pith. sign in

arxiv: 2605.31014 · v2 · pith:EHTASJRCnew · submitted 2026-05-29 · 💻 cs.LG

SDM-Q: Cost-Aware Staged Decision-Making for Multi-Omics Classification with Deep Q-Learning

Nan Mu , Yangfan Xiao , Ling Wang , Xiaoning Li , Yue Kang , Chen Zhao This is my paper

Pith reviewed 2026-06-29 00:03 UTC · model grok-4.3

classification 💻 cs.LG
keywords multi-omics classificationdeep Q-learningcost-aware decision makingreinforcement learningprecision medicinesequential decision makingadaptive modality acquisitionstaged diagnosis
0
0 comments X

The pith

Deep Q-learning decides whether to acquire additional omics modalities or stop and classify, using a terminal reward that trades off accuracy against total cost.

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

The paper recasts multi-omics diagnosis as a finite-horizon sequential decision task in which the current set of acquired modalities forms the state. At each stage an action-value function, approximated by a deep Q-network, chooses either to request one more modality or to terminate with a prediction. The only reward is delivered at termination and equals classification correctness minus cumulative acquisition cost. A backward stage-wise optimization procedure trains the network to keep policies consistent across stages. On four public datasets the resulting policy reaches competitive accuracy while acquiring far fewer modalities than the full set on average.

Core claim

SDM-Q reformulates multi-omics classification as a finite-horizon Markov decision process whose state is the subset of acquired modalities, whose actions are to request a new modality or to terminate with a class prediction, and whose sole reward is a function of final classification correctness and negative cumulative cost; a deep Q-network learns the action-value function and a backward stage-wise optimization yields the policy.

What carries the argument

The action-value function that, at each stage, evaluates whether the expected improvement in terminal reward from acquiring one more modality exceeds its cost.

If this is right

  • More than 99 percent of subjects in the BRCA dataset receive accurate classification after a single modality.
  • More than 95 percent of subjects in the KIPAN dataset receive accurate classification after a single modality.
  • Average modalities acquired stay below two on the ROSMAP and LGG datasets.
  • Classification performance remains competitive with methods that always receive the complete multi-omics input.

Where Pith is reading between the lines

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

  • The same staged-decision structure could be applied to other costly multi-modal diagnostic pipelines that combine imaging, laboratory assays, and genetic panels.
  • The backward optimization step may transfer to other finite-horizon reinforcement-learning settings that require early stopping decisions.
  • Real-world use would need prospective validation to confirm that the learned stopping thresholds align with existing clinical workflows.

Load-bearing premise

The terminal reward defined only by final classification correctness and cumulative acquisition cost, together with backward stage-wise optimization, produces policies that stay stable and clinically relevant outside the four evaluated datasets.

What would settle it

A fifth multi-omics dataset, drawn from a different disease or population, on which the trained policy either requires an average number of modalities close to the full set or falls below the accuracy achieved by any full-modality baseline.

Figures

Figures reproduced from arXiv: 2605.31014 by Chen Zhao, Ling Wang, Nan Mu, Xiaoning Li, Yangfan Xiao, Yue Kang.

Figure 1
Figure 1. Figure 1: Stage-wise computational framework of SDM-Q. Dynamic State Encoding Module. This module constructs stage-specific multi￾omics states during sequential modality acquisition through dynamic masking and zero-padding. Since SDM-Q does not assume that all omics modalities are available at the initial stage, the state representation is progressively updated according to the modalities acquired along the decision… view at source ↗
Figure 2
Figure 2. Figure 2: Cost sensitivity analysis of SDM-Q under different modality acquisition cost configurations. The top row reports the average number of modalities acquired by the model, while the bottom row shows the classification accuracy under the corresponding cost settings. From left to right, the columns correspond to BRCA, KIPAN, LGG, and ROSMAP, respectively. Each heatmap cell represents the result obtained under a… view at source ↗
read the original abstract

