Full and SingleMD (∣ρt∣∈[0.61,0.95]|\rho_t|\in[0.61,0.95]). Across the evaluated trajectories, the current two-threshold LOSO router selects a constant action (12/12 quarters) on each stream: it matches the best static policy on Full and SingleMD, incurs 1.83 pts of static regret below always-LoRA on IndepMD, and does not dynamically switch from LoRA to non-LoRA arms in SingleMD's later quarters (Q8–Q12, 2024-04 to 2025-04) once history accumulates. We therefore frame DBA strictly as a prequential diagnostic audit rather than claiming a validated dynamic controller.

2. Single Outpatient Service Scope (W3) & 2025–2026 Literature (W5)

  • Dataset scope (Sections 3.1 & 4.3): We state explicitly that all three longitudinal streams originate from one multi-clinician TCM outpatient service without cross-visit patient IDs, and that Full nests SingleMD and IndepMD as multi-scale deployment regimes (though the two physician streams share zero cases). Within available data, we add four mutually exclusive clinician audits (Tables 21–22), external static REAL-Meridian-Verified (212 audited cases; Figure 7), and public 148-way TCM-SD (54,152 cases; Table 12, which replicates static aggregation rather than temporal drift), leaving multi-hospital Western EHR longitudinal validation to future work.
  • Integration of 2025–2026 drift & continual-learning literature (Section 2): Alongside Villena et al. (2026, Clinical-ShiftEval) and Zeng et al. (2026, MedCL-Bench) (already in Section 2), revised Sections 1 and 2 integrate Subasri et al. (2025, JAMA Network Open) on hospital-scale drift detection and scheduled retraining and Schirmer et al. (2025, TMLR) on temporal test-time adaptation via state-space models. Notably, even these recent studies evaluate parameter updates or single-lever shifts in isolation, underscoring DBA's joint delayed-label audit of non-parametric memory retention versus post-retrieval aggregation.
Reviewer_zSibOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how a retrieval-based clinical classifier can adapt over time. It proposes Diagnose-Before-Adapt (DBA), a prequential audit across a single delayed-label timeline. DBA measures the gain from changing one component at a time: refreshing memory (freshness), keeping cumulative memory (retention), changing the decision rule, or updating model parameters with LoRA. A support-constrained oracle provides a ceiling for any selector that selects only from retrieved candidate labels. The method, Learned Neighborhood Aggregation (LNA), is a small gradient-boosted ranker over 14 per-class features from lexical and dense retrieval.

Critical strengths
  • Interesting to see a huge case study consisting of 117328 consultations!
  • Moreover, temporally over 51 months
  • Useful framing. The paper asks which intervention has headroom (memory, decision rule, or parameters) before spending GPU on retraining.
Critical weaknesses

It is unclear from the abstract whether DBA is the authors' proposed method. I recommend stating explicitly what DBA is and how it differs from the second contribution, LNA.

The introduction is hard to follow in parts. From line 049 onward, it jumps between topics and introduces technical details without a clear rationale.

Figure 1 introduces the motivation, but this is not sufficient because it lacks citations. I recommend that the authors not place these arguments only in figures, and instead present them in the text with proper citations.

In Figure 1, it is hard to understand where the retrieved neighbors suddenly come from. Which classifier and which feature space are used? This notion is never introduced before the figure.

Questions for authors

Where do the three streams in the Method section come from? There is no citation, and if they are the authors' own construction, the wording does not make this clear.

The results section does not describe the results; it mostly quotes numbers, which are hard for readers to parse. It would be more informative to read the results and explain why the method worked or performed well, rather than a list of numbers.

Why do the tables not include baselines from existing methods?

5 comments
AuthorsOpenReview ↗

W1-W4: Separating DBA & LNA, Clinical Walkthrough, & Pre-Fig 1 Defs

We sincerely appreciate your careful reading and candid feedback on the manuscript's exposition, as well as your recognition of our 51-month (117,328-consultation) longitudinal scale and the practical value of diagnosing where adaptation headroom lies before spending GPU compute on retraining. While keeping manuscript edits minimal (preserving all Figures 1–20 and Tables 1–3, adding two-panel Table 4 in Section 4.3, and placing audit tables in Appendix K), we organize our reply into three parts: Part 1 covers presentation clarity (W1–W4), Part 2 covers stream provenance, clinical mechanisms, and literature baselines (Q1–Q3), and Part 3 addresses the Ethics Flag.

