SAGEAgent: A Self-Evolving Agent for Cost-Aware Modality Acquisition in Multimodal Survival Prediction

Bennett A. Landman; Can Cui; Chongyu Qu; Juming Xiong; Junchao Zhu; Junlin Guo; Tianyuan Yao; Yanfan Zhu; Yuankai Huo; Yuechen Yang

An LLM agent can stop the glioma diagnostic sequence early for many patients, cutting average clinical burden by 55% while keeping survival prediction nearly as accurate as a full workup.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 02:25 UTC pith:ASJ6MS55

load-bearing objection Clean methods paper: order-respecting LLM acquisition with dual memory keeps C-index near full fusion while cutting MCDA burden ~55%; the savings number is only as solid as the hand-set cost weights that also shape the rewards. the 3 major comments →

arxiv 2607.09521 v1 pith:ASJ6MS55 submitted 2026-07-10 cs.AI

SAGEAgent: A Self-Evolving Agent for Cost-Aware Modality Acquisition in Multimodal Survival Prediction

Chongyu Qu , Can Cui , Zhengyi Lu , Junchao Zhu , Tianyuan Yao , Junlin Guo , Juming Xiong , Yanfan Zhu

show 3 more authors

Yuechen Yang Bennett A. Landman Yuankai Huo

This is my paper

classification cs.AI

keywords modality acquisitionLLM agentsurvival predictionmultimodal oncologycost-aware diagnosticsgliomaself-evolving agentsequential decision

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Not every cancer patient needs every diagnostic step for a reliable survival estimate. In glioma care the modalities arrive in a fixed order of rising burden—demographics, then radiology, then pathology, then genomics—yet most multimodal models either demand the full set or treat missing data as a passive nuisance. SAGEAgent reframes the problem as a sequential stop-or-acquire decision: at each stage a frozen language model, guided by tools that turn risk and uncertainty into text, memories of similar past cases, and rules it discovers from its own outcomes, decides whether the next invasive test is worth the cost for that patient. On a combined TCGA–BraTS cohort the agent holds concordance within 0.012 of the full-modality backbone while saving 55% of the average burden. A sympathetic reader cares because biopsy and genomics are expensive, slow, and burdensome; if selective acquisition works, many patients can be spared those steps without sacrificing prognostic quality.

Core claim

SAGEAgent formulates cost-aware modality acquisition for multimodal survival prediction as a sequential decision problem that respects the clinically mandated order. Using clinical tools that verbalize risk and calibrated uncertainty, episodic memory of similar patients and rewarded past episodes, and semantic memory of self-discovered decision rules refined by periodic reflection, the training-free LLM agent achieves a C-index of 0.813 (within 0.012 of the full-modality backbone) while reducing average acquisition burden to 0.451—a 55% saving—on 170 complete-modality glioma patients under nested 5×5 cross-validation.

What carries the argument

SAGEAgent: a frozen LLM that, at each stage of the ordered diagnostic sequence, synthesizes three text-space signals—clinical tools converting numerical risk and uncertainty into language, episodic memory retrieving similar cases and past rewarded decisions, and semantic memory of rules that self-evolve through reflection on good versus poor outcomes—to output an auditable acquire-or-stop choice without gradient updates.

Load-bearing premise

The claimed 55% savings rest on fixed burden weights for each modality derived from multi-criteria decision analysis; if real money, time, invasiveness, or infrastructure costs differ by site or patient, the stop policy and the headline reduction no longer map to clinical cost.

What would settle it

On an independent multi-site glioma cohort evaluated with the same nested protocol and the paper’s own burden weights, the agent’s majority-vote policy either drops C-index more than about 0.02 below the full-modality backbone or fails to reduce average burden by roughly 40% or more.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

A substantial fraction of glioma patients may not need invasive pathology for reliable prognostication, because the agent learns to skip it for roughly half the cohort.
Each acquisition decision is grounded in explicit chain-of-thought reasoning, making cost-aware stopping clinically auditable rather than a black-box policy.
Stopping rules discovered by self-reflection converge stably across folds and outperform chance, showing that reusable decision patterns can be learned without hand-crafted priors.
Combining episodic and semantic memory yields a markedly better accuracy–cost trade-off (higher hypervolume) than tools or either memory alone.

