acceptodds
Under review as a conference paper at ICLR 2027

ClinAdapt: Self-Adaptive Multi-Agent Learning for Longitudinal Clinical Event Modeling in Lung Cancer Detection

Abstract

Longitudinal clinical event modeling from electronic health records (EHRs) requires reasoning over sparse, noisy, and long-context multimodal sequences. Existing LLM-based multi-agent systems alleviate context-length constraints but process each patient in isolation, failing to emulate how clinicians exploit accumulated experience from similar prior cases. We introduce ClinAdapt, a self-adaptive multi-agent system with two complementary adaptation mechanisms. First, an Adaptive Experience Memory (AEM) serves as non-parametric memory, indexing rejection-sampled reasoning traces to retrieve similar patients as few-shot contexts. Second, Multi-Agent Policy Adaptation (MAPA) with reward-ranked fine-tuning parametrically optimizes inter-agent and agent-memory collaboration. A leave-one-out cross-retrieval strategy unifies these mechanisms, aligning training- and inference-time behavior under retrieval augmentation. On lung cancer prediction using up to five years of multimodal EHRs, ClinAdapt outperforms nine strong baselines on both the overall population and a challenging never-smoker population. Analyses of the adaptation dynamics reveal three key findings. First, as the AEM expands, the optimal retrieval strategy shifts from diverse to specific samples. Second, under MAPA, the manager agent’s prediction loss converges rapidly, while worker agents’ temporal reasoning continues to benefit from more verified patients. Third, the two mechanisms are complementary in predicted risk, with AEM improving specificity and MAPA improving sensitivity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Medical Image Analysis, Multi-Agent Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Sc8iOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper extends a chain of LLM agents over chunked EHR, with two additions. One is a memory of the best reasoning traces from training patients, retrieved as few-shot examples for similar patients. The other fine-tunes the worker and manager agents on rollouts that agree with the label. Task is one-year lung cancer prediction from up to five years of single-centre EHR, with an overall and a never-smoker test set.

Critical strengths
  • never-smoker early detection is maybe important (How accurate is self-reported status? I also never smoked of course...)
  • leave-one-out retrieval during training makes sense
  • detailed cohort table
Critical weaknesses
  • the stream is random training samples added over time, no time order. how is that self-adaptive?
  • only 27 cases in the never-smoker test set. bootstrap SE only, any paired test planned?
  • index date is the chest radiology exam. is the report of that exam in the input? then the model just reads the finding
  • 1:10 case-control matching, hence PPV and F1 don't mean much outside this setup
  • LCRAT is made for ever-smokers, not a fair baseline on never-smokers
  • MAPA is basically RAFT style fine-tuning on two agents, why call it MARL in the figures?
  • you store reasoning traces of real patients and show them for other patients. data governance?
  • what is k? how many GPU hours?
Questions for authors

see above

Reviewer_F12AOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces ClinAdapt, extending the Traj-CoA algorithm, to alleviate shortcomings of current LLM methods in utilising electronic health records for early lung cancer prediction. The paper introduces two algorithmic novelties, i) AEM for retrieving information regarding similar patients, and ii) MAPA, a framework to optimise inter-agent and memory collaborations. A leave-one-out retrieval procedure integrates these components during training.

Critical strengths

The paper motivates and addresses a clinically significant and challenging problem. EHRs contain heterogeneous and irregularly sampled data points, making temporal reasoning difficult. The proposing methodology is well motivated, and tackles the challenge of LLM-based architectures that treat patients as isolated, and do not consider previously encountered patients, the the say way a human clinician would.

Evaluation of their methodology on a non-smokers cohort is also clinically valuable, as this population presented different diagnostic challenges from traditional smoker based assessment.

Critical weaknesses
  1. Major Weakness: The absence of quantitive experiments. The abstract and conclusions make claims regarding the performance of the proposed methodology and AEM and MAPA modules. These are however not substantiated by any quantitative or qualitative results. There is therefore no way to tell if the proposed methodology improves over Traj-CoA, and whether each added module is beneficial in isolation or combined (no ablation study).

  2. The paper states a random train/test split of its data. This should be clarified by the authors - a temporally separated evaluation would ensure testing of generalisation to future time / predicted time, and reduce the possibility of optimistic estimates arising from similarities between train and test distributions.

  3. Evaluation concerns. The prediction index is defined as the completion of chest radiology examination, however Table 1 includes index examinations of other body sites. Further clarification is required to resolve this inconsistency. Additionally, the paper also reports 27 positive cases and 808 controls in the never-smoker test set, despite describing 1:10 case-control matching. This sampling protocol needs clarification.

  4. Reproducibility. Several implementation choices and hyper parameters are not declared. For example, the retrieval count k, and training hyperparameters.

  5. The paper would also benefit from a discussion about runtime. Without this analysis, the practical benefits over simpler alternatives remains unclear.

Questions for authors

Experimental results: The abstract and conclusion claim state-of-the-art performance and complementary benefits from AEM and MAPA, but Section 4 does not contain quantitative results. Are these results missing from the submitted version? Could the authors provide the main comparisons, uncertainty estimates, and individual component ablations?

Validity of reasoning traces: MAPA selects training traces based on whether their final predicted scores agree with the binary ground-truth outcome. Since correct predictions do not necessarily imply clinically valid reasoning, what evidence supports the claim that this procedure improves temporal reasoning and inter-agent collaboration, rather than merely reinforcing successful final predictions?

Evaluation protocol: Given the clinical setting and the emphasis on continual adaptation, how do the authors ensure that random patient-level splitting and retrieval from verified cases do not introduce temporal information leakage? Has the reported performance been assessed under a temporally separated evaluation protocol?

Reviewer_pb5BOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

I used AI to help polish my review comments.

Critical strengths
  1. The paper addresses a meaningful and clinically relevant problem in longitudinal clinical prediction. Incorporating accumulated experience from historical patient cases into multi-agent reasoning is a promising research direction.

  2. The proposed framework provides a systematic design that integrates experience memory with multi-agent policy adaptation, offering a unified approach to improving clinical prediction.

Critical weaknesses
  1. Limited methodological novelty. The paper introduces Adaptive Experience Memory (AEM) and Multi-Agent Policy Adaptation (MAPA). However, their core designs largely rely on existing techniques, including similar-patient retrieval, reward-guided trajectory selection, and supervised fine-tuning. The authors do not sufficiently clarify the distinctive contributions of this combination compared with existing methods, nor do they provide an in-depth justification for the key design choices and their necessity.

  2. Insufficient ablation studies. The paper lacks adequate ablation experiments to isolate the contributions of AEM, MAPA, and their combination. Consequently, it remains unclear whether each component provides the expected improvements and whether the two mechanisms offer meaningful complementary benefits.

  3. Insufficient details on dataset construction. Although the paper describes the prediction time window and case-control matching strategy, important details remain unclear, including patient inclusion and exclusion criteria, the procedures used to establish outcome labels, and the specific matching process. Providing a more comprehensive description would improve the transparency and reproducibility of the experiments.

  4. Minor presentation issue. On page 7, the expression "Eq. equation 10" contains a redundant equation reference. The authors are encouraged to check the consistency of equation references throughout the manuscript.

Questions for authors

See my review comments in weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.