1. Cleanly Separating DBA from LNA (W1)

We agree that introducing DBA and LNA together in the original Abstract and Introduction risked blurring which contribution is the diagnostic framework and which is the adaptation method. In the revised Abstract and Section 1 (highlighted in navy blue), we separate their roles explicitly from Paragraph 1:

  • Diagnose-Before-Adapt (DBA) is our counterfactual prequential audit framework. On identical delayed-label visits, DBA measures the signed conditional gain and separate (GPU, CPU) cost of four distinct levers: memory freshness (recent vs. frozen), memory retention (cumulative vs. bounded), candidate-retrieval support (oto_t), and post-retrieval decision aggregation.
  • Learned Neighbourhood Aggregation (LNA) is our lightweight CPU adaptation method, directly motivated by DBA's diagnosis: once cumulative memory is retained, retrieval already covers the gold meridian in 93.67%–95.96% of queries on our two large streams while standalone prior recalibration and frozen-LLM score fusion saturate (-0.17 to +1.33 pts; Section 4.2). Thus, actionable low-compute headroom sits in how retrieved neighbours are combined.

2. Concrete Three-Step Clinical Walkthrough of LNA (W2)

To make LNA's inference and update cycle concrete on first encounter without relying on Figure 1 alone, we added a step-by-step clinical walkthrough when introducing Figure 1 in Section 1 (Pages 1–2) and expanded Figure 1's caption so its left (DBA audit) and right (LNA clinical workflow) panels are self-contained:

  1. Step 1 — Dual Frozen Retrieval (<t only): When an outpatient arrives in month tt with pre-decision intake text xtx_t (chief complaint, tongue and pulse findings, and optional intake Western diagnosis), LNA queries historical state St−1S_{t-1} (restricted strictly to months <t) for the top-k=32k=32 sparse lexical matches and top-k=32k=32 dense semantic matches. Post-decision fields (herbal prescription, syndrome notes, and month-tt label yty_t) are strictly withheld (Appendix K.1, Table 25).
  2. Step 2 — 14-Feature Relational Summarization per Meridian: Rather than prompting an LLM or using fixed voting, LNA summarizes the retrieved evidence for each of the six liu-jing meridians (Taiyang, Yangming, Shaoyang, Taiyin, Shaoyin, Jueyin) into a 14-D relational vector: lexical support count and weight mass, best/mean/relative similarity, reciprocal rank, and query overlap; the corresponding dense-retrieval statistics; cross-retriever agreement; and the historical prior π<t(c)\pi_{<t}(c) (kept even when a rare class has zero retrieved neighbours).
  3. Step 3 — CPU Gradient-Boosted Ranking & Delayed Refit: A compact 200-tree histogram-gradient-boosted (HGB) ranker scores all six 14-D candidate rows on CPU (23.8 ms median latency) to output y^t\hat y_t. Only after month-tt labels lock do non-evaluation cases append to memory and refit the HGB scorer on <t leave-one-out neighbourhoods N<t(−x)(x)N^{(-x)}_{<t}(x) (3,000 balanced historical cases, 500/class) to form StS_t.

3. Motivating Clinical Drift & Defining Encoders/Ranker for Figure 1 (W3, W4)

In revised Section 1 (Pages 1–2) and the self-contained Figure 1 caption, we now motivate longitudinal clinical shift (Nestor et al., 2019; Finlayson et al., 2021; Subasri et al., 2025) and explicitly define all three building blocks when Figure 1 is introduced: (i) sparse IDF lexical space over symptom, tongue/pulse, and intake-diagnosis tokens; (ii) frozen 512-D BAAI/bge-small-zh semantic space (Xiao et al., 2024) (fp32, max length 128) capturing paraphrased clinical descriptions without encoder backprop; and (iii) 200-tree HGB tabular ranker (31 leaves, lr = 0.08) fitted on <t relational rows.

