Timely Clinical Diagnosis through Active Test Selection

We thank the reviewer for their constructive comments. We have addressed the reviewer’s concerns below:

Binary classification and diagnostic probability distributions

Binary classification in a one vs all manner is a common approach in medicine once a differential diagnosis list has been established for test selection (see response to Reviewer tmUM). ACTMED does not assume access to a ground-truth diagnostic probability distribution. Instead, it leverages LLM-generated samples to approximate plausible posteriors under different hypothetical test outcomes. This does mean performance depends on the accuracy of the LLM generated samples and probability estimates (see response to Reviewer tmUM), but we demonstrate that this approach improves downstream classification accuracy. The test that most meaningfully shifts the model’s disease probability estimate is recommended. Final predictions are then evaluated against ground truth labels, allowing comparison of test selection quality across methods.

Comparison to LLM-based imputation pipelines

While pipelines that combine LLM based imputation with classifiers (1) can work well, they typically require task specific labelled training data, risk introducing distributional biases, and do not support sequential, adaptive feature acquisition. ACTMED is designed to operate zero shot and focuses on optimizing which tests to perform rather than replacing clinician driven diagnosis. We have added a more detailed discussion of imputation methods, including those using LLMs, in the related work section.

Clinician-in-the-loop

We have now included a qualitative evaluation with experienced clinicians and senior medical students, who reviewed ACTMED’s reasoning pathways (see response to reviewer tmUM). Clinicians valued ACTMED’s ability to suggest tests when they were uncertain, its intuitive stopping rule, and its supportive nature. As detailed in the study, clinicians can examine the intermediate outputs and assess how the model is changing its risk in response to the lab value changes. Clinicians can intervene at three points: by adjusting the predicted lab values, modifying the assigned risk estimate, or overriding the final test recommendation. In practice, reviewing distributions and risk changes can be time consuming, so most clinicians under time pressure would focus primarily on evaluating the suggested test. After incorporating the clinician’s input, whether it is confirming the suggested test or selecting a different one, ACTMED updates its predictions and proceeds accordingly.

On Dataset Scale and Realism

Scaling ACTMED to very large clinical datasets is challenging due to (a) privacy restrictions on LLM usage and (b) the fact that many datasets only record a single test trajectory per patient, limiting their suitability for sequential testing studies. We address this by dynamically constructing multiple synthetic trajectories from tabular data to generate EHR style notes as tests are selected. We emphasize that the cited datasets contained roughly 100 patients only but this was the size of the test set used in the model evaluation. ACTMED is not subject to overfitting in the conventional sense, as it performs no fine-tuning and does not update its parameters based on the evaluation data. Its zero-shot design ensures it avoids learning dataset-specific shortcuts. In contrast, baselines that rely on full-feature classification are more prone to overfitting, especially when operating on small, fully observed datasets. While the datasets are older, the clinical challenges and measurements remain relevant, as discussed in Section 4.

To strengthen the evaluation, we extended the diabetes analysis to include all 768 patients (Table 1) and added a more complex evaluation on an AgentClinic-derived dataset (see response to Reviewer fFNv). Together, these bring the total number of evaluated patients to over 1,000 and cover a wider range of covariates.

Table 1: Performance of GPT-4o and GPT-4o-mini on the full diabetes dataset (n=768) against the baselines.

References

1. Active Learning with LLMs for Partially Observed and Cost-Aware Scenarios (Astorga et al. NeurIPS 2024)

Thank you for your thoughtful feedback. We appreciate your insights, which have helped improve the quality and clarity of our work. We believe ACTMED provides a practical step toward generalizable, interpretable AI in clinical diagnostics, and we welcome continued dialogue.

We thank the reviewer for their time reviewing our work and their constructive comments.

