acceptodds
Under review as a conference paper at ICLR 2027

Physiology as Topology: Intervenable Organ Graphs for Longitudinal Patient Forecasting

Abstract

Do language models forecast patient trajectories by reasoning through physiology, or by fitting surface statistics? We introduce PhysioCircuit, a multi-agent agent system that represents each organ as an agent and each physiological relation as an editable mechanism card with direction, lag, and evidence. Its expert-defined graph is frozen at inference, enabling controlled edge interventions. We also introduce PhysioCircuit-Bench, a mechanism-controlled benchmark spanning three organ systems, treatment events, and real ICU admissions. We evaluate intervention-direction agreement (IDAR), which requires the affected organ to change in the expected direction within the expected lag window. The expert graph substantially outperforms all topology controls on edge-IDAR, including fully connected, random, and reversed graphs. Independent clinician adjudication supports the counterfactual results. PhysioCircuit transfers to new organ systems by configuration alone, without retraining, while preserving inspectable and editable relations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-Agent Systems, Large Language Models, Clinical Reasoning, Predictive Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_VkdvOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PhysioCircuit, a multi-agent framework for predicting how a patient's health condition changes over time. The main idea is to represent each organ as an individual LLM agent and connect these agents through an expert-defined physiological graph. Each connection describes how one organ affects another, including the direction and timing of the effect. These connections can be modified or removed to examine how specific physiological mechanisms influence predicted patient trajectories. The authors also introduce PhysioCircuit-Bench, which includes synthetic patient cases across three physiological systems, treatment interventions, and real ICU patient records. They evaluate both forecasting accuracy and whether the model responds appropriately to changes in physiological mechanisms using a metric called Intervention-Direction Agreement (IDAR). Experimental results show that PhysioCircuit outperforms several baseline methods in prediction accuracy and achieves higher intervention agreement when using the expert-defined physiological graph.

Critical strengths
  1. One of the key strengths of this paper is the introduction of a new evaluation perspective that goes beyond conventional prediction accuracy. By proposing Intervention-Direction Agreement (IDAR), the authors assess whether modifying a physiological mechanism leads to the expected changes in organ states and their timing. This is a meaningful contribution because accurate predictions alone do not necessarily indicate whether a model captures the underlying physiological relationships.

  2. Another important strength is the flexibility of the proposed multi-agent architecture. By representing organs as individual LLM agents and their interactions through editable mechanism graphs, the framework can be adapted to different physiological systems by modifying the graph configuration without retraining the underlying LLM. This makes the approach potentially more extensible and easier to maintain than methods that require model retraining for new physiological systems.

Critical weaknesses
  1. Although the paper introduces a meaningful mechanism-oriented evaluation, most of the supporting evidence comes from synthetic patient trajectories generated using predefined physiological rules. While the authors include real ICU data from MIMIC-IV, the improvements over the strongest baseline (i.e., implicit graph learning) are relatively small (around 0.01 AUROC improvement), and the real-world experiments do not directly validate the causal correctness of the modeled physiological interactions. This raises concerns about whether the strong results on synthetic benchmarks can generalize to complex real-world clinical scenarios.

  2. The proposed IDAR metric is an interesting contribution, but the evaluation uses different intervention mechanisms for different methods. PhysioCircuit is evaluated through direct graph-edge modifications (edge-IDAR), while models without explicit graphs are evaluated through natural-language counterfactual instructions (input-IDAR). Although the authors acknowledge that these scores are not directly comparable, the results are still used to support the superiority of explicit physiological graphs. This makes it difficult to determine how much of the improvement comes from better physiological reasoning versus differences in the intervention interfaces.

Questions for authors
  1. Most of the mechanism-level evaluation relies on synthetic data, while the improvements on real MIMIC-IV data are relatively small. Could the authors clarify whether these improvements are statistically significant and to what extent the real-world experiments support the physiological validity of the proposed framework, rather than only its predictive performance?

