acceptodds
Under review as a conference paper at ICLR 2027

TAPS: Tree-Adaptive Prediction Sets via LCA-Weighted Nonconformity Scoring for Hierarchical Multi-Label Clinical Coding

Abstract

Conformal prediction (CP) provides distribution-free, finite-sample coverage guarantees. However, applying CP to hierarchical multi-label classification leads to a combinatorial explosion over possible label subsets. Recent parallel work (Mortier et al., 2026; den Hengst et al., 2025) addresses this by allowing prediction sets to include internal nodes of a taxonomy. We identify a complementary gap: prior methods either use flat leaf-level conformity scores or incorporate hierarchy without an explicit abstraction-cost penalty in the conformity score. We introduce **Tree-Adaptive Prediction Sets (TAPS)**, a CP algorithm that directly integrates hierarchical structure into the calibration score. TAPS employs a *specificity-regularized Lowest-Common-Ancestor (LCA) score*, which penalizes each tree node based on its uncertainty and a height-based abstraction cost (weighted by ). Crucially, this identical score governs both calibration and inference. We prove that the TAPS score is dominated by the flat score, ensuring calibration thresholds remain no larger than those of flat CP at the same . We also prove TAPS guarantees marginal semantic coverage under exchangeability, with calibration and inference provably evaluating the exact same event. The parameter controls an empirical compactness-specificity trade-off while conformal validity is preserved after independent selection, requiring a non-degeneracy screen to prevent trivial root-collapse. On MIMIC-IV automated ICD-10 coding (QLoRA-tuned Phi-3-mini, full training set), TAPS attains near-nominal empirical semantic joint coverage, comparable to Flat LAC, while producing prediction sets roughly 48% smaller than flat CP. It achieves this by summarising uncertainty via ancestor nodes, trading off exact leaf-level specificity. We evaluate against eleven baselines spanning flat, hierarchical, and multi-label conformal methods, and identify a common failure mode, which we call the 'free root' pathology, that causes every hierarchy-aware baseline we test lacking score-level integration to exhibit severe undercoverage, root-collapse, or substantially inflated prediction sets. We further validate TAPS under the same calibration and evaluation framework on a second, structurally distinct domain, the arXiv category taxonomy (depth , vs. ICD-10's depth ), providing evidence that both the compactness gains and the baseline failure modes are not specific to clinical coding.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
conformal prediction, hierarchical classification, multi-label classification, uncertainty quantification, prediction sets

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_idR6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper purposes Tree-Adaptive Sets (TAPS), a method for conformal prediction in hierarchical multi-label classification. The purposed method constructs a hierarchy-aware nonconformity score that combines class probabilities with a penalty for selecting less specific ancestor nodes. A conformal calibration step adapted to tree structure is then applied to construct prediction set using the highest-level nodes satisfying the calibration threshold. The paper establishes finite-sample semantic coverage guarantees under exchangeability. Experiments on clinical coding and other hierarchical classification datasets demonstrate reductions in prediction-set size compared with the considered baselines.

Critical strengths
  • The problem considered is practically relevant. Incorporating hierarchical structure into prediction sets provides a natural way to produce more compact and interpretable predictions.
  • The paper presents relatively extensive experiments across multiple datasets, including comparisons with various baselines and ablation studies.
Critical weaknesses
  • The purposed method mainly introduces a hierarchy-aware nonconformity score and a corresponding prediction-set construction, followed by standard split conformal calibration. The semantic coverage guarantee also largely follows from standard conformal arguments.
  • A natural alternative is to first construct valid prediction sets using standard conformal methods such as LAC, APS, or RAPS, and then merge selected leaf labels into ancestor nodes. Such post-processing would preserve semantic coverage while potentially reducing prediction set size. Although the paper considers a related ablation, it does not directly compare against more carefully designed hierarchical compression procedures. This makes it difficult to assess the benefit of incorporating hierarchical penalties directly into the nonconformity score.
Questions for authors

Related to the second point in the weakness section, Table 3 in the ablation section reports that flat score + hierarchical calibration always outputs the root note. Is this because, when using the flat score, the root node is guaranteed to have a lower nonconformity score than any of its descendants, and thus cannot be pruned? If so, the author should also consider better post-processing procedures applied directly to leaf-level prediction sets, to ensure a fair comparison.

