Tree of Concepts: Interpretable Continual Learners in Non-Stationary Clinical Domains

Dongkyu Cho; Rumi Chunara; Samrachana Adhikari; Xiyue Li

arxiv: 2604.17089 · v1 · submitted 2026-04-18 · 💻 cs.LG

Tree of Concepts: Interpretable Continual Learners in Non-Stationary Clinical Domains

Dongkyu Cho , Xiyue Li , Samrachana Adhikari , Rumi Chunara This is my paper

Pith reviewed 2026-05-10 06:40 UTC · model grok-4.3

classification 💻 cs.LG

keywords continual learninginterpretabilityhealthcareconcept bottleneckdecision treestability-plasticitynon-stationary datatabular benchmarks

0 comments

The pith

A fixed shallow decision tree supplies stable concepts that let models adapt to new clinical data while keeping explanations consistent.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to resolve the conflict in continual learning between the need for models to adapt to shifting data and the requirement for stable, interpretable explanations in healthcare. It does so by anchoring a concept bottleneck model to concepts extracted from a shallow decision tree that stays fixed across updates. Only the parts that map raw features to those concepts and then to labels are allowed to change, so the rules behind the explanations do not drift. On tabular healthcare datasets under standard continual-learning protocols the resulting models retain more past performance while incorporating new patterns than replay-based baselines. Readers should care because medical systems must update without producing explanations that suddenly become untrustworthy or require re-auditing.

Core claim

Tree of Concepts defines a fixed set of rule-based concepts using a shallow decision tree and trains a concept bottleneck model to predict both those concepts and the final labels. Continual updates are performed only on the feature-to-concept extractor and the concept-to-label head, leaving the tree-defined concepts and their semantics unchanged. This separation produces explanations that remain stable across sequential distribution shifts while still allowing the model to improve on new data.

What carries the argument

The shallow decision tree that supplies a fixed, rule-based concept interface to a concept bottleneck model.

If this is right

Explanations remain semantically unchanged across multiple model updates.
The stability-plasticity trade-off improves relative to replay-enhanced continual learners on the tested healthcare benchmarks.
A consistent audit interface is preserved without requiring redefinition of concepts when new data arrives.
Interpretability does not have to be sacrificed for the ability to adapt in non-stationary clinical settings.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same fixed-concept anchoring could be tested in other regulated domains where initial expert rules can be encoded as shallow trees.
Periodic checks would still be needed to confirm that the original tree concepts continue to match current domain knowledge.
If the tree misses important factors, the method would force the rest of the model to work around incomplete concepts, potentially limiting overall accuracy.

Load-bearing premise

The concepts chosen by the initial shallow decision tree stay relevant and sufficient for accurate predictions even after the patient data distribution changes.

What would settle it

After several rounds of distribution shift the fixed concepts lose predictive accuracy or produce explanations that clinicians judge as no longer meaningful on new data.

Figures

Figures reproduced from arXiv: 2604.17089 by Dongkyu Cho, Rumi Chunara, Samrachana Adhikari, Xiyue Li.

**Figure 2.** Figure 2: Stability–plasticity trade-offs on MIMIC-III mortality prediction under protocols [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

read the original abstract

Continual learning aims to update models under distribution shift without forgetting, yet many high-stakes deployments, such as healthcare, also require interpretability. In practice, models that adapt well (e.g., deep networks) are often opaque, while models that are interpretable (e.g., decision trees) are brittle under shift, making it difficult to achieve both properties simultaneously. In response, we propose Tree of Concepts, an interpretable continual learning framework that uses a shallow decision tree to define a fixed, rule-based concept interface and trains a concept bottleneck model to predict these concepts from raw features. Continual updates act on the concept extractor and label head while keeping concept semantics stable over time, yielding explanations that do not drift across sequential updates. On multiple tabular healthcare benchmarks under continual learning protocols, our method achieves a stronger stability-plasticity trade-off than existing baselines, including replay-enhanced variants. Our results suggest that structured concept interfaces can support continual adaptation while preserving a consistent audit interface in non-stationary, high-stakes domains.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Freezing an early shallow decision tree to define concepts lets the rest of the model adapt without drifting explanations, but the fixed concepts may not stay sufficient after real clinical shifts.