Where Pith is reading between the lines

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

The same ordered, memory-guided stop-or-acquire pattern could transfer to other staged diagnostic pathways such as breast or lung cancer workups.
If burden weights were made patient- or site-specific, the agent would need an online cost model to keep the claimed savings clinically meaningful.
Prospective studies could test whether the agent’s early-stop decisions align with clinician judgment and with actual patient outcomes beyond retrospective concordance.
Acquisition rules currently churn more than stopping rules; larger multi-site experience pools may be required before both rule types stabilize equally.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Clean methods paper: order-respecting LLM acquisition with dual memory keeps C-index near full fusion while cutting MCDA burden ~55%; the savings number is only as solid as the hand-set cost weights that also shape the rewards.

read the letter

Punchline: this is a real sequential-acquisition agent for multimodal survival that respects the clinical order (demo→rad→path→gen), keeps ranking accuracy close to full fusion, and actually skips expensive steps for many patients. The 55% burden cut is the headline, but it is defined by fixed MCDA scalars that also enter the episode rewards, so treat that number as conditional on those weights.

What is new is the combination, not any single piece. RL acquisition and LLM sequential diagnosis already exist. SAGEAgent freezes a multimodal Cox backbone, adds a calibrated head that predicts whether more modalities would change the risk score, and lets a frozen LLM decide acquire/stop using three text signals: tool translations of risk/uncertainty, FAISS episodic neighbors, and semantic rules mined by periodic reflection. No gradient updates on the agent. Nested 5×5 CV on 170 complete-modality glioma cases, majority vote, static fusion + RL + Reflexion baselines, component ablations, and a rule-stability table are all there. C-index 0.813 vs 0.825 for the full Self-Att backbone at burden 0.451 is a credible accuracy–cost point under their protocol. Code is linked. The dual-memory self-evolution and the explicit clinical ordering are the parts I would actually remember.

Soft spots, in proportion. The stress-test lands: burdens (0.03/0.14/0.53/0.30) are hand-normalized MCDA and appear inside R_stage and R_term, so the discovered stopping rules and the reported savings/hypervolume are optimized for that cost vector. Nested CV does not probe alternative weights. Evaluation is only the complete-modality subset; partial-modality patients are augmentation for the predictor, not the agent test set. Free knobs (τ, k, reflection cadence, reward mix, temperature scaling) are real but secondary. None of this breaks the internal ML claim; it bounds how far you can push the clinical interpretation (“skip pathology for half the cohort”).

Who it is for: people building multimodal oncology models who care about ordered acquisition and audit trails, not pure survival-methodologists chasing another 0.01 C-index. Math and citations look fine; no circular evaluation metric. I would send it to peer review. Worth a reading-group slot if your group does clinical multimodal agents; otherwise skim the ablations and Table 3.

Referee Report

3 major / 0 minor

Summary. The paper formulates cost-aware diagnostic modality acquisition for multimodal survival prediction as a sequential decision problem that respects a clinically mandated order (demographics → radiology → pathology → genomics). It proposes SAGEAgent, a training-free LLM agent that at each stage chooses acquire or stop using three text-space signals: clinical tools that verbalize a frozen multimodal Cox predictor’s risk and a calibrated uncertainty head, episodic memory of similar patients and past episodes, and semantic memory of decision rules refined by periodic reflection on episode rewards. On a glioma cohort (TCGA-LGG/GBM + BraTS), nested 5×5 CV over 170 complete-modality patients reports C-index 0.813 ± 0.046 (within ~0.012 of the full-modality Self-Att. Fusion backbone) at average burden 0.451 (55% saved vs full workup B=1.0), with ablations attributing burden reduction to dual memory and a self-evolution analysis of discovered stopping/acquisition rules.