In addition, the trade-off between prediction-set size and specificity should be discussed more thoroughly. In its current form, I find it difficult to interpret the effectiveness and gains of the proposed method, since it is unclear what constitutes the best balance between set size and specificity.

1 comment

We thank the reviewer for a thoughtful review, in particular for proposing a direct comparison with post-hoc compression of conformal sets. A revised paper and supplement have been uploaded; related weaknesses (W) and questions (Q) are grouped below. Status tags: [Addressed] changed in the revision; [Partly] addressed with a stated limitation; [Agreed] the point is accepted and the framing changed.

Summary of changes relevant to this review: new post-hoc lifts of flat LAC with leaf-expanded size, LCAD and a matched-target table (Appendix R); λ sweep (Appendix S); coverage intervals (Appendix T); MeSH correction (Appendix O).

W1. Novelty beyond standard conformal prediction. [Agreed] The coverage guarantee is standard split conformal prediction and dominance is immediate; dominance is now a proposition. The contribution is a hierarchy-aware nonconformity score used directly in calibration and prediction-set construction, as opposed to a separate representation transformation applied to an already constructed flat conformal set, together with the measured trade-offs. The lifts also use calibration quantiles; λ is selected on a separate split. The selection rule chooses λ = 0 on arXiv and DBpedia14, so the height penalty is inactive there. This separates the path-minimum construction from the penalty: the path-based score and output rule remain active at λ = 0, while the penalty matters on MIMIC-IV (λ = 0.10), where its removal yields about 70% root-only sets (Appendix S).

W2. Post-hoc compression of standard conformal sets. [Addressed] Lifts of the flat LAC set were added: L1 (lossless), L2 (fractional), L3 (probability-thresholded) and a more aggressive L3x (Appendix R). For a fixed flat set, replacing leaves by ancestors cannot uncover a label that is already semantically covered; our experiments additionally tune the lift parameters and the LAC level, so this fixed-set property is distinct from a finite-sample guarantee for the full tuned procedure. Three-seed means at α = 0.10 (nodes / leaf-expanded size / LCAD), MIMIC-IV: TAPS 17.08 / 36.0 / 2.83; L3 tuned for fewest nodes 9.57 / 39.3 / 3.12; lossless L1 17.63 / 33.0 / 2.07; L2 tuned for leaf-expanded size 15.08 / 34.1 / 2.44, at coverage 0.897 against 0.891 for TAPS. L2 attains numerically higher observed coverage and lower three-seed means for nodes, leaf-expanded size and LCAD; these are observed operating points, and the tuning of the lift does not establish a finite-sample guarantee for the full procedure. L1 differs: it has no tuned parameter, is applied at the nominal level and inherits the guarantee of flat LAC; it returns slightly more nodes than TAPS, with a smaller leaf-expanded set and lower LCAD. arXiv: TAPS 1.66 / 43.1 / 1.05; L3 1.98 / 42.1 / 0.97, at coverage 0.919 against 0.900. A carefully designed lift is therefore competitive; the paper no longer claims that TAPS is more compact or more specific.

Q1. Why Table 3 returns the root. [Agreed] The reviewer's explanation is correct: with a flat score the nonconformity score of a node is 1 − p̂_v, and the root probability is approximately 1 (Appendix I), so the root has the smallest score of all nodes and passes whenever any node passes; the topmost-node rule therefore returns the root. This ablation is degenerate by construction. It shows that replacing the hierarchy-aware TAPS score with a flat score under the same construction causes root collapse; the designed lifts show that other output rules can avoid this pathology.

