Pith. sign in

REVIEW 4 major objections 5 minor 37 references

The paper claims that a Deep Q-Network, trained inside a random-forest simulator of bladder cancer recurrence, can learn sequential treatment policies that outperform classical RL methods, achieving a cumulative reward of 63,918.87 in the s

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 · deepseek-v4-flash

2026-08-01 19:31 UTC pith:7DRIW7MX

load-bearing objection The simulator produces physically impossible tumor sizes; the headline reward is an artifact. the 4 major comments →

arxiv 2607.16916 v1 pith:7DRIW7MX submitted 2026-07-18 cs.LG

Enhancing Personalized Bladder Cancer Treatment Through Reinforcement Learning: A Recurrent Patient State Transition Decision Support Framework

Divyansh Chawla , Anshu Garg , Isshaan Singh This is my paper
classification cs.LG
keywords bladder cancerrecurrent event modelingclinical decision supportpatient state-transition simulationreinforcement learningdeep Q-networkMarkov decision processprecision oncology
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.

The paper tries to establish that recurrent bladder cancer—where a treatment's effectiveness changes across successive episodes—can be handled by a reinforcement learning agent that learns from simulated patient trajectories rather than static guidelines. The authors construct a simulator in which random forest regressors predict how tumor size and tumor count change after each treatment, then train a Deep Q-Network to choose among placebo, pyridoxine, and thiotepa across thousands of rollouts. They report that this agent accumulates a reward of 63,918.87, with an average per-episode loss of 0.0056 and a policy improvement score of 6.62%, beating value iteration, Monte Carlo, and Q-learning in the same simulator. The authors present this as evidence that sequential learning inside a state-transition simulator captures recurrent disease dynamics better than single-step prediction or tabular RL, and that the modular design lets better predictive models be installed without retraining the decision loop.

Core claim

On the paper's own terms, the central discovery is that a recurrent patient state-transition simulation—constructed from random forest models that predict the next tumor size and count given treatment and recurrence history—provides a sufficiently rich environment for a Deep Q-Network to learn a treatment policy that accrues a cumulative reward of 63,918.87, compared with -354,770 for Q-learning and similar negative totals for value iteration and Monte Carlo. The authors interpret this gap as evidence that the DQN balances exploration and exploitation across heterogeneous patient trajectories, while the classical methods get stuck in suboptimal regions. They also argue that the interpretable

What carries the argument

The load-bearing mechanism is the state-transition simulation loop. Two random forest regressors, g_size and g_number, are trained on the Bladder1 dataset to estimate changes in tumor size and count from the feature vector (treatment, number of tumors, tumor size, recurrence rate, treatment start, treatment stop). The simulator rolls this out for M iterations per patient to produce roughly 29,399 simulated episodes; the DQN then applies an epsilon-greedy policy over the three treatments, using a reward of +10 for a reduction in size or count, -10 for an increase, and -10/-100/-5 for recurrence, death from bladder cancer, or death from other causes. The random forest models are explicitly fra

Load-bearing premise

The entire reported policy quality rests on the random-forest transition models, trained on 118 patients' trial records, remaining accurate when rolled out for hundreds of steps; the paper never validates the multi-step rollouts against real held-out trajectories.

What would settle it

Run the paper's simulator from real patient starting states for the same number of iterations used in training and compare the predicted tumor size and count trajectories against the actual observed intervals in held-out Bladder1 patients; if the predicted trajectories drift after just a few steps, the reported cumulative reward of 63,918.87 is an artifact of the fitted simulator, not a clinically meaningful outcome.

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

If this is right

  • Treatment recommendations can be re-evaluated after every episode, so a patient's plan adapts to how their tumor metrics respond rather than following a fixed protocol.
  • The framework outputs complete treatment trajectories and simulation logs, giving clinicians a record of why each action was chosen.
  • Because the transition model is modular, replacing the random forest with a more accurate predictor does not require redesigning the reinforcement learning loop.
  • The reported comparison positions the framework as applicable beyond radiotherapy parameter tuning, extending DRL-based planning to longitudinal regimen selection.