Multi-omics data provide complementary molecular characterizations of disease phenotypes and play an important role in disease diagnosis and subtype classification in precision medicine. However, acquiring complete multi-omics profiles is expensive and time-consuming, while most existing deep learning methods assume full modality availability during inference, resulting in substantial redundancy and limited practicality in clinical settings. To address this issue, we propose SDM-Q, a reinforcement learning framework for adaptive and cost-aware multi-omics classification. Specifically, multi-omics diagnosis is reformulated as a finite-horizon sequential decision problem, where the currently acquired omics modalities define the diagnostic state at each stage. An action--value function determines whether to acquire an additional modality or terminate the decision process and output the final prediction. To balance diagnostic utility and acquisition cost, the reward is defined only at the terminal stage and jointly determined by classification correctness and cumulative modality acquisition cost. A backward stage-wise optimization strategy is introduced to improve policy consistency and training stability. Experiments on four public multi-omics datasets, including ROSMAP, LGG, BRCA, and KIPAN, demonstrate that SDM-Q effectively reduces redundant modality acquisition while maintaining competitive classification performance compared with methods using complete multi-omics inputs. In the BRCA and KIPAN datasets, more than 99\% and 95\% of subjects, respectively, achieve accurate classification using only a single omics modality, while the average number of acquired modalities remains below two for ROSMAP and LGG. These results suggest that cost-aware sequential decision-making provides an effective paradigm for improving the efficiency of precision medicine workflows.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper proposes SDM-Q, a deep Q-learning framework that reformulates multi-omics classification as a finite-horizon sequential decision process. At each stage the state is the set of acquired modalities; the agent chooses either to acquire one more modality or to terminate and output a classification. The terminal reward balances classification correctness against cumulative acquisition cost. A backward stage-wise optimization procedure is used to train the action-value function. Experiments on the four public datasets ROSMAP, LGG, BRCA and KIPAN are reported to show that the learned policy achieves competitive accuracy while acquiring far fewer modalities on average (e.g., >99 % of BRCA subjects classified correctly from a single modality).

Significance. If the reported empirical results hold under proper controls, the work supplies a concrete, cost-sensitive paradigm for multi-omics diagnosis that could materially reduce the expense of precision-medicine workflows. The explicit modeling of modality acquisition as a staged MDP with a terminal reward that trades accuracy against cost is a clear technical contribution, and the use of four public datasets provides a reproducible starting point for further study.

major comments (1)
  1. [Experimental evaluation section] Experimental evaluation section: the manuscript supplies no description of train/test splits, preprocessing pipelines, baseline implementations, error bars, or statistical significance tests for the four datasets. Because the central claim is that SDM-Q “maintains competitive classification performance” while reducing modality count, the absence of these controls renders the quantitative results unverifiable and is load-bearing for the empirical conclusion.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on the experimental evaluation. We agree that the current manuscript lacks sufficient detail on experimental controls, which is necessary to support the central claims regarding classification performance and modality reduction. We will revise the manuscript to address this.

read point-by-point responses

  1. Referee: [Experimental evaluation section] Experimental evaluation section: the manuscript supplies no description of train/test splits, preprocessing pipelines, baseline implementations, error bars, or statistical significance tests for the four datasets. Because the central claim is that SDM-Q “maintains competitive classification performance” while reducing modality count, the absence of these controls renders the quantitative results unverifiable and is load-bearing for the empirical conclusion.

    Authors: We acknowledge that the manuscript does not currently provide explicit descriptions of train/test splits, preprocessing pipelines, baseline implementations, error bars, or statistical significance tests. This is a valid concern that affects the verifiability of the reported results on ROSMAP, LGG, BRCA, and KIPAN. In the revised manuscript, we will expand the Experimental Evaluation section with a new subsection that includes: (i) details on train/test split ratios and any stratification or cross-validation procedures used; (ii) the full preprocessing pipelines applied to each dataset, including normalization, missing value handling, and feature selection; (iii) implementation details or references for all baselines; (iv) error bars (standard deviations across runs) for accuracy and modality count metrics; and (v) statistical significance tests (e.g., paired t-tests or McNemar’s test) with p-values comparing SDM-Q to baselines. These additions will directly support the claims of competitive performance with reduced modality acquisition. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation or claims

full rationale

The paper reformulates multi-omics classification as a standard finite-horizon MDP with states defined by acquired modalities, an action-value function for acquisition/termination decisions, and a terminal reward combining classification accuracy with cumulative cost. A backward stage-wise optimization is applied for training stability. These are conventional RL elements applied to the domain; no equation or result is shown to equal its own inputs by construction. Reported performance numbers (e.g., >99% single-modality accuracy on BRCA) are presented as experimental outcomes on four public datasets, not as fitted parameters or self-defined predictions. No self-citation load-bearing steps, uniqueness theorems, or ansatz smuggling appear in the abstract or described method. The work is self-contained empirical demonstration rather than a closed derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, mathematical axioms, or newly postulated entities.

pith-pipeline@v0.9.1-grok · 5831 in / 1081 out tokens · 28222 ms · 2026-06-29T00:03:49.654998+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    Z. Cai, R. C. Poulos, J. Liu, Q. Zhong, Machine learning for multi-omics data integration in cancer, Iscience 25 (2) (2022)