  2. The paper evaluates graph-based methods using edge-IDAR and graph-free baselines using input-IDAR. In Appendix A.5.2, the input-level intervention descriptions also appear to explicitly specify the expected direction and timing of the response. Could the authors clarify how this difference in intervention interfaces affects the fairness and interpretation of the comparison? In particular, what evidence supports the claim that the performance advantage comes from explicit physiological modeling rather than differences in how interventions are presented to each method?

  3. Figure 1 and 2 in the paper appear to potentially AI-generated. However, the AI Use Statement only mentions the use of LLM tools for language editing and programming, without specifying whether generative AI was involved in figure creation. Could the authors clarify whether AI tools were used to generate or modify these figures?

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

This paper proposes PhysioCircuit, a multi-agent longitudinal forecasting framework in which each organ is represented by an LLM agent and each directed organ-to-organ relation is encoded as an editable mechanism card containing direction, lag, evidence, counterevidence, and treatment bindings. The paper introduces PhysioCircuit-Bench, consisting of synthetic mechanism-controlled cases, treatment events, counterfactual pairs, and a MIMIC-IV retrospective pillar, and evaluates both predictive accuracy and intervention-direction agreement with timing constraints through IDAR.

Critical strengths
  1. The paper identifies an important limitation of standard clinical trajectory prediction: high predictive accuracy alone does not establish whether a model’s intermediate explanation respects clinically meaningful organ interactions, directions, and time lags.

  2. The explicit mechanism-card abstraction is potentially useful for auditability, because it exposes source organs, target organs, allowed directions, lag ranges, evidence conditions, and treatment bindings in a form that clinicians can inspect and edit.

Critical weaknesses
  1. Could the authors clarify how the central mechanistic claim avoids circularity, given that the truth team creates the executable state-transition generator, the intervention labels, and the expected direction and lag window WiW_i, while PhysioCircuit is evaluated against trajectories generated from the same underlying physiological assumptions? Even if the mechanism-card team is separated from the truth team, the benchmark may still reward agreement with a finite author-designed simulator rather than faithful reasoning about real physiology.

  2. Why should edge-IDAR be interpreted as evidence of causal reasoning rather than compliance with explicit instructions embedded in mechanism cards? The graph edit directly changes a card that specifies source, target, allowed direction, lag, activation evidence, counterevidence, and treatment bindings. Please provide controls in which card semantics are preserved while edge identity is permuted, or in which semantically matched but clinically incorrect cards are used, to isolate whether the model reasons over causal structure rather than following card-level textual cues.

  3. Please provide a matched edge-IDAR baseline for all major graph controls and an input-IDAR version of PhysioCircuit under exactly the same intervention text. For example, an edge-bearing learned graph with the same message schema, a card-permuted graph, and an oracle graph with anonymized edge names would better isolate contributions from graph topology, mechanism-card wording, typed message format, and programmatic gating.

Questions for authors

1.How are the arbitration decisions “accept,” “partial accept,” and “reject” produced, calibrated, and evaluated independently? Since the receiving organ agent chooses the effective update coefficient α\alpha, the framework may encode most of its behavior in the arbitration prompt rather than in the proposed topology. Please report arbitration accuracy against clinician annotations, distributions of α\alpha, acceptance rates by mechanism type, and ablations in which α\alpha is fixed, randomized, or computed by a non-LLM deterministic rule.

2.How clinically valid are the mechanism cards when their expected effects are often context dependent, nonlinear, and multi-pathway? A card contains a single allowed direction and lag range, but many real physiological relations can reverse direction under disease severity, compensatory states, treatment dose, baseline reserve, or co-interventions. Please report how many cards allow bidirectional effects, how context-specific activation conditions are represented, and failure analyses for cases where multiple mechanisms produce competing effects on the same target organ.