Where Pith is reading between the lines

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

  • The paper's headline numbers are only as meaningful as the simulator: a direct check would be to roll out the random forest models from real patient starting states for many steps and compare predicted tumor trajectories to the actual held-out intervals; divergence after a few episodes would make the 63,918.87 reward a property of the fitted model, not of clinical reality.
  • The reward function deliberately excludes side effects and quality of life, so the learned policy minimizes tumor burden alone; adding toxicity or patient-reported outcomes to the reward could visibly change which treatments are preferred.
  • The action space holds only the three trial arms, so the framework's clinical reach depends on extending the transition model to modern agents; that extension is a natural next test and would require new data rather than any change to the RL machinery.
  • The paper acknowledges that the MDP's first-order Markov assumption may be too simple for recurrence biology; testing recurrent or attention-based state representations against the current random-forest rollouts would settle whether longer temporal dependencies matter.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a decision-support framework for personalized bladder cancer treatment that couples a Random Forest (RF) state-transition simulator with a Deep Q-Network (DQN). The RF models are trained on the Bladder1 dataset to predict changes in tumor size and count, and the DQN is trained inside this simulator to choose among placebo, pyridoxine, and thiotepa over multiple simulated episodes. The authors report a cumulative reward of 63,918.87, an average training loss per episode of 0.0056, a policy improvement score of 6.62%, and much larger tumor-size/count reductions for DQN than for value iteration, Monte Carlo, or Q-learning. They also compare against DRL radiotherapy planning frameworks and perform sensitivity analyses over learning rate, discount factor, and batch size. The conclusion is that the recurrent state-transition simulation plus DQN forms a flexible, modular decision-support tool for longitudinal treatment planning.

Significance. If the central performance claim were credible, the framework would make a modest contribution: it decouples the predictive transition model from the RL optimizer, is modular, and provides explicit simulation logs for interpretability. The paper also performs feature-importance diagnostics and hyperparameter sensitivity checks, which are good practices. However, every headline result is produced inside a simulator that is fit to the same small dataset used to define the task, and the simulator itself produces physically impossible tumor-size reductions. The main claim is therefore not externally substantiated, and the internal reward signal is contaminated by fictional state trajectories. The framework may be useful as a methodological prototype, but the evidence presented does not support the claimed effectiveness or robustness.