Significance. If the accuracy–burden trade-off holds under realistic cost variation and broader cohorts, the work is a meaningful step beyond static multimodal fusion and black-box RL acquisition: it couples clinical ordering, transparent chain-of-thought decisions, and experience-driven rule discovery without gradient updates to the LLM. Strengths include nested CV with majority-vote aggregation, comparison to static fusion, RL, and Reflexion baselines, component ablations (Table 2), rule-stability analysis across folds (Table 3), public code, and an explicit uncertainty head with reported calibration (AUROC 0.908, ECE 0.110). The framing that not every patient needs full workup is clinically relevant and under-explored in multimodal survival AI.

major comments (3)

Sec. 3.1 and the reward definitions in Sec. 2.3: the headline 55% burden reduction and HV trade-off treat fixed MCDA scalars (Demo 0.03, Rad 0.14, Path 0.53, Gen 0.30) as ground truth, and the same b(m) enter R_stage and R_term that label episodes for semantic reflection. Table 3 stopping rules and the majority-vote policy are therefore optimized for this cost vector. Without sensitivity analysis over plausible reweightings of monetary cost, time, invasiveness, and infrastructure (or site-specific costs), the quantitative burden claim and the clinical interpretation that pathology can be skipped for half the cohort do not generalize. Please add sensitivity sweeps and report how C-index, burden, and rule stability change under alternative weight vectors.
Sec. 3.1 evaluation protocol: all agent metrics are computed only on the 170 complete-modality patients, while 792 partial-modality cases are used solely as predictor augmentation. Complete-modality patients are a selected subset; early-stop decisions and the skip-pathology claim may not transfer to the broader incomplete population that the clinical workflow actually faces. Please either evaluate acquisition policies under realistic missingness patterns (e.g., prefixes that stop before unavailable modalities) or clearly bound the claim to complete-workup candidates and discuss selection bias.
Sec. 2.2–2.3 and Table 1: the uncertainty head is trained with labels from full-vs-prefix risk differences on complete patients (threshold τ=0.3), then exposed as a primary tool signal. While C-index remains an external metric, the agent’s stop decisions are tightly coupled to a signal defined with knowledge of the full-modality outcome. Please report (i) agent performance when the uncertainty tool is ablated or replaced by a non-oracle uncertainty estimate, and (ii) whether RL baselines were given the same calibrated uncertainty and clinical-order constraints; the large C-index gap (~0.07) may partly reflect unequal information rather than method class alone.

Circularity Check

1 steps flagged

No load-bearing circularity: C-index and burden are external empirical outcomes; mild self-use of authors' prior fusion backbone and of reward-defined episode labels does not force the 55% claim by construction.

specific steps

self citation load bearing [Sec. 2.2 / Table 1 (Self-Att. Fusion† backbone; cite [8] MMD)]
"The prediction backbone is a Transformer-based multimodal encoder that handles missing modalities via learnable mask tokens [9,8]. ... Self-Att. Fusion† 0.825±.050 1.000 (0%) ... †=predictor backbone used by SAGEAgent."

The frozen multimodal survival encoder is taken from the authors' prior MMD/Self-Att. line [8] (overlapping authors Cui/Huo). This is ordinary reuse of a backbone, not a uniqueness or derivation claim: C-index of the backbone is re-reported here, and the paper's central claim is the acquisition policy on top, not a re-derivation of fusion accuracy. Flagged only as mild non-load-bearing self-citation; does not force the 55% burden result.

full rationale

SAGEAgent is an empirical systems paper, not a first-principles derivation. The headline result (C-index 0.813 at burden 0.451, 55% saved) is measured on held-out complete-modality patients under nested 5×5 CV against actual survival labels and a fixed MCDA cost vector; neither quantity is algebraically identical to the agent's inputs. The uncertainty head is trained on full-vs-prefix risk differences (Eq. 1, τ=0.3) and then exposed only as a text tool signal—not as the evaluation metric—so it is an intermediate feature, not a self-definitional prediction. Episode rewards R_stage and R_term reuse the same burden scalars that appear in the reported average burden; that makes the policy cost-vector-specific (a robustness concern), but the 55% figure remains an empirical count of which modalities the agent stopped after, not a quantity forced by definition. Semantic rules are mined from training-split episode rewards and applied at test time under nested CV—standard experience-based learning, not circular renaming. The frozen Self-Att./MMD-style backbone cites overlapping prior work [8], but its C-index is re-measured in this paper and the acquisition policy is the novel claim; that self-citation is not load-bearing for the 55% result. No uniqueness theorem, ansatz smuggled via citation, or fitted parameter re-labeled as an independent prediction appears in the derivation chain. Score 1 only for the mild, non-forcing self-use of the authors' prior backbone and of reward-defined logs.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 4 invented entities