Q2. Size-specificity balance. [Partly] The paper now reports leaf-expanded size and LCAD alongside node count, and a λ frontier for MIMIC-IV. No method is uniformly preferable: flat sets retain fine resolution, whereas hierarchical sets yield a shorter node list. No clinician study has been conducted, and no claim is made that either approach reduces coding effort.

Corrections found during our audit [identified on our side while preparing this rebuttal; not raised in the reviews]. (i) The MeSH loader inserted a second dot into tree numbers that were already dotted, so that 99.14% of MeSH labels were dropped during evaluation. MIMIC-IV, arXiv and DBpedia14 are unaffected. After correcting the loader and retraining (one seed), the corrected MeSH result is unfavourable to TAPS (coverage 0.891 with 61.7 nodes covering all 282 leaves, against 0.907 with 10.7 nodes for LAC + L3x); the claim that TAPS generalises to MeSH is withdrawn (Appendix O). (ii) Across three seeds, observed MIMIC-IV joint coverage ranges from 0.8886 to 0.8927 for TAPS and from 0.8873 to 0.8924 for flat LAC. Separate document-level bootstrap 95% intervals exclude 0.90 but may underestimate uncertainty due to repeated admissions per patient. The cause remains unresolved (Appendix T).

We hope that the revision addresses the reviewer's concerns, and we would welcome any further comments.

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

This paper proposes TAPS, a conformal prediction framework for hierarchical multi label classification, with applications to clinical ICD coding. The method incorporates tree structure into the nonconformity score using a regularized lowest common ancestor formulation. The authors establish finite sample semantic coverage guarantees and evaluate the approach on MIMIC IV and other hierarchical datasets. The results show that TAPS produces prediction sets about 48% smaller than flat conformal prediction while maintaining near nominal semantic coverage.

Critical strengths

The paper addresses an interesting problem of uncertainty quantification for hierarchical multi label prediction. Incorporating hierarchy directly into the nonconformity score is a reasonable approach, and the calibration procedure preserves the standard conformal coverage guarantee.

The experimental evaluation is fairly comprehensive, with comparisons against several conformal baselines. The ablation studies are useful, and the reduction in prediction set size is encouraging.

Critical weaknesses

The main concern is that the improvement in prediction set size comes at a substantial cost to prediction specificity. In Table 2, TAPS reduces the average set size from 33.08 to 17.08 compared with Flat LAC, but the LCA distance increases from 0.069 to 2.833. Since predicting an ancestor counts as correct, the reported coverage guarantee is considerably weaker than exact code coverage. It is unclear whether these smaller sets are actually more useful for clinical coding, where the specific diagnosis code matters.

The theoretical contributions seem somewhat incremental. The score dominance result in Theorem 1 follows directly from including the true leaf in the minimization, while the coverage guarantee follows from standard split conformal prediction. The compactness result also primarily reflects replacing multiple leaves with their ancestors. Furthermore, some hierarchical baselines are adapted from single label settings, making it difficult to interpret their poor performance as evidence of a fundamental limitation of existing approaches.

Questions for authors

Can the authors provide a comparison that accounts for both prediction set size and specificity? For example, how does TAPS compare with flat conformal prediction when the hierarchical predictions are expanded into their corresponding leaf codes?

How should the significant increase in LCA distance be interpreted in practical clinical coding? Is there evidence that clinicians would prefer the smaller but less specific prediction sets?

Table 2 reports joint coverage of 0.8905 against a nominal target of 0.90. Can the authors clarify whether this difference is explained by sampling variability and provide uncertainty estimates for the empirical coverage?

1 comment

We thank the reviewer for a clear and careful review. A revised paper and supplement have been uploaded; related weaknesses (W) and questions (Q) are grouped below. Status tags: [Addressed] changed in the revision; [Partly] addressed with a stated limitation; [Limitation] acknowledged but not resolved.

Summary of changes relevant to this review: leaf-expanded size and LCAD reported alongside node count, post-hoc lifts and a matched-target table (Appendix R); λ sweep (Appendix S); coverage intervals (Appendix T); MeSH correction (Appendix O); baseline provenance (Section 5.1, Appendices G to I, Table 6). A clinician study remains a limitation.