read the letter

The core idea is to lock concept definitions in a shallow decision tree trained on initial data, then train a concept bottleneck model where only the extractor and label head update under continual learning. This keeps the rule-based explanations consistent across tasks while allowing some plasticity on tabular healthcare benchmarks. The approach is a direct response to the tension between opaque deep models that adapt well and brittle interpretable ones that do not. It is new in how it combines a fixed concept interface with selective updates rather than replay or regularization alone. The paper does a clean job framing the problem for high-stakes domains where auditability matters and reports better stability-plasticity trade-offs than replay baselines. That framing is useful even if the numbers need checking. The main soft spot is the untested assumption that the initial tree's concepts remain relevant and complete after distribution shifts. If later tasks require new clinical concepts not captured early, the model can only adapt inside a limited space, which could cap accuracy or force the explanations to stay superficial. The abstract gives no sign that the tree was validated on future tasks or that the chosen shifts preserve the concept set. Experimental details are also thin: no dataset splits, no statistical tests, and no clear description of how baselines were implemented or whether post-hoc choices affected the results. This makes the central claim hard to assess from the given text. The work is aimed at researchers who care about interpretable continual learning in regulated settings like medicine. A reader focused on deployment constraints would find the architecture discussion worth their time, though the empirical side needs more transparency before it changes practice. I would send it to peer review so the full experiments and any concept-relevance checks can be examined.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Tree of Concepts, an interpretable continual learning framework for non-stationary clinical domains. A shallow decision tree is fixed to define a stable, rule-based concept interface; a concept bottleneck model then maps raw tabular features to these concepts, with continual updates restricted to the concept extractor and label head while concept semantics remain frozen to prevent explanation drift. The central claim is that this yields a stronger stability-plasticity trade-off than existing baselines (including replay variants) on multiple tabular healthcare benchmarks under continual learning protocols.

Significance. If the fixed-concept assumption holds, the framework provides a concrete mechanism for non-drifting, auditable explanations during continual adaptation, which is valuable for high-stakes non-stationary domains. The approach usefully combines the interpretability of decision trees with the adaptability of neural concept extractors. Credit is due for the explicit design choice of a frozen, structured interface rather than post-hoc explanation methods.

major comments (2)

[Abstract and §3] Abstract and §3 (method): the stability claim rests on the fixed shallow decision tree remaining sufficient and semantically stable after distribution shifts, yet no validation, sensitivity analysis, or discussion of concept coverage on later tasks is provided; if the initial tree misses concepts required post-shift, the reported trade-off may reflect an incomplete feature space rather than the framework's merit.
[§4] §4 (experiments): the abstract states improved stability-plasticity results on benchmarks but supplies no details on dataset splits, statistical tests, exact baseline configurations, or controls for post-hoc choices, leaving the central empirical claim with limited verifiable support.

minor comments (2)

[§3.1] Clarify in §3.1 whether the decision tree is fit exclusively on the first task or uses any form of concept selection across tasks.
[Related Work] Add explicit comparison to prior concept-bottleneck continual learning work and decision-tree-based continual methods in the related-work section.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive report. We address each major comment below and outline revisions that will strengthen the manuscript without altering its core claims.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (method): the stability claim rests on the fixed shallow decision tree remaining sufficient and semantically stable after distribution shifts, yet no validation, sensitivity analysis, or discussion of concept coverage on later tasks is provided; if the initial tree misses concepts required post-shift, the reported trade-off may reflect an incomplete feature space rather than the framework's merit.

Authors: We agree that the stability of explanations depends on the initial tree providing adequate concept coverage. The manuscript presents the fixed tree as a design choice that prevents semantic drift, with the concept extractor adapting to new data distributions while the rule-based interface remains unchanged. However, we acknowledge the absence of explicit sensitivity analysis or coverage checks on later tasks. In revision we will add a dedicated paragraph in §3 discussing the assumption that the initial tree spans the relevant concept space for the clinical domain, together with a new ablation in §4 that measures performance and explanation consistency when the initial tree is deliberately restricted to a subset of concepts. This will clarify the conditions under which the reported stability-plasticity trade-off holds. revision: yes
Referee: [§4] §4 (experiments): the abstract states improved stability-plasticity results on benchmarks but supplies no details on dataset splits, statistical tests, exact baseline configurations, or controls for post-hoc choices, leaving the central empirical claim with limited verifiable support.