Choice of datasets

The chosen datasets were selected using strict criteria:

1. Open source (to avoid restrictions on LLM use and ensure reproducibility).
2. Primarily numerical features (as sampling narrow categorical features offers little benefit from LLM reasoning).
3. Well-defined clinical meaning and units, enabling realistic simulation.
4. The datasets intentionally do not include a single definitive diagnostic test (e.g., a PCR test for COVID-19), as such features would trivialize the decision-making task.

While these differ from typical benchmarks such as MedQA or AgentClinic, few open-source datasets meet these criteria and support dynamic, sequential test acquisition. Traditional benchmarks generally lack the necessary structure for sequential testing, as ACTMED is designed for the second step of diagnosis (optimal test selection once differentials are known) rather than the initial generation of differentials (see response to reviewer tmUM).

To address concerns about simplicity, we created a dataset derived from MedQA, selecting 120 patients with sufficient test results. For each, we provide a brief EHR-style summary, including synthetic controls with normal lab values, and compare ACTMED’s performance against baselines. Results (Table 1) show ACTMED achieves performance gains across varied diseases, though differing test panels limit detailed population-level sampling analysis.

Table 1: Performance of ACTMED on the AgentClinic derived dataset against baselines.

Novelty of BED with LLMs

While Bayesian Experimental Design (BED) and KL-divergence–based selection are well-established tools, ACTMED is, to our knowledge, the first framework to integrate BED with large language models as generative priors for clinical reasoning in a zero-shot setting. Unlike prior reinforcement learning or decision tree–based approaches, which demand extensive task-specific training and retraining to accommodate new diseases or tests, ACTMED uses LLMs to generate plausible, patient-specific test outcomes on demand. This enables rapid, adaptive reasoning across diverse tasks and clinical domains without any fine-tuning or reliance on large, labelled datasets—a capability not demonstrated in prior work. By decoupling diagnostic reasoning from model retraining, ACTMED fills a critical gap: enabling sequential, uncertainty-aware diagnostic decisions in contexts where labelled data is scarce or clinical conditions shift rapidly. To our knowledge, the only related work applies BED with entropy gains to coding task disambiguation (1), but not to medicine, and does not employ KL divergence.

Diagnostic test cost modelling

As the reviewer notes, clinical decision-making often involves balancing the diagnostic value of a test against factors such as invasiveness, availability, and cost. In practice, these trade-offs are typically determined by panels of experienced clinicians rather than by automated systems.

In the current study, ACTMED serves solely as a decision-support tool, recommending tests based on their estimated diagnostic benefit alone. For simplicity and comparability, we assumed unit cost for all tests, as the baseline methods used for comparison do not incorporate cost modelling.

Nevertheless, our framework is fully compatible with cost-aware decision-making. Institutions adopting ACTMED could assign test-specific costs, and the model can directly integrate these weights into its test selection procedure.

Clinician involvement

The performance of LLMs against humans in diagnostic settings has been evaluated in prior studies suggesting that experienced clinicians still perform better in complex scenarios (2) and this is discussed in the paper. The authors would like to reemphasize that ACTMED is designed as a test selection tool to support clinicians not as a standalone diagnostic tool. To strengthen this, we conducted a qualitative clinician review of ACTMED’s reasoning pathways (see response to reviewer tmUM), which confirmed its alignment with expert reasoning under uncertainty.

Evaluation of LLM-generated samples

We appreciate the reviewer’s concerns regarding the quality of LLM-generated samples. We emphasize that the purpose of the sampling procedure is not to exactly predict the most likely test result for a given patient but to draw plausible values under both disease and non-disease assumptions. A well-calibrated test will naturally produce roughly half of its simulated values deviating significantly from the true outcome.

The “percentage in reasonable range” metric is intended to detect hallucinations rather than assess per-sample accuracy. The reasonable range is empirically defined based on observed data, and the LLM generates samples that fall within this range for 99% of cases, confirming that outputs remain realistic. The full distributions produced are also shown in Appendix E.

