acceptodds
Under review as a conference paper at ICLR 2027

PaST-Q: Learning Future-Queryable Patient States for Tool-Augmented Clinical Agents

Abstract

While clinical guidelines provide population-level recommendations, personalized cancer treatment requires considering both longitudinal patient history and future clinical evolution under candidate therapies. Existing approaches remain limited in either the use of longitudinal disease and treatment context or the richness of future predictions. To address these challenges, we propose PaST-Q, a patient modeling and multi-agent framework that connects historical context with future simulation for guideline-grounded treatment planning. PaST-Q organizes heterogeneous records into temporally ordered Observation-Action blocks and combines hierarchical set encoding with causal temporal modeling to construct a shared patient state. A unified predictor combines Future Query Tokens, a shared backbone, and role-specific experts to directly forecast future observation and action representations at specified horizons. By incorporating candidate treatments into the patient history, PaST-Q predicts patient-specific clinical evolution under each treatment. Through a shared tool interface, a multi-agent system translates these predictions into clinical evidence, compares candidate treatments, and generates treatment recommendations. We evaluate PaST-Q on the public MMRF CoMMpass dataset and a private cervical cancer cohort. Comparisons of patient history integration and future simulation methods, together with clinician evaluation, demonstrate the value of longitudinal context and patient-specific forecasting for treatment recommendation and clinical decision support.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Longitudinal Patient Representation Learning, Predictive Latent Modeling, Tool-Augmented Multi-Agent Systems, Clinical Decision Support

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

PaST-Q is a patient-modeling and multi-agent framework for guideline-grounded cancer treatment planning. Patient records are organized into temporally ordered Observation–Action blocks with unordered events within blocks; role-specific set encoders and a causal Transformer produce a shared patient state S_t. A Future-Query predictor answers requests of the form (role, horizon) directly — a query token combines a role embedding, a time embedding, and the current state, feeds shared cross-attention and role-specific experts, and predicts the future latent of the requested role, with a persistence-plus-residual prediction head. Candidate treatments are appended as Action blocks and the frozen encoder re-computes the state, yielding treatment-conditioned forecasts at 7/30/90 days. Lightweight readouts translate latents into clinical statements; a learned ranker combines a historical treatment-choice prior with predicted-outcome features; a multi-agent layer adds guideline retrieval and synthesizes recommendations. Evaluation covers history-integration and future-simulation ablations on MMRF CoMMpass (public) and a 6,449-patient private cervical-cancer cohort, latent-forecasting and readout benchmarks, a full-system agent benchmark, and two small physician studies. Ethics and AI-use statements are complete.

Critical strengths

S1 — Exemplary evidence discipline for a clinical ML system. Patient-level disjoint splits; a frozen preprocessing pipeline that respects when information actually becomes available (retrospective summaries timestamped at completion); explicit accounting of supervision eligibility per horizon; persistence controls at every evaluation level; component-wise ablations that isolate history integration and future simulation separately; and scrupulously honest reporting of every cell where the method is second best on the public cohort. This protocol makes the paper's measured claims trustworthy, and it is a template worth propagating.

S2 — A reusable architectural abstraction with practical throughput. The future-queryable patient state — encode the history once, then answer arbitrary (role, horizon) queries, including candidate-treatment-conditioned ones, without retraining the encoder for new readout tasks — is a clean interface that other clinical decision-support systems can adopt. The resulting agent completes all 200 benchmark tasks at a mean of 1.76k LLM tokens, which, whatever the comparison caveats, is a practically meaningful cost profile.

Critical weaknesses

W1 — The counterfactual framing exceeds the evidence, and the core validity question is unaddressed. Predicting outcomes under treatments a patient never received, from observational data, is a causal-inference problem; the paper cites the treatment-effect literature (Bica et al.; Melnychuk et al.; Lim) but engages none of its machinery — no identifiability assumptions, no overlap diagnostics, no off-policy evaluation, no negative controls. The deployed metric (agreement with recorded treatment) cannot distinguish a correct dynamics model from a physician-mimicry prior, and Eq. (5)'s log p_choice term is never ablated from the residual. For a system pitched as patient-specific treatment comparison, this silence is consequential.