AuthorsOpenReview ↗

Q1-Q3: Stream Provenance, Clinical Mechanism, & Canonical Baselines

Building on Part 1, we very much value your three questions probing stream provenance (Q1), the clinical mechanism behind our empirical results (Q2), and canonical literature baselines in the tables (Q3)—answering them made our empirical narrative much sharper.

1. Provenance & Construction of the Three Longitudinal Streams (Q1)

In revised Section 3.1, we clarify upfront that all three chronological delayed-label streams derive from the de-identified outpatient consultation archive introduced in Luo et al. (2025, ViTCM-LLM) (2021-04 to 2025-06, 12 recording clinicians, closed six-meridian target). Whereas Luo et al. (2025) studied static multimodal RAG, we organize the timestamped visits into three nested deployment regimes (after a frozen 12-month warm-up window, 2021-04 to 2022-03):

Longitudinal StreamConsultationsClinician ScopeSpan (Eval)Diagnostic Role in DBA
HOSP-Stream-Full117,328All 12 clinicians (26,099 + 79,535 + 11,694)51 mo (39 eval)Clinic-wide deployment under combined patient, seasonal, & staffing shift
HOSP-Stream-SingleMD26,099Single MD (f3ec4945, all 51 mo)51 mo (39 eval)Holds clinician constant to isolate intra-physician drift (Fig. 2)
HOSP-Stream-IndependentMD79,535Disjoint MD (8056e996, zero overlap with SingleMD)50 mo (38 eval)High-volume independent clinician stream (Tables 21–22 audit all 4 eligible MDs)

2. Clinical Interpretation of Cumulative Memory + HGB Ranking Gains (Q2)

One possible interpretation of these results in outpatient clinical context (noted in Sections 3.4 and 4.2–4.3) connects the observed gaps to domain structure:

  1. Why cumulative retention beats windows and reservoirs (Table 2): Pruning history to a 6-month window (BL3) loses 3.32, 1.89, and 4.57 pts across Full, SingleMD, and IndepMD, and a 20,000-case reservoir (BL4) loses 7.46 and 5.68 pts on the two >20k streams (0.00 on SingleMD), whereas retaining all 107,136 cases takes ~304 MiB (Table 16). Clinically, outpatient visits may exhibit annual seasonal recurrences (e.g., summer-dampness or winter-cold patterns) alongside long-tail meridians (Shaoyang, Shaoyin), so discarding multi-year cases may reduce retrieval support for recurring or rare patterns.
  2. Why HGB ranking beats fixed voting, recalibration, and neural poolers: Empirically, a rolling prior alone accounts for 6.4%–38.7% of LNA's gain (Figure 8b), and rare-class-weighted LNA+ lifts rare Shaoyang recall on SingleMD by +17.37 pts (Table 23). In TCM semiology, high dense similarity on broad complaints (fatigue, insomnia) may be noisy unless corroborated by specific tongue/pulse lexical terms, and decision-tree splits over ψ\psi may help capture threshold interactions—such as discounting dense similarity when sparse semiology disagrees or promoting rare syndromes when top matches concur against a dominant prior.

3. Connecting Table 1 & Table 4b Controls to Named Literature Methods (Q3)

To make prior-method comparisons prominent in the 10-page main text, we added canonical inline citations directly to Table 1 (tab_main) and added Table 4b in Section 4.3 (Page 10) summarizing controlled supervised aggregators:

  • Memory & concept-drift baselines (Table 1): frozen and cumulative kNN/RAG memory (Khandelwal et al., 2020; Lewis et al., 2020), sliding windows (Gama et al., 2014), and reservoir replay (Vitter, 1985).
  • Label-shift & post-hoc calibration (Table 1): prior-shift EM/reweighting (Saerens et al., 2002; Alexandari et al., 2020) and temperature scaling (Guo et al., 2017).
  • LLM RAG & parameter adaptation (Table 1): frozen Qwen3.5-4B RAG fusion (Brown et al., 2020; Luo et al., 2025), drift-triggered LoRA (Hu et al., 2022; Subasri et al., 2025), and monthly full-history LoRA retraining.
  • Supervised aggregators (Table 4b, Figure 11a) & public encoders (Table 12): on identical <t dual-retriever matrices (B=3,000B=3,000), we benchmark LNA–HGB against Logistic, Two-layer MLP, and DeepSets (Zaheer et al., 2017), plus published fine-tuned RoBERTa and ZY-BERT on 148-way TCM-SD (Table 12, where Base LNA reaches 82.72% Acc / 56.25% Macro-F1 vs. 82.26% / 47.55% and 82.19% / 51.01%):
Supervised Aggregator (Table 4b, B=3,000B=3,000)FullSingleMDIndepMDAdvantage of LNA–HGB
Logistic ranker76.1162.7474.12+1.63 to +2.71 pts
Two-layer MLP76.4863.6174.49+1.26 to +2.34 pts
DeepSets (Zaheer et al., 2017)69.8859.2769.31+5.92 to +7.86 pts
LNA–HGB (ours)77.7465.1976.83Best on all streams (Full/IndepMD match Table 10; SingleMD = active eval)
AuthorsOpenReview ↗

Ethics Flag: IRB Governance, De-identification, & Offline Protocol

Finally, we appreciate your raising the ethics flag so we could make our clinical data governance, de-identification, and non-redistribution safeguards explicit in the manuscript.

Dedicated Ethics Statement Added to the Revised Manuscript

We have added a standalone, factual Ethics Statement (immediately preceding the AI Use and Reproducibility statements on Page 11) documenting four governance safeguards:

  1. Source IRB Approval & Upstream PII Removal: Our hospital streams reuse the de-identified outpatient corpus introduced and documented by Luo et al. (2025, IEEE BIBM), who report institutional review board (IRB) approval at the participating hospitals and upstream removal of all direct personal identifiers (patient names, national/hospital ID numbers, phone numbers, and addresses) prior to research access, alongside the public TCM-SD benchmark (Ren et al., 2022).
  2. Pseudonymized Pipeline Indexing: Within our analysis pipeline, records are keyed solely by consultation code (case_id), calendar month (year_month), and a salted cryptographic hash of the recording clinician (doctor_id, e.g., f3ec4945, 8056e996). No cross-visit patient identifier is present, and no re-identification or linkage was attempted.
  3. Strictly Retrospective, Non-Interventional Evaluation: All experiments are conducted offline on historical records (2021-04 to 2025-06) and had zero interaction with live clinical care or treatment decisions.
  4. Strict Non-Redistribution of Private Clinical Text: Our anonymous reproducibility bundle releases only evaluation code, frozen configs, regression tests, and aggregate summary JSONs (Appendix J); raw consultation text, embeddings, and unsalted clinician identifiers remain strictly on the hospital host and are never redistributed.

Comments on the paper

1 comment

Ahead of the end of the discussion period, we wanted to leave a brief note for the Area Chair and Reviewers bringing together the new experiments and log audits added since our initial revision, along with the scope boundaries we will carry into the final version.

1. Initial revision and reviewer threads In our uploaded revision (Table 4 on p. 10 and Appendix K, Tables 25-29), we separated DBA from LNA, added a three-step clinical walkthrough and pre-Figure 1 definitions, added inline baseline citations to Table 1, and included an Ethics Statement (Luo et al., 2025), after which Reviewer zSib switched the inadvertent ethics flag to No. For Reviewers Vs5e and cxui, we unified Definition 1's notation, reported size-matched freshness slopes (+5.60 and +5.38 pts/yr, Table 6) and negative memory-aggregation substitution (I=−9.86,−8.44,−7.81I = -9.86, -8.44, -7.81 pts, Prop. 2, Table 17), separated multi-candidate recall oto_t (Table 26) and constant 12/12 LOSO routing (Table 27) from single-label claims, compared fixed-neighbourhood aggregators (Table 4b: LNA-HGB beating logistic, MLP, and DeepSets by 1.26-7.86 pts), documented pre-decision field roles (Table 25), and withdrew the confounded hospital budget sweep (Figure 17, Table 29). Under moving-block and Newey-West HAC inference (Table 4a, Table 28), Base LNA gains over BL2b (78.85/57.25/76.39 vs. 76.03/56.20/72.86) hold on Full (+2.83, NW-12 [+2.05, +3.61], p63<10−4p_{63} < 10^{-4}) and IndepMD (+3.53, NW-12 [+2.34, +4.71], p63<10−4p_{63} < 10^{-4}) at 0.0020-0.0111 vs. 2.28-2.64 GPU-h for 4B-LLM LoRA/retraining.