The “minimum distance” metric will, as noted, tend toward zero if the model can generate sufficiently diverse samples and enough draws are taken (Appendix E). Since ACTMED is constrained to 10 draws per query, this metric serves primarily to ensure that the number of samples drawn for the experiment is large enough for enabling reliable BED computations.

Addressing the reviewer’s suggestion, we conducted a distribution-level evaluation of generated versus true samples using Wasserstein and Energy distances (see Table 2). We have also updated Figures 3 and 4 to represent these distribution metrics. The manuscript text has been revised for clarity to reflect these results.

Table 2: Normalized Distribution Matching Metrics (Lower is Better)

Interpretation of model results and stoppage criterion

CKD results, originally included only in the appendix due to near-perfect performance across all methods, are now presented in Figure 5 for completeness. Figure 7 examines why the implicit baseline underperforms relative to ACTMED. For datasets where ACTMED shows clear performance gains (diabetes and hepatitis), we further analyzed feature selections relative to the globally optimal sets identified prior to observing any data.

ACTMED’s stopping criterion is not based on the model’s state but is determined by the KL divergence between successive belief states. For example, with γ = 0.5, the framework will acquire another test only if the expected shift in belief is at least half the distance to the decision boundary. If this threshold is not met, the system deems itself sufficiently confident that no additional test would meaningfully alter the prediction and halts testing.

Clinical experts involved in the review noted that this rule closely mirrors their own decision-making in practice, where additional testing is pursued only when it is likely to significantly change diagnostic confidence.

Comparison of ACTMED and full information baseline

We agree with the reviewer that ACTMED outperforming the full features baseline, which is not limited to 3 diagnostic tests but has access to all information in the dataset may seem surprising. We have conducted an experiment involving classification using sequential increases in the number of features to support this (see response to reviewer pTAQ).

Formatting and minor comments

We standardized mathematical notation (per Reviewer pTAQ’s suggestions) and clarified figure content. Figure 1 now clearly indicates that the distributions shown represent prior and posterior disease probabilities. The green box notation was simplified for clarity. We clarified that “ML” and “DL” in Table 3 refer to “machine learning” and “deep learning,” with examples, and corrected the note on the diabetes dataset: we do not claim BMI is diagnostic, only that OGTT is absent, necessitating indirect metrics. The paragraph on line 354 has been rewritten to better explain how ACTMED supports clinicians.

References

1. Active Task Disambiguation with LLMs (Kobalczyk et al., ICLR 2025)
2. Evaluation and mitigation of the limitations of large language models in clinical decision-making (Hager et al, Nature Medicine 2024)

We greatly appreciate the time and effort you’ve dedicated to reviewing our paper. We hope that the resulting additional experiments will strengthen the contributions of this paper. We’d be happy to engage in further discussions.

Thanks for the detailed responses. No further questions from my side. I'm willing to raise the score to 5.

We thank the reviewer for their helpful feedback, particularly on enhancing the clarity of the mathematical formulation.

Formulation of the BED using KL divergence

We now clearly distinguish between priors and posteriors over disease probabilities versus disease labels, and we clarify that the KL divergence in Eq.​(2) is computed between probability distributions (not between the labels themselves).

1. Epistemic vs. Aleatoric in Active Learning

In Active Learning (common BED framework), we aim to reduce the epistemic uncertainty by acquiring the label of the most informative sample $x_o$, which.

$\underbrace{I(Y;Z\mid, x_o, d)}_{\text{epistemic}} = H(Y\mid d, x_o)- \underbrace{H(Y\mid Z,d, x_o)}\_{\text{aleatoric}} ,$

where $Z$ is a R.V. representing the latent hypothesis (e.g. parameters distribution), $d$ denotes current data, $H(\cdot)$ is entropy and $I(\cdot)$ mutual information.