W2 — Reproducibility and statistical support are below the paper's evidentiary standard. The training objective is prose-only with all four λ weights missing; there is no hyperparameter table, no code release, no MMRF cohort statistics, and no candidate-pool sizes for the retrieval metrics; no confidence intervals anywhere; and the physician studies (18 and 8 assessments) are framed as validation when they are feasibility anecdotes — the largest reported gain rests on three assessments. These are fixable, but they currently block independent verification of the main quantitative claims.

Questions for authors

Q1 (clarification, affects the central claim). What evidence distinguishes the candidate-treatment simulations from confounded observational prediction? Can you provide any proxy validation — e.g., calibration or ranking accuracy restricted to patient subgroups who actually received each candidate, negative-control outcomes, or overlap/positivity diagnostics for the candidate pool?

Q2 (clarification, affects Eq. (5)). How much of the ranking performance is carried by the historical treatment-choice prior log p_choice versus the outcome-feature residual R_ψ? An ablation separating the two terms would clarify what future simulation contributes to the final recommendation.

Q3 (clarification, affects reproducibility). Can you provide the formulas for the five loss terms in Eq. (3), the values of λ_c, λ_Δ, λ_r, λ_o, the target-window size, the training hyperparameters, and a code release for the MMRF pipeline (which uses only public data)?

Q4 (clarification, affects Table 6). How many distinct target candidates enter the retrieval ranking for each role and cohort, and can you report confidence intervals for Tables 1, 2, and 6?

Q5 (clarification, affects Section 4.4). Given 18 and 8 paired assessments, will you reframe the physician studies as feasibility evidence, or expand them? As written, a single changed assessment flips the Q1 comparison.

Q6 (suggestion, future work). A comparison with causal treatment-effect models (e.g., the Causal Transformer line you cite) on overlapping patient subgroups would connect this system to the literature whose problem it adopts.

1 comment
AuthorsOpenReview ↗

Response to Reviewer fqXM

Thank you for valuable feedback! We will address your concerns below and hope they resolve your issues.

Response to W2 and Q3 can be found in general response G1, response to W1 can be found in G2, and response to Q1 can be found in G3.

Q2: Clarification on Eq. (5)

We added a score-term ablation of Eq. (5) on 179 patients from the supplementary Private-CC benchmark. With the model and candidate sets fixed, we compared prior-only, residual-only, and combined scores. Top-1 evaluates treatment-component agreement. The results are shown below.

Scoring ruleTop-1MRRTreatment F1
Prior only52.51%0.651358.44%
Residual only5.03%0.160821.95%
Prior + residual54.19%0.657461.36%

The prior provides the main ranking foundation, and adding the residual raises Top-1 by 1.68 percentage points and F1 by 2.92 points. The residual combines state, candidate, and forecast features to refine this ranking. Table 2 separately evaluates future simulation: enabling it raises F1 from 58.10% to 60.65% on Private-CC and from 72.02% to 72.73% on MMRF. Together, these results show how forecast-informed refinement complements the historical choice prior.

Q4: Confidence intervals.

For Table 6, the full Observation/Action retrieval pools contain 606/108 candidates for MMRF and 18,099/3,123 for Private-CC, shared across methods and horizons. We have added 95% confidence intervals for the PaST-Q results below.

Tables 1 and 2: treatment recommendation (%, estimate [95% CI]).

CohortPrecisionF1Acc@1
Private-CC63.11 [56.51, 69.61]60.65 [54.19, 67.08]55.50 [48.57, 62.22]
MMRF67.72 [61.66, 73.60]72.73 [66.48, 78.53]35.90 [26.15, 46.97]

Table 6: future-latent forecasting (estimate [95% CI]).

MetricPrivate-CCMMRF
RMSE0.508 [0.501, 0.516]0.722 [0.699, 0.747]
MAE0.370 [0.364, 0.376]0.546 [0.528, 0.564]
R@10 (%)6.51 [6.07, 6.97]40.49 [37.53, 43.44]
R@20 (%)10.04 [9.45, 10.66]55.53 [52.40, 58.55]
R@30 (%)12.74 [12.03, 13.45]65.62 [62.38, 68.71]
MRR0.031 [0.029, 0.033]0.201 [0.182, 0.221]

Q5: About physician assessment.

We present both physician studies as feasibility and exploratory evidence. Their value is in proving our method receives more favorable assessments, and could provide additional assistance to the physicians. We will review the paper to ensure that the claims consistently reflect this exploratory scope and avoid overclaiming.