Authors: We appreciate the call for greater experimental transparency. While §4 and the supplementary material already specify the continual-learning protocol, benchmark datasets, and baseline implementations, we recognize that these details are not presented at the level of granularity required for immediate reproduction. In the revised manuscript we will expand §4 with (i) explicit descriptions of temporal/patient-based splits used for each clinical dataset, (ii) the statistical tests applied (paired Wilcoxon signed-rank tests with Bonferroni correction), (iii) hyperparameter grids and selection procedures for all baselines including replay variants, and (iv) controls confirming that post-hoc explanation methods were not used. We will also release the full experimental configuration files alongside the code. revision: yes

Circularity Check

0 steps flagged

No circularity; derivation is self-contained via novel architecture and external benchmarks

full rationale

The abstract introduces Tree of Concepts as an architectural proposal: a shallow decision tree defines a fixed concept interface, a concept bottleneck model predicts these concepts from raw features, and continual updates are restricted to the extractor and label head. The claimed stronger stability-plasticity trade-off is presented as an empirical outcome on tabular healthcare benchmarks versus baselines (including replay variants). No equations, fitted parameters, or derivations are shown that reduce this outcome to its inputs by construction. The non-drifting explanations follow directly from the design choice of freezing the tree, but this is not framed as a derived prediction or first-principles result that collapses to self-reference. The full text is referenced but contains no load-bearing steps matching the enumerated circularity patterns; the central claim rests on external comparisons rather than internal self-definition or self-citation chains.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim depends on the domain assumption that a fixed concept interface can remain useful without drift; no explicit free parameters or invented physical entities are stated in the abstract.

axioms (1)

domain assumption Concept semantics defined by the initial shallow decision tree remain stable and sufficient across non-stationary clinical data shifts
Invoked to ensure explanations do not drift while the model adapts.

invented entities (1)

Tree of Concepts framework no independent evidence
purpose: Provide an interpretable continual learning interface via fixed concepts
Newly proposed architecture combining decision tree and concept bottleneck components.

pith-pipeline@v0.9.0 · 5489 in / 1324 out tokens · 51010 ms · 2026-05-10T06:40:40.649424+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

Mind the gap: Examining the self-improvement capabilities of large language models.arXiv preprint arXiv:2412.02674, 2024

URLhttps://arxiv.org/abs/2412.02674. Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class-incremental learning. InEuropean conference on computer vision, pages 398–414. Springer, 2022. Zidian Xie, Olga Nikolayeva, Jiebo Luo, and Dongmei Li. Building risk prediction models for type 2 diabetes using mac...

work page arXiv 2022
[2]

arXiv preprint arXiv:2205.15480 , year=

URLhttps://arxiv.org/abs/2205.15480. 14 Tree of Concepts Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class-incremental learning: A survey.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 46(12):9851–9873, 2024. 15 Tree of Concepts Table 6: Cohort transparency summary. We report public source dataset to...

work page arXiv 2024

[1] [1]

Mind the gap: Examining the self-improvement capabilities of large language models.arXiv preprint arXiv:2412.02674, 2024

URLhttps://arxiv.org/abs/2412.02674. Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class-incremental learning. InEuropean conference on computer vision, pages 398–414. Springer, 2022. Zidian Xie, Olga Nikolayeva, Jiebo Luo, and Dongmei Li. Building risk prediction models for type 2 diabetes using mac...

work page arXiv 2022

[2] [2]

arXiv preprint arXiv:2205.15480 , year=

URLhttps://arxiv.org/abs/2205.15480. 14 Tree of Concepts Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class-incremental learning: A survey.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 46(12):9851–9873, 2024. 15 Tree of Concepts Table 6: Cohort transparency summary. We report public source dataset to...

work page arXiv 2024