The central accuracy–burden claim rests on a frozen multimodal Cox predictor, a supervised uncertainty head, fixed MCDA burden weights, hand-set agent hyperparameters, and the premise that LLM chain-of-thought over tool text plus dual memory yields clinically ordered stop/acquire decisions. No new physical entity is postulated; the invented pieces are algorithmic (agent, memories, rewards, uncertainty head).

free parameters (5)

uncertainty change threshold τ
Binary label for the uncertainty head is 1 when |r_current − r_full| > τ; paper sets τ=0.3 by calibration quality on inner validation.
MCDA modality burdens
Normalized burdens 0.03/0.14/0.53/0.30 define total B and the 55% savings headline; chosen via multi-criteria scoring, not learned from outcomes.
episodic retrieval k and reflection cadence
k=3 neighbors, reflection every 10 patients, up to 10 active rules, 3 episodes per training patient control memory and self-evolution behavior.
stage and terminal reward coefficients
R_stage = −b(m_{t+1}) + max(0, u_t − u_{t+1}) and R_term = (1−u_t) − B_t define what counts as a good episode for retrieval and rule mining.
post-hoc temperature scaling for uncertainty
Calibration (AUROC 0.908, ECE 0.110) uses temperature scaling on validation; affects tool text the LLM sees.

axioms (5)

domain assumption Diagnostic modalities must be considered in a fixed clinical order (demographics → radiology → pathology → genomics) with binary acquire-or-stop decisions at each stage.
Problem formulation Sec. 2.1; ordering is taken from clinical practice citations rather than learned.
domain assumption A scalar MCDA burden in [0,1] per modality is an adequate proxy for monetary cost, time, invasiveness, and infrastructure.
Sec. 3.1; all burden and HV results depend on this scalarization.
domain assumption Cox proportional hazards on a Transformer fusion embedding is a valid survival ranking model under modality dropout and mask tokens.
Sec. 2.2; backbone is frozen and supplies risk scores for tools and C-index.
ad hoc to paper A frozen instruction-tuned LLM can make reliable acquire/stop decisions from natural-language tool outputs plus retrieved cases and rules without gradient updates.
Core of SAGEAgent Sec. 2.3; not independently proven outside this empirical setup.
domain assumption Episodic vs semantic memory from cognitive science is an appropriate organizing principle for case retrieval and rule accumulation.
Sec. 2.3 citing Tulving; structures the dual-memory design.

invented entities (4)

SAGEAgent sequential acquisition agent no independent evidence
purpose: Unify clinical tools, episodic memory, and semantic self-evolution to decide modality acquisition under clinical order.
Primary proposed system; evaluated only inside this paper’s glioma protocol.
Calibrated uncertainty head for future modality value no independent evidence
purpose: Predict whether acquiring more modalities would change the risk score enough to matter.
Eq. 1; trained on prefix vs full risk differences, not an external clinical endpoint.
Stage and terminal reward definitions for episode labeling no independent evidence
purpose: Score past decisions for retrieval ranking and rule effectiveness.
Sec. 2.3; internal to the agent’s self-evolution loop.
Semantic decision rules discovered by periodic reflection no independent evidence
purpose: Store reusable stop/acquire patterns indexed by stage and direction.
Table 3; rules are mined from the same training episodes they later guide.

pith-pipeline@v1.1.0-grok45 · 12839 in / 3636 out tokens · 48314 ms · 2026-07-13T02:25:04.442769+00:00 · methodology