• Epistemic $=$predictive $-$aleatoric

2. Active Feature Acquisition (AFA)

Our setting is closer to Active Feature Acquisition, where we ask how much a candidate feature $F_j$ reduces predictive uncertainty in $Y$:

$\displaystyle I(Y;F_j)=H(Y)-H(Y\mid F_j).$

Because $H(Y)$ is constant across features, maximizing $I(Y;F_j)$ is equivalent to minimizing $H(Y\mid F_j)$; the most informative test is the one that lowers entropy most.

3. Our KL‑Based Criterion and Its Derivation

When computing expected information gain we may work interchangeably in entropy or KL‑divergence space (1). Equation (2) evaluates

$E[\mathrm{KL}] = -H(Y \mid F_j) + \mathrm{CE}\bigl(Y, Y_{\text{prior}}\bigr).$

where

• Term A ($-H(Y\mid F_j)$) is exactly the uncertainty‑reduction term in AFA.
• Term B is a cross‑entropy that rewards tests which move beliefs away from an uninformative prior, thus avoiding redundant queries.

Note on notation: the prior probability $p_{\text{prior}}$ is constant and no longer carries an index $j$, this has been corrected in the manuscript.

Concrete derivation

For a single Monte‑Carlo sample we compute

$p_{\text{post}}\log \frac{p_{\text{post}}}{p_{\text{prior}}}+ (1-p_{\text{post}})\log \frac{1-p_{\text{post}}}{1-p_{\text{prior}}}$

Re‑arranging,

$= \bigl[p_{\text{post}}\log p_{\text{post}} +(1-p_{\text{post}})\log(1-p_{\text{post}})\bigr]$ $- \bigl[p_{\text{post}}\log p_{\text{prior}} +(1-p_{\text{post}})\log(1-p_{\text{prior}})\bigr]$

Where the term in the first brackets corresponds to $A$ and the second to term $B$. Averaging over $M$ such samples yields the expectation in Eq.​(2). The full derivation has now been included in Appendix B.

Clarity on the utility function notation

First, Equation (1) is now expressed in standard minimisation form as suggested by reviewer fFNv:

$\mathcal{L}(y,u_t,\lambda)=\sum_t \mathbb{I}[y\neq y_{d_t}]+\lambda\sum_t c(u_t)$

Second, to prevent symbol overload we separate the random variable from its realisation:
throughout Section 2 (lines 110–114), $u_t$ denotes the type of diagnostic test (a random variable), while $u_t'$ denotes the observed outcome of that test.

We relax the dual objective with
$\mathcal{F}(u_t^i)=\frac{I(u_t^i)}{c^*(u_t^i)},$

taking $c(u)=\log c^{u}$ to model diminishing marginal penalty. All experiments assume unit costs $(c^*=1)$ so baselines remain directly comparable; real monetary or invasiveness costs can be inserted transparently. Normalizing information gain by cost gives an intuitive “information‑per‑dollar” score and removes the need to hand‑tune a dataset‑specific $\lambda$. The gain per unit cost formulation is also common in medicine (2).

Use of KL Divergence versus Entropy for BED

In Appendix B, we present a theoretical example illustrating how an entropy-based criterion can fail to identify the most informative test, specifically in cases where entropy is high but concentrated on outcomes that provide little diagnostic value. Both entropy and KL divergence can, in principle, be used for test selection (1), and there is no universal proof favoring one over the other; their relative effectiveness depends on the specific setting and data distribution. However, in our experiments, we provide both a theoretical justification and empirical evidence supporting KL divergence for our application. Across all datasets, the results (Table 1) demonstrate that KL-based selection consistently improves diagnostic accuracy and reduces redundant testing relative to entropy-based selection.

Table 1: Performance of ACTMED using KL divergence versus entropy for test selection.

Performance Degradation with All Features

We confirm that ACTMED’s accuracy can decline when too many features are presented at once. Following the reviewer’s suggestion, we added an experiment examining performance as features are added sequentially (Table 2).