major comments (4)
  1. [Section III.B, Eqs. (6)-(8); Table 2] The reported DQN total tumor size reduction of 41,437.87 is physically impossible. Section IV.B states that Bladder1 has 118 patients with baseline largest-tumor size in single-digit centimeters; the maximum possible total reduction over nonnegative trajectories is on the order of a few hundred. Equation (7) defines ΔS_i^(j)=S_i^(j-1)-S_i^(j), and Eq. (8) gives +10 for any positive ΔS. If the RF model predicts negative S_i^(j), the accumulated reductions become arbitrary large positive numbers. Algorithm 1 imposes no non-negativity constraint on predicted tumor size, and the paper reports no check of predicted states. Thus the headline cumulative reward and reductions are artifacts of an invalid simulator, not evidence of treatment efficacy. This is a load-bearing internal inconsistency.
  2. [Section III.B, Eqs. (8)-(10)] The reward definition is internally inconsistent. Equation (8)'s third branch reads '0 if ΔS<0 (No change)', but the intended condition is ΔS=0. Equation (9), described as an additional reward for tumor count changes, uses ΔS_i^(j) in the conditions instead of ΔN_i^(j). Equation (10) adds penalties for R=1,2,3, but earlier in Eq. (6) R is used as a recurrence feature, and the mapping from state to recurrence status is unspecified. Moreover, Eq. (5) defines R_t = Δsize + Δcount, while Eqs. (8)-(10) add discrete ±10/0 bonuses; Section III.D repeats R=(tumor size change)+(tumor count change), which is not the same reward. The reported cumulative reward is therefore not well-defined, and the implementation of the reward cannot be verified from the paper.
  3. [Section IV.B and Section IV.D] The DQN is trained and evaluated exclusively within a simulator whose transition model is fit to the same Bladder1 dataset. The authors argue that 29,399 simulated episodes 'amplify the learning space' beyond the original 118 patients, but the simulated episodes are generated by a model trained on those same 118 patients. If the RF rollouts diverge from real disease progression over multiple steps—which is unexamined—the learned policy and the cumulative reward are meaningless. The paper never validates multi-step rollouts against held-out real patient trajectories nor reports distributional fidelity of predicted states. This circularity is load-bearing because all performance claims are computed inside this fitted environment.
  4. [Section IV.F, Table 4] The comparison with DRL-based radiotherapy planning frameworks is not meaningful. The frameworks differ in disease, action space, state space, and reward scale, so raw cumulative reward (63,918.87 vs. 9,350–50,706), training loss, and reward standard deviation are not comparable across environments. In addition, Table 4 lists 'Convergence Episode' as 1 for every method, while the text states that the proposed framework reaches stable performance after approximately 1,000 episodes and that other frameworks converge faster. This internal contradiction undermines the comparative claims in Section IV.F.
minor comments (5)
  1. [Abstract / Introduction] There is a duplicated sentence in the Motivation subsection: 'The motivation behind this research is to bridge the gap between predictive modeling and RL in oncology...' appears twice nearly verbatim. Section IV.F also contains a repeated opening sentence.
  2. [Section III.E, Eqs. (24)-(25)] The notation for Total Tumor Size Reduction and Total Tumor Count Reduction sums over i=1 to M, but M is the number of simulation iterations per patient and i indexes patients. The sum should be over patients (n), not over iterations, and the index is misleading.
  3. [Algorithm 1] Algorithm 1 does not define M, the number of simulation iterations, nor a termination criterion for each patient's episode. The reader must infer M from the text; this should be stated explicitly in the algorithm preamble.
  4. [Section IV.B] The dataset description says 'Number represents the initial tumor count, ranging up to 8, with 8 denoting "eight or more tumors"' but the later feature-importance discussion refers to 'number' as a feature. Please clarify whether the predicted count is capped or treated as a continuous variable, since this affects the plausibility of count reductions.
  5. [Section IV.A] The text says experiments used 'Google Collaboratory'; the correct product name is 'Google Colaboratory' or 'Colab'. Also, reporting both a Windows 10 system and Google Colaboratory is confusing—please specify where the experiments were actually run.

Circularity Check

1 steps flagged

The headline DQN reward is a sum of the same random-forest predictions used to build the simulator, so the central performance claim reduces to the fitted transition model rather than to any independent clinical outcome.

specific steps
  1. fitted input called prediction [Section III.A, Eqs. (2)-(5); Section III.D, Eq. (17); Section IV.B; Section IV.D, Table 2]
    "We define the target variables ysize and ynumber as the expected changes in tumor size and count, respectively. ... Tumor Size Model: ŷ_size = g_size(X) (3) ... Reward Rt = Δsize + Δcount (5) ... R = (tumor size change) + (tumor count change) (17)."

    The random-forest models g_size and g_number are trained to output the expected changes in tumor size/count from the Bladder1 data. The DQN reward is defined as exactly those changes (Eqs. 5 and 17), and Algorithms 1-2 generate every state transition and Δ value from the same fitted models. The reported cumulative reward (63,918.87) and cumulative tumor-size reduction (41,437.87) in Table 2 are therefore sums of the fitted models' own outputs over M-step rollouts. Calling these numbers evidence of 'robust treatment optimization' presents the fitted simulator's output as an independent prediction, when the evaluation metric and the learned environment are the same fitted quantity by construction.