0 comments

read the original abstract

Does every cancer patient truly need a complete diagnostic workup for accurate survival prediction? In multimodal clinical oncology, diagnostic modalities follow a clinically mandated order of escalating burden -- from demographics collected at intake to genomic profiling requiring specialized tissue analysis. Current multimodal survival methods either assume all modalities are available or passively handle missing data, but none actively reason about whether acquiring the next modality is justified for a given patient along this ordered workflow. We formulate this as a sequential decision problem and propose SAGEAgent (Sequential Acquisition Guided by Experience), a self-evolving LLM-based clinical agent that decides which diagnostic modalities to acquire for each patient, balancing predictive accuracy against clinical invasiveness. SAGEAgent reasons about each patient's evolving diagnostic state through clinical tools that translate numerical predictions into text, an episodic memory that retrieves similar past cases, and a semantic memory that accumulates reusable decision patterns from experience. Experiments on a glioma cohort combining TCGA-LGG, TCGA-GBM, and BraTS with four diagnostic modalities demonstrate that SAGEAgent achieves competitive survival prediction accuracy while reducing average acquisition burden by 55%.

Figures

Figures reproduced from arXiv: 2607.09521 by Bennett A. Landman, Can Cui, Chongyu Qu, Juming Xiong, Junchao Zhu, Junlin Guo, Tianyuan Yao, Yanfan Zhu, Yuankai Huo, Yuechen Yang, Zhengyi Lu.

**Figure 2.** Figure 2: SAGEAgent architecture and reasoning. SAGEAgent Flow [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 3 linked inside Pith

[1]

arXiv preprint arXiv:2506.13474 (2025)

Bani-Harouni, D., Pellegrini, C., Özsoy, E., Keicher, M., Navab, N.: Language agents for hypothesis-driven clinical decision making with reinforcement learning. arXiv preprint arXiv:2506.13474 (2025)

arXiv 2025
[2]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Bernardino, G., Jonsson, A., Loncaric, F., Martí Castellote, P.M., Sitges, M., Clarysse, P., Duchateau, N.: Reinforcement learning for active modality selection during diagnosis. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 592–601. Springer (2022)

2022
[3]

Nature Reviews Cancer 22(2), 114–126 (2022)

Boehm, K.M., Khosravi, P., Vanguri, R., Gao, J., Shah, S.P.: Harnessing mul- timodal data integration to advance precision oncology. Nature Reviews Cancer 22(2), 114–126 (2022)

2022
[4]

In: International confer- ence on medical image computing and computer-assisted intervention

Braman, N., Gordon, J.W., Goossens, E.T., Willis, C., Stumpe, M.C., Venkatara- man, J.: Deep orthogonal fusion: multimodal prognostic biomarker discovery inte- grating radiology, pathology, genomic, and clinical data. In: International confer- ence on medical image computing and computer-assisted intervention. pp. 667–677. Springer (2021)

2021
[5]

IEEE Transactions on Medical Imaging41(4), 757–770 (2020)

Chen, R.J., Lu, M.Y., Wang, J., Williamson, D.F., Rodig, S.J., Lindeman, N.I., Mahmood, F.: Pathomic fusion: an integrated framework for fusing histopathology and genomic features for cancer diagnosis and prognosis. IEEE Transactions on Medical Imaging41(4), 757–770 (2020)

2020
[6]

In: Proceedings of the IEEE/CVF interna- tional conference on computer vision

Chen, R.J., Lu, M.Y., Weng, W.H., Chen, T.Y., Williamson, D.F., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for survival pre- diction in gigapixel whole slide images. In: Proceedings of the IEEE/CVF interna- tional conference on computer vision. pp. 4015–4025 (2021)

2021
[7]

Cox,D.R.:Regressionmodelsandlife-tables.Journaloftheroyalstatisticalsociety: Series B (methodological)34(2), 187–202 (1972)

1972
[8]

In: International Conference on Medi- cal Image Computing and Computer-Assisted Intervention