W1 / Q1. Set size versus specificity; comparison on expanded leaf codes. [Addressed] The reviewer is correct that the reduction in nodes is accompanied by a large increase in LCAD (0.069 to 2.833), and that coverage through an ancestor is semantically coarser than exact-code coverage. TAPS expands to 36.0 leaf codes against 33.1 for flat LAC; it therefore returns fewer hierarchy nodes, not fewer implied leaf codes. Leaf-expanded size and LCAD are now reported alongside node count for flat LAC, TAPS and four post-hoc lifts on every dataset, together with a λ sweep for MIMIC-IV (Appendices R and S). The paper now states that users who require exact codes should prefer flat conformal prediction.

Post-hoc lifting (request for a fair comparison). [Addressed] Three-seed means at α = 0.10 (nodes / leaf-expanded size / LCAD), MIMIC-IV: flat LAC 33.08 / 33.1 / 0.07; TAPS 17.08 / 36.0 / 2.83; L3 tuned for fewest nodes 9.57 / 39.3 / 3.12; L2 tuned for leaf-expanded size 15.08 / 34.1 / 2.44, at coverage 0.897 against 0.891 for TAPS. L2 attains numerically higher observed coverage and lower three-seed means for nodes, leaf-expanded size and LCAD; these are observed operating points, and the tuning of the lift does not establish a finite-sample guarantee for the full procedure. arXiv: TAPS 1.66 / 43.1 / 1.05; L3 1.98 / 42.1 / 0.97, at coverage 0.919 against 0.900. Any claim that TAPS is the most compact or the most specific method is withdrawn.

Q2. Clinical interpretation of LCAD. [Limitation] No clinician study has been conducted, and one cannot be completed before the deadline; no claim is therefore made that coders prefer the smaller sets. LCAD, leaf-expanded size and the lifts serve as proxies, and the absence of a user study is stated as a limitation.

Q3. Coverage 0.8905 against 0.90. [Addressed, cause not established] Per-seed TAPS coverage on MIMIC-IV is 0.8886, 0.8927 and 0.8903 (Appendix T); flat LAC behaves identically (0.8873 to 0.8924), so the shortfall is not specific to TAPS. The document-level bootstrap 95% intervals exclude 0.90, but they reflect test-set sampling only, whereas the guarantee is also marginal over the calibration draw. The three seeds share a data split and are not independent replicates. Document-level intervals do not fully account for calibration variability or repeated admissions per patient. Consequently, we cannot conclusively attribute the observed shortfall to sampling variability or distribution shift. We report this discrepancy transparently and distinguish empirical coverage from the theoretical guarantee under exchangeability. The arXiv (0.8999) and DBpedia14 intervals include 0.90.

W2a. Incremental theory. [Agreed] Dominance follows directly from including the true leaf in the minimisation, the coverage guarantee is standard split conformal prediction, and compactness reflects the replacement of leaves by ancestors. Dominance is now presented as a proposition; the contribution lies in the score and the empirical analysis.

W2b. Baselines adapted from single-label settings. [Partly] Agreed: CRSVP, HCC and CSP were designed for other settings, so their failures do not indicate a limitation of the original ideas. Section 5.1 and Appendices G to I describe each adaptation, the MeSH table (Table 6) marks every method as native or adapted, and the corresponding claims have been moderated.

Correction found during our audit [identified on our side while preparing this rebuttal; not raised in the reviews]. The MeSH loader inserted a second dot into tree numbers that were already dotted so that 99.14% of MeSH labels were dropped during evaluation. MIMIC-IV, arXiv and DBpedia14 are unaffected. After correcting the loader and retraining (one seed), the corrected MeSH result is unfavourable to TAPS (coverage 0.891 with 61.7 nodes covering all 282 leaves, against 0.907 with 10.7 nodes for LAC + L3x). The claim that TAPS generalises to MeSH is withdrawn, and tall hierarchies with many high internal labels are listed as a limitation (Appendix O).