3.Could the authors clarify the fairness of the baseline comparison in Table 3, especially regarding computation and access to structured information? PhysioCircuit uses separate organ agents, mechanism cards, typed messages, fourteen deterministic gates, arbitration, and deterministic macro aggregation, consuming 41.741.7–42.4K42.4\mathrm{K} tokens per case, whereas the strongest learned-graph baseline uses 18.918.9–19.2K19.2\mathrm{K} tokens. The “compute-matched, no graph” baseline uses 34.834.8–35.4K35.4\mathrm{K} tokens but may not match the number of agents, prompts, retrieval contexts, or deterministic post-processing steps.

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

The paper presents PhysioCircuit, a multi-agent system in which each organ is an off-the-shelf LLM and organs exchange messages along a clinician-written graph whose edges carry a direction and a lag. Nothing is trained, and the graph is an input. The paper also introduces a benchmark of three synthetic organ systems generated by a rule-based truth generator, plus 600 MIMIC-IV admissions, scored mainly by whether editing a named edge moves the affected organ in the expected direction and time. The expert graph outperforms random and reversed graphs on this metric and leads baselines on direction accuracy.

Critical strengths

S1. The paper proposes a useful test: edit one organ-to-organ edge and check whether the forecast changes the way medicine says it should.

  • Accuracy alone cannot show whether a forecaster used the right mechanism. IDAR checks direction and timing of the response to a named edit.
  • The DA/TA split works as a diagnostic. With all edges reversed, timing stays at 78.4 while direction drops to 33.5, so the system is reading the graph rather than guessing from the patient.

Why this matters to my score: This protocol is the one part of the paper I would reuse.

S2. The authors tried to guard against circular benchmark.

  • The rule generator and the mechanism cards are written by different people, an injection probe tests whether the model copies the generator, and three outside clinicians re-judge all 78 counterfactual pairs.
  • A new organ system is added by rewriting 127 to 154 configuration lines with a 2.6 to 3.8 point loss, where the learned graph loses 7.8 to 17.6.

Why this matters to my score: Credit for attempting the checks. They do not raise my score, because W3 shows each check leaves the circularity in place.

Critical weaknesses

W1. The method is a prompted multi-agent system over a hand-written graph.

  • Each organ is an off-the-shelf LLM. Edges are clinician-written rules with direction and lag. Messages follow the rules and the receiver mixes them in with a fixed coefficient. Nothing is trained. The graph is an input.
  • Table 1 separates PhysioCircuit from prior work only by where rules are stored and whether an edge can be edited.

Why this matters to my score: I do not see a technical contribution. The system is a configuration of existing parts.

W2. The main experiment does not answer the opening question.

  • The abstract asks whether the LLM forecasts by reasoning through physiology or by pattern matching on the text. Every experiment holds the LLM fixed and varies the graph, so nothing is attributed to the LLM.
  • Messages can only travel along existing edges, so the gap between the correct graph and the random or reversed graphs follows from the routing. A rule script with no LLM would give the same ordering.

Why this matters to my score: The edge-IDAR gap shows the system executes the graph it is given. The question in the abstract is very confusing. I am not asking for additional experiment without LLMs which is not reasonable. I am asking if the logic flow and the writing could be refined.

W3. The answer key and the system under test come from the same source.

  • The rule generator produces the patients, the gold labels, and the expected response to every intervention. The mechanism cards the system reads are written from the same textbooks. The benchmark asks whether one copy of the rules agrees with another copy of the rules.
  • The three defenses do not close this gap. The below-30% overlap is a rater estimate with no stated metric. The 18-rule injection probe shows the model does not copy the generator's text, not that the benchmark separates mechanism reasoning from executing the cards. The external clinicians reach κ 0.51 and majority agreement on only 63% of conflict pairs, and conflict pairs are 40% of the test set by design.

Why this matters to my score: Where the paper says arbitration matters most, there is no agreed answer key, so a high edge-IDAR there cannot be interpreted.

W4. The presentation is rough.