The results show that ACTMED avoids this “information overload” by stopping early when further tests offer little diagnostic value and only selecting informative tests. Similar degradation with excessive input has been observed in prior work in different contexts (3, 4).

One contributing factor is that the models tend to assume higher disease risk as more lab results are presented, even when results are normal (Table 3). This bias is especially evident with GPT‑4o mini.

Table 2: Accuracy versus Number of Sequentially Added Tests

Table 3: Impact of Additional Tests on Risk Estimation

LLM Performance within ACTMED

We agree that the performance of ACTMED depends on the underlying LLM’s ability to generate plausible test outcomes. As noted in our response to Reviewer tmUM, a 7B parameter Biomistral model produced unrealistic samples, limiting ACTMED’s gains. While fine‑tuning could improve small models for specific tasks, our results suggest that ACTMED benefits most from larger, general models capable of diverse, accurate sampling.

Test Stopping Criterion

If no test meets the KL divergence threshold, ACTMED stops acquiring tests. This reflects many clinical situations, where a confident diagnosis can be made based on history and presentation alone (often referred to as diagnosed clinically by doctors).

In our experiments, this situation was rare, as each patient was provided with only basic demographic and symptom information initially, so additional testing was typically required. However, in principle it would be possible for the model to diagnose a patient “clinically” without requesting further tests.

Writing and Presentation Improvements

Sections 2 and 3 have been revised to include the more consistent notation. Section 4.1 has also been changed to incorporate the distribution analysis metrics as suggested by Reviewer fFNv. Figures 1 and 2 have been simplified and re-captioned as requested by Reviewer fFNv. Figures 3 and 4 now reflect distributional evaluation metrics.

References

1. Modern Bayesian Experimental Design (Rainforth et al. Statist. Sci. 2024)
2. Cost-effectiveness Thresholds Used by Study Authors, 1990-2021 (Neumann et al. JAMA 2023)
3. Large Language Models Can Be Easily Distracted by Irrelevant Context (Shi et al., ICML 2023)
4. Inverse Scaling in Test-Time Compute (Gema et al., arXiv 2025)

We appreciate the time and effort you have dedicated to reviewing our paper and hope these additions and clarifications address your concerns satisfactorily. We’d be happy to engage in further discussions.

We thank the reviewer for their positive assessment of our work and for recognizing both the intuitive methodology and overall presentation. Below, we address the specific concern regarding potential LLM prediction errors.

Quantifying LLM Uncertainty and Error Rates

We agree that the reliability of the framework depends on the accuracy of the LLM-generated diagnostic predictions. To address this, we now provide two complementary evaluations:

1.Repeated sampling and bootstrapping:

For each patient, we perform 10 independent diagnostic for risk classification and report the mean prediction accuracy with 95% confidence intervals using Bayesian bootstrapping (1). This approach captures the variability of model outputs and allows us to estimate worst-case bounds on prediction accuracy. Table 1 shows the mean, standard deviation, and Bayesian bootstrap confidence intervals for both models across the three datasets with patients stratified by true disease label. Table 2 summarizes the metrics across the entire dataset. A more detailed discussion is included in the Appendix E.

Table 1: Average model prediction accuracy across the hepatitis, diabetes, and CKD datasets (positive and negative cases, both models), with corresponding 95% confidence intervals.

2.Distribution-level validation:

We compute Wasserstein and Energy distances between the generated and true test distributions (see also response to reviewer fFNv), providing a measure of how well the LLM replicates the observed data beyond point predictions. Across all three datasets, ACTMED maintains low distributional distances (Wasserstein ≤ 0.15) and stable performance under repeated sampling. For smaller models that fail to produce realistic distributions, overall accuracy declines and ACTMED does not yield performance gains, as effective Bayesian Experimental Design depends on accurate sampling (see also response to reviewer tmUM).