We hope that the revision addresses the reviewer's concerns, and we would welcome any further comments.

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

TAPS is a split-conformal method for hierarchical multi-label classification, applied to ICD-10 coding. Its nonconformity score is the minimum, over the true label and its ancestors, of a node score. The same score is used for calibration and set construction, and the output keeps the maximal passing nodes. TAPS reaches semantic joint coverage with less nodes than flat methods, but with lower specificity.

Critical strengths
  • relevant topic, natural and clean idea, interesting contribution, flat conformal sets become unusably large for hierarchical clinical coding
  • well described algorithm , definitions defined compact and explained well, design choices such as height vs depth are justified
  • careful and broad evaluation
  • loss of specificity is reported openly
Critical weaknesses
  • semantic coverage and node-count set size favor the method (see ablation arm 2, with 36 leaf codes against 33 for flat LAC)
  • the free-root pathology is a consequence of the metric, several baselines were not designed for semantic coverage, providing for example single labels only
  • the experiments use only the top-50 codes (out of 70.000), covering about a third of instances, but hierarchical abstraction is more useful for the long tail
  • missing literature: “Hedging Your Bets” (Deng, Krause, Berg & Fei-Fei, CVPR 2012) “Hierarchical classification at multiple operating points” (Valmadre, 2022)
  • limited theoretical novelty: theorem 1 follows in one line from the score definition, coverage guarantee is standard split-conformal validity for a fixed score
Questions for authors
  • Can you report the leaf-expanded set size (as in arm 2) for all methods in table 2?
  • Is a set of 17 nodes at depth 3 more useful to a coder than 33 leave nodes? a clinician assessment would be helpful
  • How does TAPS behave on the full label space or at least a larger subset (top-500)?
2 comments

We thank the reviewer for a careful and constructive review and for the two references. A revised paper and supplement have been uploaded. Related weaknesses (W) and questions (Q) are grouped below. Status tags: [Addressed] changed in the revision; [Partly] addressed with a stated limitation; [Limitation] acknowledged but not resolved.

Summary of changes relevant to this review: leaf-expanded size and LCAD reported alongside node count, post-hoc lifts and a matched-target table (Appendix R); λ sweep (Appendix S); coverage intervals (Appendix T); MeSH correction (Appendix O); baseline provenance (Section 5.1, Appendices G to I, Table 6).

W1 / Q1. Node count and semantic coverage favour the method; leaf-expanded size for all methods. [Partly] On MIMIC-IV, TAPS returns 17.08 nodes against 33.08 for flat LAC, but expands to 36.0 leaf codes against 33.1, and LCAD increases from 0.069 to 2.83. The paper now refers to fewer nodes rather than smaller sets, and reports leaf-expanded size and LCAD alongside every node count. Leaf-expanded size is reported for flat LAC, TAPS and four post-hoc lifts on all main datasets and for all methods on MeSH. For methods that output leaf codes (flat LAC, APS, RAPS, BR-MLCP, Fisch et al.) it equals the set size already reported in Table 2 (33.1, 46.7, 46.7, 22.8, 39.0); it has not been recomputed for the hierarchical-output baselines (CRSVP, HCC, CRC-Hierarchical, DSB) on MIMIC-IV, so the comparison is not exhaustive. A lift tuned for leaf-expanded size (L2) attains lower three-seed means than TAPS for node count (15.08 vs. 17.08), leaf-expanded size (34.1 vs. 36.0) and LCAD (2.44 vs. 2.83) on MIMIC-IV, with a numerically higher observed coverage (0.897 vs. 0.891). These are observed operating points; the tuning of the lift does not establish a finite-sample guarantee for the full procedure. The paper therefore no longer claims that TAPS is the most compact or specific method.