Q5: About suggestions on future work.

Thank you for this constructive suggestion. Comparing with causal treatment-effect models such as Causal Transformer is a valuable next step.

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

PaST-Q encodes longitudinal observations and treatments into patient states that support action-conditioned future queries. Forecasts and clinical readouts are supplied to a tool-augmented agent for treatment recommendation. Evaluation uses the Private-CC cohort of 6,449 patients and MMRF-CoMMpass with 995 patients, alongside representation, forecasting, agent, and small physician studies.

Critical strengths
  • The separation between patient-state encoding, future queries, and clinical readouts provides an interface for integrating longitudinal models with language agents.
  • The evaluation includes two clinical cohorts and tests history modeling, future simulation, and laboratory readouts separately. These analyses provide more information than treatment-recommendation accuracy alone.
Critical weaknesses
  • The evidence does not establish that the learned forecasts support reliable comparison of alternative treatments. Main recommendation metrics measure agreement with recorded treatment choices, and the ranker is trained on those choices. Observational future prediction can also reflect treatment-selection bias. Without addressing this distinction, improved historical agreement cannot establish that forecast-based ranking identifies better treatment options. The additional readout experiments do not close that gap.
  • The evidence for the proposed predictor's practical advantage is limited. In Table 2, Private-CC Acc@1 is 55.5 versus 55.0 for CLARITY-adapted; on MMRF, it is 35.9 versus 37.18. These differences lack uncertainty estimates. The physician comparison contains 18 paired assessments, and the exploratory assistance study contains eight. These results are too limited to establish the claimed decision-support value with confidence.
Questions for authors
  • What assumptions justify comparing forecasts under alternative treatments, and how are treatment-selection bias and candidates with little observational support handled?
  • With the treatment prior, candidate set, and ranker held fixed, how much benefit comes specifically from action-conditioned forecasts rather than historical treatment prediction?
  • Can you report paired patient-level uncertainty for the main comparisons and all physician-assessment outcomes, including disagreement, errors, and unsuccessful assistance cases?
  • For Table 9, what explains the difference between 488.332k tokens for ReAct with the same tools and 1.756k for PaST-Q? Please separate language-model usage from training and tool-computation costs.
1 comment
AuthorsOpenReview ↗

Response to Reviewer KySt

Thank you for valuable feedback! We will address your concerns below and hope they resolve your issues.

Response to W1 can be found in general response G2, and response to Q1 can be found in G3.

W2: About the advantage of our method.

Our evidence addresses both the contribution of future forecasting and the usefulness of presenting these forecasts to physicians. In Table 2, PaST-Q achieves the highest Precision, F1, and Acc@1 on Private-CC and competitive performance on MMRF. Across both cohorts, it improves all three metrics over both “No future simulation” and “No action conditioning,” providing consistent evidence for the contribution of candidate-conditioned forecasting within the evaluated workflow. Private-CC’s rich, heterogeneous longitudinal records may offer particularly useful context for this approach, resulting a better performance on the dataset.

The physician studies were designed as initial evaluations of clinical interpretability and potential usefulness. The exploratory study also identified instances in which the recommendations prompted consideration of previously underconsidered treatment-tolerability factors. Our claim is that PaST-Q offers a feasible forecasting-based workflow with potential complementary value for clinical reasoning. We will revise the claim in the paper.

Q2: About action-conditioned forecasts.

To isolate the contribution of action conditioning, we performed a paired, frozen-system experiment on Private-CC and MMRF. We fixed the encoded state, historical treatment prior, patient-specific candidates, outcome readouts, and downstream selector, without retraining. We replaced only the candidate-action-conditioned future latent with a history-conditioned, action-unconditioned latent shared across candidates. We report the results in the table below, Δ = metric with action-conditioned future latents − metric with action-unconditioned future latents. Action conditioning increased all three point estimates in both cohorts.

DatasetΔ PrecisionΔ F1Δ Acc@1
Private-CC+0.50+0.60+0.50
MMRF+0.75+1.85+1.28

Q3: About physician assessment.

We report all 18 paired physician assessments below.

Assessment itemPaST-Q only acceptable / betterAgentEHR only acceptable / betterBoth acceptable / equalNeither acceptable / both inadequate
Clinical appropriateness and completeness7623
Prediction interpretation and treatment comparison10701
Monitoring and reassessment recommendations8820
Overall decision-support value and specificity10701

Q4: About LLM token usage.