Strengthening Experimental Validation

In addition, we have incorporated several experiments requested by other reviewers, including:

1. Evaluations using open-source models.
2. Multi-disease diagnostic tasks derived from the AgentClinic dataset.
3. Expanded clinician-in-the-loop evaluations.

References

1. The Bayesian Bootstrap (Donal B. Rubin, The Annals of Statistics, 1981)

We appreciate the time and effort you have dedicated to reviewing our paper and hope these additions and clarifications address your concerns satisfactorily. We’d be happy to engage in further discussions.

We thank the reviewer for their constructive feedback and insightful comments. We appreciate that the reviewer recognizes both the importance and clinical relevance of our work, as well as the detailed evaluation of its performance across three datasets. Below, we address each concern raised.

Experimental evaluation and model choice

As we emphasize in the paper, ACTMED is inherently model-agnostic: it does not rely on any model-specific fine-tuning or internal access, but only on generated distributions of test outcomes and disease risk estimates. Our contribution is a test-time computational improvement — selecting tests to maximize information gain — rather than a modification of the underlying language model. Consequently, the primary consideration is not whether the model is open-source or closed-source, but whether it is capable of generating sufficiently accurate and diverse outcome distributions.

To clarify this point, we reran all experiments using Biomistral-7B, an open-source biomedical LLM. The results (Tables 1 and 2) show that while ACTMED is compatible with Biomistral, its diagnostic performance depends heavily on the quality of the generated distributions. Biomistral’s outputs frequently diverged from real data — for example, predicting non-integer values for the number of prior pregnancies in the diabetes dataset — leading to weaker downstream performance.

Table 1: Wasserstein distance between the true empirical distributions and the model-generated test outcome distributions.

Lower values indicate that the generated samples more closely reflect the real-world data. GPT models produce distributions that closely match the data, while Biomistral-7B shows substantially higher distances, suggesting its generated outcomes are less realistic.

Table 2: Classification accuracy using all features versus ACTMED-selected and sequential implicit-selected features for each model.

ACTMED improves diagnostic accuracy when paired with models that generate realistic outcomes (low distances in Table 1). For models like Biomistral-7B, where outputs diverge significantly from the data, ACTMED cannot provide meaningful gains.

Binary vs. Multi-Label Diagnosis

Clinical diagnosis generally follows a two-step process (1). First, the clinician conducts an initial assessment, forming a primary diagnosis alongside a list of differential diagnoses. Second, tests are ordered to both confirm the leading diagnosis and rule out competing conditions.

This process naturally reduces to a series of one-vs-all evaluations, where each condition is assessed independently. Such an approach is common in clinical practice and machine learning when handling complex multi-class problems. Widely used clinical scores, such as the Wells score for deep vein thrombosis (2) and the CHA₂DS₂ VASc score for atrial fibrillation-related stroke (3), each estimate the likelihood of a single disease. Clinicians routinely apply these sequentially, allowing patients to be assessed for multiple conditions in parallel.

ACTMED addresses the second step by optimizing test selection across diseases, independent of how the differentials are generated. This distinction has been clarified in the revised manuscript, and we include results on a multi-disease dataset derived from AgentClinic (see response to reviewer fFNv), where ACTMED continues to perform strongly.

Handling of Categorical and Free-Text Features

Numerical and categorical variables are converted into natural language descriptions resembling EHR notes (see Appendix D). This ensures that values are presented in a clinically interpretable way, accounting for differences in units and context. At the same time, ACTMED relies on a tabular structure so each feature can be sampled and recombined consistently, allowing us to dynamically generate clinical vignettes and counterfactual diagnostic pathways for the same patient — something not feasible with free-text records alone.

To address the reviewer’s concern, we added an experiment using a dataset derived from AgentClinic that combines structured features with free-text EHR-style notes. ACTMED operates effectively in this mixed-modality setting without modification, demonstrating its flexibility.