2. New experiments and hospital log audits added during discussion Over three rounds with Reviewer cxui—who concluded that our bounded-budget framing and scope limits "now accurately reflect the available evidence" with "no further questions"—we evaluated direct classifiers trained on frozen queries alone in 148-way TCM-SD (43,180 / 5,486 / 5,486) and audited our archived hospital logs:

Public TCM-SD (Acc / Macro-F1 %)3001,2003,0006,00012,000All 43,180
Base LNA (IDF + BGE neighbours)79.57 / 51.6481.81 / 55.1282.97 / 56.0182.88 / 54.85*83.65 / 56.0983.78 / 54.16
Direct Logistic (IDF + BGE)61.85 / 22.8870.56 / 26.0575.74 / 33.1578.31 / 38.3081.39 / 48.4984.25 / 57.51
Direct MLP (IDF + BGE)57.13 / 11.2469.65 / 26.6073.39 / 33.3776.85 / 36.83*79.15 / 45.0782.59 / 56.99

*300-12,000 use shared cross-class-shuffled draws with fixed sweep settings and 200-round HGB (early_stopping=False); validation-tuned 6k runs give 83.05/55.48 for Base LNA (+4.74/+17.18 over logistic) and 77.32/38.20 for MLP.

  • Where retrieval helps vs. saturates on TCM-SD: Pure 512-D bge-small-zh query heads fall well short even at 43,180 rows (70.69/42.82 logistic; 74.15/36.92 MLP). With sparse IDF + BGE features, Base LNA scales smoothly across 300-12,000 rows and leads direct logistic under every bounded cap (+7.23 Acc / +22.86 Macro-F1 at 3,000 rows, matching our monthly hospital cap, and +2.26 [1.39, 3.14] / +7.60 [4.47, 9.99] at 12,000). At full 43,180-row stationary IID scale, dual logistic matches 43k Base LNA in accuracy (LNA minus logistic: -0.47 [-1.20, +0.24]) and exceeds unweighted Base LNA in Macro-F1 (-3.35 [-5.26, -1.13]; 82.40/53.19 vs. 82.87/56.13 on 4,933 unseen test texts).
  • Archived hospital log replay: On the two 39-month freshness-probed streams (Table 6, Year-1 gaps +24.04 and +5.45), locking the best Year-1 arm in {BL2b, Base LNA, LNA+} over Months 13-39 achieved +1.11 [0.18, 2.03] over 4B-LoRA on Full (LNA+) and -1.20 [-2.32, -0.09] on SingleMD (BL2b), saving 1.76 GPU-h each.

3. Scope boundaries for the final manuscript We will fold these tables and audits into the final paper and frame our claims around four explicit limits: (i) on static TCM-SD, LNA's gain over dual IDF+BGE linear models holds under bounded refit budgets (<=12k rows) rather than full-corpus IID fitting, while hospital direct heads remain an unexecuted host check (Sec. 4.3, App. E) and hospital Figure 17 stays withdrawn; (ii) Full (117,328 visits) nests IndepMD (79,535; 1,481/7,600 shared eval visits) and SingleMD (26,099) rather than acting as an independent cohort, and the other 10 MDs (11,694 visits) reach at most 3,074 <t training cases (0/38 mo >=8k) after 200/mo eval sequestration; (iii) SingleMD's 39m policy (NW-12 [-0.44, +2.53]; 20 active mo +1.96, p63=1.0000p_{63} = 1.0000) and 4-MD HK [-0.71, +4.17] / PI [-5.16, +8.62] (Tables 21-22) cross zero, and unseen-patient validation still requires patient_id (Table 3); and (iv) DBA is an offline counterfactual audit rather than a dynamic controller (Table 27).

We sincerely thank all three reviewers and the Area Chair for their constructive engagement, and remain available for any final questions before the window closes.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.