We define LLM usage as the sum of input and output tokens across all generative LLM calls associated with a request, including any LLM-based tool calls and recorded retries. The difference reflects how the two workflows use the language model. ReAct repeatedly submits accumulated tool observations during its reasoning–action loop, whereas PaST-Q executes forecasting, ranking, and evidence assembly through its tool workflow and uses one LLM call to organize the resulting evidence claims. The input/output breakdown is 483.845k/4.487k tokens for ReAct and 1.620k/0.136k for PaST-Q per task. Thus, the difference is primarily attributable to repeated input processing.

Reviewer_cRH6OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work introduces PaST-Q, a framework of modelling and learning from patient histories for enabling clinical decision support by agents. The framework models patient history by blocks of actions and observations, where observations and actions within each block may not ordered. Block information is aggregated in action and observation embeddings which are then subsequently fed into a causal transformer for extracting time-dependent information. The transformer summarizes patient state at time tt with the vector StS_t. This state is used for predicting future observations and actions based on the given time horizon Δt\Delta t. The model is trained by using these predictions of future states by combining five different loss terms.

Agents have access to i) model state StS_t and can use ii) future prediction interface with a time horizon Δt\Delta t as a tool call, finally iii) have access to clinical guides and physician statements (as per line 312) for predictions. Agents use i) and ii) for generating suggestions and ranking, and use iii) for summarizing their findings.

The work is evaluated on two longitudinal observational datasets MMRF-CoMMpass and Private-CC and compare to similar clinical decision-making agents such as AgentEHR, which is the main baseline. They also make use of clinicians for evaluating agent outputs. Some evaluation is done on design choices such as using future predictions for training and block-wise modelling of history.

Critical strengths
  • I was very interested to see evaluation by clinicians. However, this was a rather limited comparison of AgentEHR and the proposed method rather than a more general evaluation of agents in the data (I couldn't find whether the evaluation was done on MMRF or Private-CC). From my understanding, this evaluation cannot be used by future works.
Critical weaknesses

I will separate my comments into minor and major weaknesses.

Minor Points:

  • Line 124: I did not independently check whether the authors' claim is true. However, from the details of the claim, it appears to me that the contributions are minor.
  • Eq (3): over what domain is the loss L defined in (3)? Rather than letting the reader guess, this needs to be formally introduced.
  • Again for Eq (3) each loss term needs clear explanations.
  • Problem Formulation: mathematical notation is unintuitive, to name just one point the choice of index JtJ_t for blocks (line 214) makes it appear as though blocks are not sequentially ordered—which is somewhat misleading.
  • What is relative availability τj\tau_j? (line 215). My guess was that the relevance of a given block for future times, i.e. a block is not used for prediction if τj\tau_j time passes. This needs to be properly introduced and also made clear in the mathematical notation such as in future predictions.
  • A follow up question on previous point: How well does the model handle relative availability of different blocks?
  • It was good to see an ablation of patient history. Authors mentioned the irregularity of clinical records as a problem earlier (line 266). However, an analysis about the impact of this was lacking in the ablation.

Major Points:

  • Authors barely justify any decision-making in the problem formulation. For example, how is the formulation of blocks here different than a Markov decision process (MDP)? It'd be nice to see an exact comparison as a way to understand the modelling of time dependence. This is important as the authors seem to have considered irregular clinical records in their design choice.
  • The work introducing the baseline AgentEHR uses MIMIC data for evaluation. I would have liked to see empirical results on MIMIC for better comparison. This is important as AgentEHR is introduced as the primary baseline.
  • The loss at (3) is very involved with five different terms. Yet there is no ablation study comparing the importance of any of the terms.
  • There is no substantive on the choice of hyperparameters other than a single sentence in the appendix (line 863).

Overall the methodology introduced in this work is not well-motivated, the mathematical notation is imprecise, and empirical evaluation lacks significantly both in breadth (e.g. evaluation on MIMIC data) and depth (e.g. ablation on the loss function).

Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer cRH6 (1/2)

Thank you for valuable feedback! We will address your concerns below and hope they resolve your issues.

Response to Minor2, Minor3, Minor4, Minor5, and Major4 can be found in general response G1.

Minor1: Explanation on our contribution.