  • Across three organ systems, gains over each baseline agree to within 0.2 points. Per-run numbers are needed. Secs. 5.3 and 5.4 repeat the same values.
  • Many paragraphs are only a few words long, and Figure 1 is drawn at an odd aspect ratio.

Why this matters to my score: Definitely not good enough for human reading.

Questions for authors
  1. With the correct graph fixed, what is edge-IDAR when a deterministic rule engine replaces the LLM, and when card text is scrambled across edges?
  2. How is baseline L trained, given no fine-tuning and 20 development cases per system? Please also give per-run accuracies for Table 3.
  3. On MIMIC-IV, how are per-organ directions and macro phases derived from raw records, and is direction consistency ever compared with the observed outcome rather than the card expectation?
Reviewer_bPt2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PhysioCircuit for predicting the direction of change in each organ and the patient's overall disease course from longitudinal clinical records. The method represents organs as LLM agents, with expert-written mechanism cards forming a directed graph; each mechanism card specifies the direction of effect, time delay, and supporting evidence. Structured messages between organs are updated through rule checking and arbitration before predictions are made, and graph edges can also be edited to test the model's intervention responses. The paper evaluates the method on three synthetic physiological systems and 600 real ICU records, reporting prediction results superior to several baselines, as well as intervention consistency in which correct graphs outperform incorrect topologies.

Critical strengths

1.The research question of this paper is valuable: even if patient disease course predictions are accurate, they may rely on incorrect organ interactions. This paper attempts to test the model's response to changes in specific physiological mechanisms, making such errors—which are difficult to detect based on prediction accuracy alone—detectable.

2.The experimental evaluation in this paper is relatively comprehensive, covering three synthetic physiological systems, topology controls, component ablations, cross-model tests, and 600 real ICU records. In all three synthetic systems, the prediction accuracy is superior to the compared baselines, and the intervention consistency of the correct mechanism graph is significantly higher than that of incorrect topologies.

Critical weaknesses
  1. The evidence for mechanism faithfulness in this paper is not yet sufficient. edge-IDAR primarily relies on synthetic data constructed by the authors as the basis for scoring. That the correct graph outperforms incorrect graphs shows that the model responds to graph structure, but it does not yet show that these responses conform to the physiological processes of real patients. Clinicians independently judged 78 counterfactual pairs; however, in the more critical conflict cases, only 63.3% reached a majority consensus on direction and time delay.

  2. The two types of intervention results should not be used to directly compare the strengths and weaknesses of methods: explicit graph methods accept interventions by modifying graph edges, whereas graph-free baselines receive textual descriptions. The paper also points out that the scores of edge-IDAR and input-IDAR cannot be directly compared, so the current results are insufficient to determine how much of the performance gap comes from the graph structure itself.

3.The transfer experiment section needs clearer explanation. For example, Figure 5 labels the off-diagonal results as zero-shot reuse, while the main text describes 87.5% and 88.7% of them as results after rewriting the configuration. In addition, the digestive system shares no organs with the other two systems, so the paper needs to explain how the original graph is used in the target system; otherwise, it is difficult to judge the actual effect of "transfer without retraining."

Questions for authors

1.Among the 78 counterfactual pairs, how do the 17 cases on which clinicians did not reach consensus correspond to and affect the overall edge-IDAR? If the conflict cases on which consensus was reached are evaluated separately, is the model's performance still consistent with the overall conclusion?

2.The paper demonstrates through same-channel comparisons between the correct graph and incorrect topologies that the topology setting affects the results. However, when comparing with the prompt-only method, the intervention channel, structured messages, and arbitration mechanism also change simultaneously. How do the authors determine how much of this performance gap comes from the explicit graph structure itself?

3.Are the 87.5% and 88.7% in Figure 5 the results of zero-shot reuse or of rewritten configurations? When the source system and the target system share no common organs, how are the nodes and edges of the original graph used for prediction?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.