Cui, C., Liu, H., Liu, Q., Deng, R., Asad, Z., Wang, Y., Zhao, S., Yang, H., Land- man, B.A., Huo, Y.: Survival prediction of brain cancer with incomplete radiology, pathology, genomic, and demographic data. In: International Conference on Medi- cal Image Computing and Computer-Assisted Intervention. pp. 626–635. Springer (2022)

2022
[9]

In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers)

Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidi- rectional transformers for language understanding. In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers). pp. 4171–4186 (2019)

2019
[10]

IEEE Transactions on Big Data (2025)

Douze, M., Guzhva, A., Deng, C., Johnson, J., Szilvasy, G., Mazaré, P.E., Lomeli, M., Hosseini, L., Jégou, H.: The faiss library. IEEE Transactions on Big Data (2025)

2025
[11]

European journal of operational research109(2), 501–521 (1998)

Guitouni, A., Martel, J.M.: Tentative guidelines to help choosing an appropriate mcda method. European journal of operational research109(2), 501–521 (1998)

1998
[12]

In: International conference on machine learning

Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International conference on machine learning. pp. 1321–1330. PMLR (2017) 10 C. Qu et al

2017
[13]

Nature medicine30(9), 2613–2622 (2024)

Hager, P., Jungmann, F., Holland, R., Bhagat, K., Hubrecht, I., Knauer, M., Viel- hauer, J., Makowski, M., Braren, R., Kaissis, G., et al.: Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nature medicine30(9), 2613–2622 (2024)

2024
[14]

Jama247(18), 2543–2546 (1982)

Harrell, F.E., Califf, R.M., Pryor, D.B., Lee, K.L., Rosati, R.A.: Evaluating the yield of medical tests. Jama247(18), 2543–2546 (1982)

1982
[15]

Cancer cell40(10), 1095–1110 (2022)

Lipkova, J., Chen, R.J., Chen, B., Lu, M.Y., Barbieri, M., Shao, D., Vaidya, A.J., Chen, C., Zhuang, L., Williamson, D.F., et al.: Artificial intelligence for multimodal data integration in oncology. Cancer cell40(10), 1095–1110 (2022)

2022
[16]

Neuro-oncology 23(8), 1231–1251 (2021)

Louis, D.N., Perry, A., Wesseling, P., Brat, D.J., Cree, I.A., Figarella-Branger, D., Hawkins, C., Ng, H., Pfister, S.M., Reifenberger, G., et al.: The 2021 who classification of tumors of the central nervous system: a summary. Neuro-oncology 23(8), 1231–1251 (2021)

2021
[17]

IEEE transactions on medical imaging 34(10), 1993–2024 (2014)

Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., Burren, Y., Porz, N., Slotboom, J., Wiest, R., et al.: The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging 34(10), 1993–2024 (2014)

1993
[18]

arXiv preprint arXiv:2506.22405 (2025)

Nori, H., Daswani, M., Kelly, C., Lundberg, S., Ribeiro, M.T., Wilson, M., Liu, X., Sounderajah, V., Carlson, J., Lungren, M.P., et al.: Sequential diagnosis with language models. arXiv preprint arXiv:2506.22405 (2025)

Pith/arXiv arXiv 2025
[19]

The cancer imaging archive2(2016)

Pedano, N., Flanders, A.E., Scarpace, L., Mikkelsen, T., Eschbacher, J.M., Hermes, B., Ostrom, Q.: Radiology data from the cancer genome atlas low grade glioma [tcga-lgg] collection. The cancer imaging archive2(2016)

2016
[20]

In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion

Saadat, A., Hashemi, N., Vaseli, H., Tsang, M.Y., Luong, C., Van de Panne, M., Tsang, T.S., Abolmaesumi, P.: Precise-as: Personalized reinforcement learning for efficient point-of-care echocardiography in aortic stenosis diagnosis. In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion. pp. 474–484. Springer (2025)

2025
[21]

Scarpace, L., Mikkelsen, T., Cha, S., Rao, S., Tekchandani, S., Gutman, D., Saltz, J.H., Erickson, B.J., Pedano, N., Flanders, A.E., et al.: The cancer genome atlas glioblastomamultiformecollection(tcga-gbm).TheCancerImagingArchive(2016)