Our central contribution is a future-queryable patient state for treatment planning. Prior work establishes the value of longitudinal EHR context for clinical prediction and history-dependent treatment-response forecasting. PaST-Q connects these capabilities through a shared longitudinal state interface: Future Query Tokens specify the target role and horizon, while candidate Action augmentation enables treatment-conditioned forecasting. Extensive experiments demonstrate the effectiveness of our method.

Minor6 & Minor7: About irregularity of clinical records.

By irregularity, we refer to the unequal time intervals between clinical records. Lab examinations, treatments, and follow-up assessments occur at different times according to the patient’s clinical course, producing unevenly spaced observations and interventions, and this is often the case for real EHR data collected from hospitals. Our formulation accommodates this recording pattern through explicit elapsed-time features and horizon-conditioned queries. Window-based future targets additionally allow supervision from records collected around the requested prediction time, without requiring an observation at that exact time.

Constructing a regularly sampled counterpart would require resampling records, imputing unobserved information, or modifying event timestamps. These operations can change the available clinical information and its temporal relationship to treatment, so the resulting comparison would also reflect preprocessing assumptions. We therefore preserve the recorded timelines and evaluate the model under the irregular sampling conditions encountered in the clinical data.

Major1: Our difference with MDP.

PaST-Q is not formulated as a MDP. The differences concern three aspects. First, we encode a history of timestamped Observation–Action blocks without assuming that an individual block is Markov-sufficient. An MDP requires a state that contains sufficient information for the next transition, conditional on the action. Our blocks therefore serve as record-organizing units, rather than Markov states.

Second, we directly predict role-specific content at a requested horizon, P(St,Δt,r)P(S_t,\Delta t,r). An MDP defines a next-state transition kernel, p(sk+1∣sk,ak)p(s_{k+1}\mid s_k,a_k). Our formulation specifies a history-conditioned forecasting query rather than a sequence of decision-step transitions.

Third, we train forecasts against future records and rank treatments against recorded choices. An MDP specifies rewards and a cumulative-return decision objective. Our formulation contains neither a reward function nor return-based policy optimization.

PaST-Q is therefore formulated as a supervised forecasting and treatment-ranking framework, not an MDP.

Major2: About experiments on MIMIC data.

Our dataset choice follows the capability being evaluated: comparing candidate cancer treatment plans using longitudinal histories and forecasts at specified future horizons. MIMIC is often used for the prediction of clinical events, such as diagnoses, tests, and prescriptions. Our task compares candidate cancer treatments using patient-specific, treatment-conditioned forecasts at specified future horizons, and our evaluation requires aligned treatment courses and subsequent disease assessments for regimen-conditioned forecasting. We therefore use public MMRF CoMMpass, which explicitly tracks successive therapies and longitudinal disease assessments. Establishing the same oncology evaluation on MIMIC would require additional cohort construction, endpoint annotation, and follow-up coverage validation. We will clarify this task-driven dataset choice and the adaptation relative to AgentEHR’s original benchmark.

AuthorsOpenReview ↗

Response to Reviewer cRH6 (2/2)

Major3: Ablation on the loss.

We evaluated 7 objectives on each cohort: the full objective, five leave-one-term-out variants, and point-only training, using fixed cohort-specific architectures and validation-selected checkpoints.

Compared with point-only training, the full objective improves MRR by 40.7% on Private-CC and 7.1% on MMRF, with RMSE increases of only 1.4% and 2.1%. Removing point supervision sharply increases RMSE, whereas removing ranking supervision lowers Recall@10 on both cohorts. Removing centering, temporal-change, or occurrence supervision yields smaller MRR decreases. These results support the overall best performance of our design.

ObjectivePrivate-CC RMSE ↓Private-CC R@10 (%) ↑Private-CC MRR ↑MMRF RMSE ↓MMRF R@10 (%) ↑MMRF MRR ↑
Full objective0.50856.510.03100.722440.490.2008
w/o point loss2.89208.390.03841.063738.220.1966
w/o centering loss0.50726.470.02980.715440.450.1998
w/o change loss0.50846.700.03060.715541.030.2003
w/o ranking loss0.50164.790.02280.709638.990.1977
w/o occurrence loss0.50676.460.03050.712039.230.1973
Point only0.50134.640.02200.707838.310.1875
Reviewer_koaNOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PaST-Q, a treatment recommendation framework that combines longitudinal patient modeling with a multi-agent system. The framework forecasts future patient states under candidate treatments and integrates these predictions with retrieved clinical guidelines to generate recommendations. It is evaluated on the public MMRF CoMMpass dataset and a private cervical cancer cohort.