W2. Free root and baselines not designed for semantic coverage. [Partly] Agreed. The free-root behaviour is a consequence of the semantic coverage criterion, and CRSVP, HCC, CSP and Tree-CP were designed for other settings; their failures therefore indicate limited transferability rather than deficiencies of the underlying ideas. Section 5.1 and Appendices G to I describe how each method was adapted to the multi-label joint protocol, the MeSH table (Table 6) marks every method as native or adapted, and the wording on free-root pathologies has been moderated.

W3 / Q3. Top-50 only; top-500. [Limitation] The evaluation covers 33.9% of MIMIC-IV ICD-10 instances, and no claim is made about the long tail; this is stated as a limitation. A top-500 evaluation requires retraining the encoder with a substantially larger output layer, which cannot be completed within the discussion period. The arXiv taxonomy (177 nodes) and the corrected MeSH tree (593 nodes) extend the evaluation beyond the MIMIC-IV top-50 hierarchy; however, the corrected MeSH result does not support generalisation of the selected TAPS operating point.

W4. Missing literature. [Addressed] Deng et al. (CVPR 2012) and Valmadre (NeurIPS 2022) are now cited. These works trade accuracy against specificity using single-node outputs and operating curves; neither provides a distribution-free finite-sample guarantee for set-valued multi-label outputs.

W5. Limited theoretical novelty. [Agreed] Dominance follows in one line from including the true leaf in the minimisation, and the coverage guarantee is standard split conformal prediction. Dominance is now presented as a proposition; the contribution lies in the score, the free-root and compactness analysis, and the empirical evaluation.

Q2. Clinical usefulness. [Limitation] No clinician study has been conducted, and no claim is made that coders prefer 17 nodes to 33 leaf codes. The absence of a user study is a stated limitation.

Corrections found during our audit [identified on our side while preparing this rebuttal; not raised in the reviews]. (i) The MeSH loader inserted a second dot into tree numbers that were already dotted, so that 99.14% of MeSH labels were dropped during evaluation. MIMIC-IV, arXiv and DBpedia14 are unaffected. After correcting the loader and retraining (one seed), the corrected MeSH result is unfavourable to TAPS (coverage 0.891 with 61.7 nodes covering all 282 leaves, against 0.907 with 10.7 nodes for LAC + L3x), and the claim that TAPS generalises to MeSH is withdrawn (Appendix O). (ii) Across three seeds, observed MIMIC-IV joint coverage ranges from 0.8886 to 0.8927 for TAPS and from 0.8873 to 0.8924 for flat LAC. Separate document-level bootstrap 95% intervals exclude 0.90 but may underestimate uncertainty due to repeated admissions per patient. The cause remains unresolved (Appendix T).

We hope that the revision addresses the reviewer's concerns, and we would welcome any further comments.

Thank you again for your review and for asking whether the findings extend beyond the top-50 ICD codes (Q3). We have now run the requested 500-leaf experiment, which updates our earlier answer to W3/Q3 (we had said it could not be completed in the discussion period). Its completion was delayed by Google Colab runtime issues on our side. The full setup, all results and limitations are in our shared comment to all reviewers and in Appendix U of the revised manuscript.

In brief, at α=0.10\alpha = 0.10 on 500500 leaf codes (1,1101{,}110-node hierarchy), TAPS reaches joint coverage 0.8990.899 with 81.181.1 nodes, versus 0.9040.904 and 297.7297.7 nodes for flat LAC (approximately 7373\\% fewer nodes). TAPS has a larger leaf-expanded size (348.7348.7 versus 297.7297.7) and higher LCAD (3.053.05 versus 0.030.03). The direction matches the top-50 experiment, but the magnitudes differ. Post-hoc lifts are competitive: L3 tuned for leaf-expanded size gives 0.9100.910 coverage, 95.495.4 nodes, 349.6349.6 leaves and LCAD 2.972.97, while lifts tuned for fewest nodes (15.715.7 to 24.024.0 nodes) are far more compact than TAPS but overcover (0.9480.948 to 0.9780.978). We therefore do not claim that TAPS outperforms lifting.