Comparison with DRL-Based Approaches

We thank the reviewer for pointing out additional related work. In the revised manuscript, we now cite Ling et al. (4) along with related DRL-based clinical decision-making systems by Yu et al. (5) and Zhong et al. (6). These works demonstrate how reinforcement learning can iteratively guide test selection and decision-making in healthcare, and we view them as complementary to our approach.

ACTMED differs in that it combines Bayesian Experimental Design with LLM-based generative priors to simulate plausible test outcomes at inference time. This allows ACTMED to operate without task-specific retraining, whereas DRL-based methods typically require substantial training to adapt to new tasks.

Clinician-in-the-Loop Evaluation

We conducted a clinician-in-the-loop evaluation with three experienced clinicians and two senior medical students, evaluating 450 test decisions (10 traces across 3 datasets, each assessed by 5 evaluators with 3 decisions per trace). Experts judged ACTMED’s test selections and the resulting risk adjustments as clinically reasonable in 94.5 ± 1.4 % of cases, underscoring its reliability as a decision-support tool.

Clinicians emphasized that they would be reluctant to trust diagnostic suggestions from a black-box LLM without transparent reasoning. ACTMED’s step-by-step decision process was repeatedly praised for building confidence in the model reasoning. Several clinicians noted that ACTMED’s Bayesian decision framework aligns with the reasoning they themselves use when selecting tests under uncertainty, particularly in the absence of clear guidelines.

They also pointed out cases where clinical decision-making diverges from a purely Bayesian approach — for example, when clear guidelines mandate a specific test (such as PCR for COVID 19 diagnosis) or when certain tests are routinely performed despite limited informativeness, such as screening to rule out cancer. However, clinicians agreed that decision support is only needed for the less straightforward cases, where test selection involves genuine uncertainty and no clear protocol exists.

Writing and Presentation

In response to Reviewer pTAQ and others, we have significantly improved the manuscript’s clarity and presentation. Sections 2 and 3 were rewritten with clearer, standardized notation, while Figures 1 and 2 were clarified for readability. Section 4.1 now reports Wasserstein and Energy distances to better characterize the generated sample distributions, addressing requests for more rigorous evaluation. Additional reviewer suggestions have also been incorporated throughout, resulting in a clearer, more consistent, and more informative presentation.

References

1. Improving Diagnosis in Health Care (Committee on Diagnostic Error in Health Care National Academies Press (US) 2015)
2. Derivation of a simple clinical model to categorize patients probability of pulmonary embolism: increasing the model’s utility with the SimpliRED D-dimer (Wells et al. Thrombosis and haemostasis 2000)
3. Validation of risk stratification schemes for predicting stroke and thromboembolism in patients with atrial fibrillation: nationwide cohort study (Olesen et al. BMJ 2011)
4. Learning to Diagnose: Assimilating Clinical Narratives using Deep Reinforcement Learning (Ling et al., IJCNLP 2017)
5. Deep Reinforcement Learning for Cost-Effective Medical Diagnosis (Yu et al. ICLR 2023)
6. Hierarchical reinforcement learning for automatic disease diagnosis (Zhong et al. Bioinformatics 2022)

Thank you for your thoughtful feedback. We appreciate your insights, which have helped improve the quality and clarity of our work. We believe ACTMED provides a practical step toward generalizable, interpretable AI in clinical diagnostics, and we welcome continued dialogue.

Thanks for the detailed responses, very helpful. Few comments:

• For the evaluation and model choice aspect, it would be useful to also consider comparisons with some open-source general purpose models (e.g., Llama) to further clarify the key contributions of the paper based on a diverse set of models, because in its current setting it's still somewhat difficult to understand the usefulness of the proposed methods based on models from two very opposite dimensions (closed source general purpose vs. open source domain-specific models).

• Also, as you acknowledged based on other reviewers' comments that "the performance of ACTMED depends on the underlying LLM’s ability to generate plausible test outcomes.", so I am not quite sure how to interpret the "ACTMED is inherently model-agnostic" claim.