2016
[22]

arXiv preprint arXiv:1707.06347 (2017)

Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

Pith/arXiv arXiv 2017
[23]

Advances in neural information processing systems36, 8634–8652 (2023)

Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., Yao, S.: Reflexion: Lan- guage agents with verbal reinforcement learning. Advances in neural information processing systems36, 8634–8652 (2023)

2023
[24]

Organization of memory1(381- 403), 1 (1972)

Tulving, E., et al.: Episodic and semantic memory. Organization of memory1(381- 403), 1 (1972)

1972
[25]

arXiv preprint arXiv:2505.09388 (2025)

Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al.: Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025)

Pith/arXiv arXiv 2025

[1] [1]

arXiv preprint arXiv:2506.13474 (2025)

Bani-Harouni, D., Pellegrini, C., Özsoy, E., Keicher, M., Navab, N.: Language agents for hypothesis-driven clinical decision making with reinforcement learning. arXiv preprint arXiv:2506.13474 (2025)

arXiv 2025

[2] [2]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Bernardino, G., Jonsson, A., Loncaric, F., Martí Castellote, P.M., Sitges, M., Clarysse, P., Duchateau, N.: Reinforcement learning for active modality selection during diagnosis. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 592–601. Springer (2022)

2022

[3] [3]

Nature Reviews Cancer 22(2), 114–126 (2022)

Boehm, K.M., Khosravi, P., Vanguri, R., Gao, J., Shah, S.P.: Harnessing mul- timodal data integration to advance precision oncology. Nature Reviews Cancer 22(2), 114–126 (2022)

2022

[4] [4]

In: International confer- ence on medical image computing and computer-assisted intervention

Braman, N., Gordon, J.W., Goossens, E.T., Willis, C., Stumpe, M.C., Venkatara- man, J.: Deep orthogonal fusion: multimodal prognostic biomarker discovery inte- grating radiology, pathology, genomic, and clinical data. In: International confer- ence on medical image computing and computer-assisted intervention. pp. 667–677. Springer (2021)

2021

[5] [5]

IEEE Transactions on Medical Imaging41(4), 757–770 (2020)

Chen, R.J., Lu, M.Y., Wang, J., Williamson, D.F., Rodig, S.J., Lindeman, N.I., Mahmood, F.: Pathomic fusion: an integrated framework for fusing histopathology and genomic features for cancer diagnosis and prognosis. IEEE Transactions on Medical Imaging41(4), 757–770 (2020)

2020

[6] [6]

In: Proceedings of the IEEE/CVF interna- tional conference on computer vision

Chen, R.J., Lu, M.Y., Weng, W.H., Chen, T.Y., Williamson, D.F., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for survival pre- diction in gigapixel whole slide images. In: Proceedings of the IEEE/CVF interna- tional conference on computer vision. pp. 4015–4025 (2021)

2021

[7] [7]

Cox,D.R.:Regressionmodelsandlife-tables.Journaloftheroyalstatisticalsociety: Series B (methodological)34(2), 187–202 (1972)

1972

[8] [8]

In: International Conference on Medi- cal Image Computing and Computer-Assisted Intervention

Cui, C., Liu, H., Liu, Q., Deng, R., Asad, Z., Wang, Y., Zhao, S., Yang, H., Land- man, B.A., Huo, Y.: Survival prediction of brain cancer with incomplete radiology, pathology, genomic, and demographic data. In: International Conference on Medi- cal Image Computing and Computer-Assisted Intervention. pp. 626–635. Springer (2022)

2022

[9] [9]

In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers)

Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidi- rectional transformers for language understanding. In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers). pp. 4171–4186 (2019)

2019

[10] [10]

IEEE Transactions on Big Data (2025)

Douze, M., Guzhva, A., Deng, C., Johnson, J., Szilvasy, G., Mazaré, P.E., Lomeli, M., Hosseini, L., Jégou, H.: The faiss library. IEEE Transactions on Big Data (2025)

2025

[11] [11]

European journal of operational research109(2), 501–521 (1998)