Critical strengths

The patient trajectory model explicitly accounts for irregular time intervals in longitudinal clinical records.

The evaluation includes physician assessments of recommendation quality, extending beyond quantitative treatment-agreement metrics.

Critical weaknesses

Insufficient methodological detail limits assessment of soundness and reproducibility. The framework has two main components: a patient trajectory model learned from observational data and a multi-agent system that integrates its predictions with clinical guidelines. Both require sufficiently detailed descriptions to assess and reproduce. Although Sections 3.2–3.3 and Appendix A.2 provide an architectural overview and some implementation details, important aspects of the training procedure and hyperparameter selection remain unclear. In particular, Eq. (3) lists several loss terms without providing their mathematical definitions or explaining how their weights are selected. The agent implementation in Section 3.4 also remains largely conceptual, leaving unclear how agents coordinate, invoke tools, and combine evidence into final recommendations. A code release would substantially improve reproducibility.

The evaluation does not establish whether the recommendations improve patient outcomes. The main experiments assess agreement with recorded treatments and physician-rated recommendation quality. However, these evaluations do not establish that the model accurately predicts outcomes under alternative treatments or that its recommendations offer benefits over existing clinical practice. Treatment agreement is informative about the ability to reproduce historical decisions, but a model that perfectly imitates those decisions would score optimally without necessarily improving care. This limitation is particularly relevant because the treatment ranker is itself “trained against recorded treatment choices”. Futhermore, the authors should add off-policy evaluation experiments to estimate outcomes under the proposed treatment and compare them with those under existing clinical practice. These experiments are needed to demonstrate the benefit of using the proposed system for treatment recommendation.

Questions for authors

Please provide the mathematical definition of each loss term in Eq. (3), including how the loss weights are selected. Please expand the implementation details of the patient trajectory model, as well as the training procedure and hyperparameter search.

Please evaluate trajectory forecating accuracy and compare with existing patient forecasting method. It seems that Appendix A.3 provide evaluation of latent forecasting. Please elaborate this benchmark. I found it hard to understand the term ‘latent forecasting’. Why not evaluate directly on the observed data but the latent embedding?

Please elaborate on the agent implementation in Section 3.4. For example, the sentence “The execution layer combines rankings, pairwise predicted-outcome differences, …, laboratory changes, and adverse-event risks across horizons.” does not specify how this integration is implemented. Details about agent roles, prompts, tool interfaces, coordination, and decision rules would improve reproducibility, as would a code release.

Please add off-policy evaluation experiments to assess whether the proposed system recommends treatments that improve patient outcomes compared with existing clinical practice or existing methods mentioned in related work. This would help establish the quality of its recommendations beyond agreement with recorded treatments.

Please add explicit references to the relevant appendix sections when discussing methods and evaluations in the main text. The current text rarely directs readers to these sections, making it easy to overlook the additional methodological details and experimental results.

1 comment
AuthorsOpenReview ↗

Response to Reviewer koaN

Thank you for valuable feedback! We will address your concerns below and hope they resolve your issues.

Response to W1, Q1, Q3 and Q5 can be found in general response G1, and response to W2 can be found in G2.

Q2: About latent forecasting.

PaST-Q predicts future patient state by querying a latent representation, which provide a shared representation of heterogeneous clinical observations and interventions for downstream tasks. Accordingly, we have evaluated forecasting at two connected levels.

Appendix A.3.1 evaluates the predicted future representations themselves. We compare each predicted latent with the representation encoded from the corresponding future records, using regression and retrieval metrics. This directly measures the accuracy and discriminability of the model’s forecasting output before task-specific decoding.

However, the latent representation cannot be directly used by the multi-agent system. Thus, we translate the representations into natural language (eg. laboratory changes and trends) through a readout step. This is the evaluation in Appendix A.3.2: the decoded predictions are compared directly with recorded clinical measurements. These results assess the complete forecasting and readout pipeline, whose performance depends on both the future representations and the readout heads. PaST-Q achieves the lowest laboratory-change error and highest trend Macro-F1 among the compared methods on both cohorts.