    2022

  2. [2]

    Zhang, Y

    J. Zhang, Y . Che, R. Liu, Z. Wang, W. Liu, Deep learning–driven multi-omics analysis: enhancing cancer diagnostics and therapeutics, Briefings in bioinfor- matics 26 (4) (2025) bbaf440

    2025

  3. [3]

    D. Xu, Y . Tang, J. Luo, C. Wen, Computational approaches to multimodal data integration in rheumatoid arthritis: from data landscape to clinical translation, Briefings in Bioinformatics 27 (1) (2026) bbag073

    2026

  4. [4]

    J. L. Ballard, Z. Wang, W. Li, L. Shen, Q. Long, Deep learning-based approaches for multi-omics data integration and analysis, BioData Mining 17 (1) (2024) 38

    2024

  5. [5]

    Teerapittayanon, B

    S. Teerapittayanon, B. McDanel, H.-T. Kung, Branchynet: Fast inference via early exiting from deep neural networks, in: 2016 23rd international conference on pattern recognition (ICPR), IEEE, 2016, pp. 2464–2469

    2016

  6. [6]

    Contardo, L

    G. Contardo, L. Denoyer, T. Artières, Sequential cost-sensitive feature acquisi- tion, in: International symposium on intelligent data analysis, Springer, 2016, pp. 284–294

    2016

  7. [7]

    C. An, Q. Zhou, S. Yang, A reinforcement learning guided adaptive cost- sensitive feature acquisition method, Applied Soft Computing 117 (2022) 108437

    2022

  8. [8]

    Z. Miao, H. Luo, M. Li, J. Zhang, Colaformer: Communicating local–global features with linear computational complexity, Pattern Recognition 157 (2025) 110870

    2025

  9. [9]

    Liang, H

    W. Liang, H. Wang, K. Zhang, J. Gong, Y . Gao, X. Tan, L. Ma, From static to adaptive multi-view: Nuanced expert prompt tuning for fine-grained image retrieval, Pattern Recognition (2026) 113884

    2026

  10. [10]

    Chaudhary, O

    K. Chaudhary, O. B. Poirion, L. Lu, L. X. Garmire, Deep learning–based multi- omics integration robustly predicts survival in liver cancer, Clinical cancer re- search 24 (6) (2018) 1248–1259

    2018

  11. [11]

    T. Wang, W. Shao, Z. Huang, H. Tang, J. Zhang, Z. Ding, K. Huang, Mogonet integrates multi-omics data using graph convolutional networks allowing pa- tient classification and biomarker identification, Nature communications 12 (1) (2021) 3445. 27

    2021

  12. [12]

    R. B. Tanvir, M. M. Islam, M. Sobhan, D. Luo, A. M. Mondal, Mogat: a multi- omics integration framework using graph attention networks for cancer subtype prediction, International Journal of Molecular Sciences 25 (5) (2024) 2788

    2024

  13. [13]

    N. Mu, H. Yang, C. Zhao, An uncertainty-aware dynamic decision framework for progressive multi-omics integration in classification tasks, Computer Meth- ods and Programs in Biomedicine (2025) 109179

    2025

  14. [14]

    Z. Wu, T. Nagarajan, A. Kumar, S. Rennie, L. S. Davis, K. Grauman, R. Feris, Blockdrop: Dynamic inference paths in residual networks, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 8817– 8826

    2018

  15. [15]

    R. S. Sutton, A. G. Barto, et al., Reinforcement learning: An introduction, V ol. 1, MIT press Cambridge, 1998

    1998

  16. [16]

    Pathak, P

    D. Pathak, P. Agrawal, A. A. Efros, T. Darrell, Curiosity-driven exploration by self-supervised prediction, in: International conference on machine learning, PMLR, 2017, pp. 2778–2787

    2017

  17. [17]

    D. A. Bennett, J. A. Schneider, Z. Arvanitakis, R. S. Wilson, Overview and find- ings from the religious orders study, Current Alzheimer Research 9 (6) (2012) 628–645

    2012

  18. [18]

    A Bennett, J

    D. A Bennett, J. A Schneider, A. S Buchman, L. L Barnes, P. A Boyle, R. S Wil- son, Overview and findings from the rush memory and aging project, Current Alzheimer Research 9 (6) (2012) 646–663

    2012

  19. [19]

    Cancer Genome Atlas Research Network, et al., The cancer genome atlas pan- cancer analysis project, Nat