Limitations. This is a single seed and split with a different eligible cohort from the top-50 study. The 500500-code vocabulary is chosen from full-dataset code frequencies, which our fixed-vocabulary guarantee does not cover. We make no claim about the full ICD-10-CM label space or the long tail.

Sincerely,

The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Additional experiment: 500-leaf scale check on MIMIC-IV (to all reviewers)

Additional experiment: 500-leaf scale check on MIMIC-IV (to all reviewers)

Dear Reviewers,

We thank all three reviewers for their assessments. Following our individual replies, we completed an additional experiment motivated by Reviewer PzHx's question about performance beyond the top-50 ICD codes. We share it with all reviewers because it also bears on the questions about hierarchical compression, specificity and post-hoc lifting.

Setup. We selected the 500 most frequent target codes that are leaves in the full ICD-10-CM hierarchy (the main experiment's 50-code list retains one internal category, E87.2), yielding a prediction hierarchy with 1,1101{,}110 nodes. Frequencies are counted over the complete dataset before splitting, as in the main experiment. We retained the model architecture, QLoRA training recipe (512512 tokens, two epochs), nominal error level α=0.10\alpha = 0.10, and calibration framework. After removing admissions without a selected target code, subjects were partitioned 80/10/1080/10/10 so that no patient appears in more than one partition, giving 96,63596{,}635 training, 11,89511{,}895 calibration and 11,85911{,}859 test admissions. This is a new cohort and partition, so the results are not a paired comparison with the top-50 experiment.

Results (joint coverage / nodes / leaf-expanded size / LCAD; all lifts shown):

  • TAPS (λ=0.10\lambda = 0.10): 0.8990.899 / 81.181.1 / 348.7348.7 / 3.053.05
  • Flat LAC: 0.9040.904 / 297.7297.7 / 297.7297.7 / 0.030.03
  • LAC + L1: 0.9070.907 / 173.3173.3 / 299.9299.9 / 1.451.45
  • LAC + L3 (minimum leaf-expanded size): 0.9100.910 / 95.495.4 / 349.6349.6 / 2.972.97
  • LAC + L2 (minimum nodes): 0.9780.978 / 24.024.0 / 438.8438.8 / 3.433.43
  • LAC + L3 (minimum nodes): 0.9480.948 / 15.715.7 / 448.6448.6 / 3.463.46
  • LAC + L3x (minimum leaf-expanded size): 0.9100.910 / 95.495.4 / 350.8350.8 / 2.982.98
  • LAC + L3x (minimum nodes): 0.9740.974 / 15.715.7 / 468.9468.9 / 3.483.48

Interpretation. TAPS emits approximately 7373\\% fewer nodes than flat LAC, but its leaf-expanded size (348.7348.7 versus 297.7297.7) and LCAD are higher. The direction matches the top-50 experiment (4848\\% fewer nodes, 99\\% larger leaf-expanded set), but the magnitudes differ. Lifts tuned for fewest nodes (15.715.7 to 24.024.0 nodes) are much more compact than TAPS but overcover (0.9480.948 to 0.9780.978) and have larger leaf-expanded sets. The L3 lift optimized for leaf-expanded size gives a similar leaf-expanded size and LCAD with higher observed coverage, although it emits more nodes. We therefore do not claim superiority over post-hoc lifting.

Scope and limitations. This is a single-seed, single-split experiment, and the methods are not compared at matched coverage: flat LAC uses the full calibration partition, whereas the lifts use the smaller DconfD_{\mathrm{conf}} partition and separate tuning data. TAPS' observed coverage (0.8990.899) is slightly below the nominal 0.900.90, as in the main MIMIC-IV experiment. The target vocabulary is selected using code frequencies over the full dataset, including examples later assigned to calibration and test; our fixed-vocabulary guarantee does not account for that selection step. We therefore present this as an empirical larger-label-space check, not as a new theoretical validation.

The protocol, results and limitations are in Appendix U of the revised manuscript.

Sincerely,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.