Guitouni, A., Martel, J.M.: Tentative guidelines to help choosing an appropriate mcda method. European journal of operational research109(2), 501–521 (1998)

1998

[12] [12]

In: International conference on machine learning

Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International conference on machine learning. pp. 1321–1330. PMLR (2017) 10 C. Qu et al

2017

[13] [13]

Nature medicine30(9), 2613–2622 (2024)

Hager, P., Jungmann, F., Holland, R., Bhagat, K., Hubrecht, I., Knauer, M., Viel- hauer, J., Makowski, M., Braren, R., Kaissis, G., et al.: Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nature medicine30(9), 2613–2622 (2024)

2024

[14] [14]

Jama247(18), 2543–2546 (1982)

Harrell, F.E., Califf, R.M., Pryor, D.B., Lee, K.L., Rosati, R.A.: Evaluating the yield of medical tests. Jama247(18), 2543–2546 (1982)

1982

[15] [15]

Cancer cell40(10), 1095–1110 (2022)

Lipkova, J., Chen, R.J., Chen, B., Lu, M.Y., Barbieri, M., Shao, D., Vaidya, A.J., Chen, C., Zhuang, L., Williamson, D.F., et al.: Artificial intelligence for multimodal data integration in oncology. Cancer cell40(10), 1095–1110 (2022)

2022

[16] [16]

Neuro-oncology 23(8), 1231–1251 (2021)

Louis, D.N., Perry, A., Wesseling, P., Brat, D.J., Cree, I.A., Figarella-Branger, D., Hawkins, C., Ng, H., Pfister, S.M., Reifenberger, G., et al.: The 2021 who classification of tumors of the central nervous system: a summary. Neuro-oncology 23(8), 1231–1251 (2021)

2021

[17] [17]

IEEE transactions on medical imaging 34(10), 1993–2024 (2014)

Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., Burren, Y., Porz, N., Slotboom, J., Wiest, R., et al.: The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging 34(10), 1993–2024 (2014)

1993

[18] [18]

arXiv preprint arXiv:2506.22405 (2025)

Nori, H., Daswani, M., Kelly, C., Lundberg, S., Ribeiro, M.T., Wilson, M., Liu, X., Sounderajah, V., Carlson, J., Lungren, M.P., et al.: Sequential diagnosis with language models. arXiv preprint arXiv:2506.22405 (2025)

Pith/arXiv arXiv 2025

[19] [19]

The cancer imaging archive2(2016)

Pedano, N., Flanders, A.E., Scarpace, L., Mikkelsen, T., Eschbacher, J.M., Hermes, B., Ostrom, Q.: Radiology data from the cancer genome atlas low grade glioma [tcga-lgg] collection. The cancer imaging archive2(2016)

2016

[20] [20]

In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion

Saadat, A., Hashemi, N., Vaseli, H., Tsang, M.Y., Luong, C., Van de Panne, M., Tsang, T.S., Abolmaesumi, P.: Precise-as: Personalized reinforcement learning for efficient point-of-care echocardiography in aortic stenosis diagnosis. In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion. pp. 474–484. Springer (2025)

2025

[21] [21]

Scarpace, L., Mikkelsen, T., Cha, S., Rao, S., Tekchandani, S., Gutman, D., Saltz, J.H., Erickson, B.J., Pedano, N., Flanders, A.E., et al.: The cancer genome atlas glioblastomamultiformecollection(tcga-gbm).TheCancerImagingArchive(2016)

2016

[22] [22]

arXiv preprint arXiv:1707.06347 (2017)

Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

Pith/arXiv arXiv 2017

[23] [23]

Advances in neural information processing systems36, 8634–8652 (2023)

Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., Yao, S.: Reflexion: Lan- guage agents with verbal reinforcement learning. Advances in neural information processing systems36, 8634–8652 (2023)

2023

[24] [24]

Organization of memory1(381- 403), 1 (1972)

Tulving, E., et al.: Episodic and semantic memory. Organization of memory1(381- 403), 1 (1972)

1972

[25] [25]

arXiv preprint arXiv:2505.09388 (2025)

Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al.: Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025)

Pith/arXiv arXiv 2025