    J. Cancer Genome Atlas Research Network, et al., The cancer genome atlas pan- cancer analysis project, Nat. Genet 45 (10) (2013) 1113–1120

    2013

  20. [20]

    C. G. A. Network, et al., Comprehensive molecular portraits of human breast tumours, Nature 490 (7418) (2012) 61–70

    2012

  21. [21]

    Fix, Discriminatory analysis: nonparametric discrimination, consistency properties, V ol

    E. Fix, Discriminatory analysis: nonparametric discrimination, consistency properties, V ol. 1, USAF school of Aviation Medicine, 1985

    1985

  22. [22]

    Cortes, V

    C. Cortes, V . Vapnik, Support-vector networks, Machine learning 20 (3) (1995) 273–297

    1995

  23. [23]

    Tibshirani, Regression shrinkage and selection via the lasso, Journal of the Royal Statistical Society Series B: Statistical Methodology 58 (1) (1996) 267– 288

    R. Tibshirani, Regression shrinkage and selection via the lasso, Journal of the Royal Statistical Society Series B: Statistical Methodology 58 (1) (1996) 267– 288

    1996

  24. [24]

    T. K. Ho, Random decision forests, in: Proceedings of 3rd international confer- ence on document analysis and recognition, V ol. 1, IEEE, 1995, pp. 278–282

    1995

  25. [25]

    W. S. McCulloch, W. Pitts, A logical calculus of the ideas immanent in nervous activity, The bulletin of mathematical biophysics 5 (4) (1943) 115–133. 28

    1943

  26. [26]

    M. A. Van De Wiel, T. G. Lien, W. Verlaat, W. N. van Wieringen, S. M. Wilting, Better prediction by use of co-data: adaptive group-regularized ridge regression, Statistics in medicine 35 (3) (2016) 368–381

    2016

  27. [27]

    Singh, C

    A. Singh, C. P. Shannon, B. Gautier, F. Rohart, M. Vacher, S. J. Tebbutt, K.-A. Lê Cao, Diablo: an integrative approach for identifying key molecular drivers from multi-omics assays, Bioinformatics 35 (17) (2019) 3055–3062

    2019

  28. [28]

    Z. Han, C. Zhang, H. Fu, J. T. Zhou, Trusted multi-view classification with dy- namic evidential fusion, IEEE transactions on pattern analysis and machine in- telligence 45 (2) (2022) 2551–2566

    2022

  29. [29]

    D. Hong, L. Gao, N. Yokoya, J. Yao, J. Chanussot, Q. Du, B. Zhang, More di- verse means better: Multimodal deep learning meets remote-sensing imagery classification, IEEE Transactions on Geoscience and Remote Sensing 59 (5) (2020) 4340–4354

    2020

  30. [30]

    Gated Multimodal Units for Information Fusion

    J. Arevalo, T. Solorio, M. Montes-y Gómez, F. A. González, Gated multimodal units for information fusion, arXiv preprint arXiv:1702.01992 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  31. [31]

    Z. Han, F. Yang, J. Huang, C. Zhang, J. Yao, Multimodal dynamics: Dynam- ical fusion for trustworthy multimodal classification, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 20707–20717

    2022

  32. [32]

    J. Liu, B. Zhuang, Z. Zhuang, Y . Guo, J. Huang, J. Zhu, M. Tan, Discrimination- aware network pruning for deep model compression, IEEE Transactions on Pat- tern Analysis and Machine Intelligence 44 (8) (2021) 4035–4051

    2021

  33. [33]

    C. Zhao, A. Liu, X. Zhang, X. Cao, Z. Ding, Q. Sha, H. Shen, H.-W. Deng, W. Zhou, Clclsa: Cross-omics linked embedding with contrastive learning and self attention for integration with incomplete multi-omics data, Computers in biology and medicine 170 (2024) 108058

    2024

  34. [34]

    F. Chen, Y . Zhang, Y . ¸ Senbabao˘glu, G. Ciriello, L. Yang, E. Reznik, B. Shuch, G. Micevic, G. De Velasco, E. Shinbrot, et al., Multilevel genomics-based tax- onomy of renal cell carcinoma, Cell reports 14 (10) (2016) 2476–2489

    2016

  35. [35]

    Y . Yuan, E. M. Van Allen, L. Omberg, N. Wagle, A. Amin-Mansour, A. Sokolov, L. A. Byers, Y . Xu, K. R. Hess, L. Diao, et al., Assessing the clinical utility of cancer genomic and proteomic data across tumor types, Nature biotechnology 32 (7) (2014) 644–652. 29

    2014