The two evaluations therefore provide complementary evidence: A.3.1 establishes fidelity to the encoded future clinical content, while A.3.2 demonstrates that the predicted representations retain information that can be translated into clinically interpretable forecasts of observed variables. Together, they connect the accuracy of our model’s native prediction output to its usefulness for clinical forecasting. We will make this relationship explicit in the main text and Appendix A.3.

Q4: About off-policy evaluation.

We conducted model-based off-policy evaluation of our recommendation method. We compare PaST-Q with recorded clinical practice, Prior-only, No-future simulation, and CLARITY-adapted on the same patient states. This is off-policy because clinician-generated historical data are used to estimate outcomes under alternative recommendation policies.

“Comparison policy” denotes either recorded clinical practice (the treatment actually administered) or Prior-only (the candidate treatment with the highest historical selection probability), each serving as a baseline for comparison with PaST-Q. Each policy’s selected treatment is evaluated using a common outcome model. We use the original readout and two separately trained evaluators, MLP and XGBoost. The table reports PaST-Q minus the comparison policy. Negative values indicate lower predicted endpoint values than our method, probability differences are in percentage points.

Comparison policyPredicted endpointOriginal readoutMLP evaluatorXGBoost evaluator
Clinical practicePrivate-CC: 21-day ANC <1.0 event probability−0.60−1.25−0.91
Clinical practiceMMRF: 90-day recorded-death probability−0.77−0.63−1.16
Prior-onlyPrivate-CC: 21-day ANC <1.0 event probability−0.78−0.53−0.49
Prior-onlyMMRF: 90-day M-protein, g/dL−0.106−0.220−0.107

These findings show consistent favorable model-estimated outcomes for PaST-Q on selected clinical endpoints. Their persistence across separately trained evaluators strengthens the outcome-based assessment of our recommendations beyond treatment agreement.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank all reviewers for the constructive feedback! In the general response, we address 3 most common concerns raised by the reviewers. We will also reply to your other questions separately.

G1: Insufficient methodological details.

We have revised our paper for additional information about implementation details.

  • We clarified the problem formulation in Section 3.1 by distinguishing chronological block indices from recording times and explicitly defining the available history, prediction cutoff, and role- and horizon-conditioned forecasting interface.
  • We expanded Appendix A.2 with target construction, the motivation and mathematical definitions of the five forecasting losses, prediction-head details, and cohort-specific training settings. The added workflow description and Figure 6 explain how patient-state tools support candidate generation, simulation, ranking, and reassessment, how their outputs are assembled into evidence claims for language-model ordering and response rendering, and how optional starting-dose refinement reuses this workflow.
  • We provide the evidence-ordering system prompt, identify where it is used in the workflow, and describe the user inputs and output validation.
  • We specify the treatment ranker’s features, residual scoring function, and recorded-choice training objective, clarifying its distinction from the predictor’s latent-target ranking loss.

G2: Concerns on evaluating our recommendations.

We thank the reviewers for highlighting the distinction between agreement with historical treatment decisions and improvement in patient outcomes. We agree that our current evaluation does not fully establish that following PaST-Q recommendations would improve outcomes over existing clinical practice. In our retrospective setting, outcomes are observed only under the treatments actually received, and treatment-selection bias remains a limitation when comparing alternatives.

To provide additional outcome-related evidence, we plan to supplement the evaluation with survival analyses using the available follow-up data. These analyses will examine the association between recommendation–treatment concordance and observed survival.

We will also revise the relevant claims throughout the manuscript to align them with the evidence provided. We position PaST-Q as a framework that integrates longitudinal patient modeling, treatment-conditioned forecasting, and guideline evidence to support clinical decision-making. We do not interpret improved historical agreement as improved patient outcomes.

G3: Causal validity and observational support

PaST-Q incorporates longitudinal history conditioning, a training-derived treatment catalogue with stage/history restrictions, and guideline checks with uncertainty-aware ranking. These designs account for recorded patient differences and constrain candidate selection. Our factual-prediction results, action-conditioning ablation, and clinician evaluations provide complementary evidence for predictive performance and recommendation quality.

To strengthen this evidence, future work will evaluate prediction and calibration within treatment groups, quantify candidate support among comparable patients, and use clinically justified negative controls to investigate residual treatment-selection bias. We will clarify the assumptions underlying candidate comparisons and focus our claims on prediction-informed decision support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.