full rationale

The paper contains no load-bearing self-citation chain: the DRL frameworks cited for comparison [29]-[32] are external works, and the state-transition models are fit to the public Bladder1 dataset. The circularity is in the evaluation loop, not in the references. The DQN is trained and evaluated entirely inside a simulator whose transition function is the random-forest model fit to the same data used to define the task; the reward that the agent maximizes (Eq. 17) is the model's own predicted Δsize+Δcount, and the headline metrics are sums of those same predictions. The paper's own limitation statement concedes that 'evaluating the framework across multiple clinical datasets would provide additional evidence of robustness, such datasets were not accessible within the scope of this study.' The internal inconsistency reinforces this: with 118 patients and single-digit-centimeter initial tumor sizes, a reported total tumor-size reduction of 41,437.87 is only achievable if rollouts produce implausible, effectively negative sizes, so the reward driving the DQN is dominated by the fitted model's drift rather than by a measured treatment effect. This is a partial circularity rather than a full tautology because the framework does contain independent algorithmic content (RF-based transition modeling plus DQN, interpretability logs) and the paper is transparent that the environment is simulated; but the central numerical claims reduce to the fitted simulator by construction.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 0 invented entities

The central claim rests on the random-forest transition model's ability to generate realistic multi-step trajectories, plus hand-chosen reward weights and simulation length. The DQN results are entirely defined by these choices. No new physical or clinical entities are postulated.

free parameters (8)
  • Reward weights = +10/-10/-100/-5
    Hand-chosen in Eq (8)-(10) for size/count reduction, recurrence, and death. These directly determine the cumulative reward metric and the learned policy.
  • Number of simulation iterations M = ~249 (inferred from 29,399 episodes / 118 patients)
    M is not stated explicitly. The paper says '29,399 simulated episodes' in Section IV.B; if this is M per patient, M≈249. This is a key driver of cumulative reward magnitude and assumes multi-step validity of the transition model.
  • Discount factor gamma = 0.99
    Chosen in Section III.C to balance immediate and future rewards.
  • Learning rate = 0.001
    Adam optimizer learning rate for DQN training.
  • Batch size = 32
    Mini-batch size for experience replay.
  • Replay buffer capacity = 10,000
    Stores past experiences for sampling.
  • Target network update frequency = every 10 episodes
    Synchronizes target Q-network weights.
  • Random Forest hyperparameters = unspecified
    Number of trees, max depth, etc. not reported; these affect transition-model quality and therefore all downstream RL numbers.
axioms (4)
  • domain assumption Markov property: future tumor state depends only on current features and treatment
    Section III.C states 'future patient states depend primarily on the current clinical state and treatment action'; acknowledged as a limitation in Section VI.
  • domain assumption Tumor size and count are sufficient state variables
    State vector x_i includes treatment, number of tumors, tumor size, recurrence, start, stop; no other clinical variables are used.
  • domain assumption RF models trained on Bladder1 data can be iteratively applied to generate valid long-horizon trajectories
    This is the core load-bearing assumption for the simulation and DQN evaluation, entered in Section III.B and used throughout.
  • domain assumption The Bladder1 dataset is representative for learning treatment-response dynamics
    The authors acknowledge in Section V that evaluating across multiple datasets was not possible, so the framework rests on this single small cohort.

pith-pipeline@v1.3.0-alltime-deepseek · 25928 in / 12389 out tokens · 125160 ms · 2026-08-01T19:31:38.255860+00:00 · methodology

0 comments
read the original abstract

Bladder cancer treatment requires personalized and adaptive decision-making, particularly for recurrent disease, where treatment effectiveness changes across successive clinical episodes. Conventional clinical decision support systems typically rely on static treatment guidelines or single-step predictive models, limiting their ability to capture disease progression over time. This paper presents a recurrent patient state-transition simulation framework for bladder cancer treatment planning that integrates predictive state-transition modeling with a Markov Decision Process (MDP) and a Deep Q-Network (DQN) reinforcement learning environment. The predictive module estimates changes in tumor characteristics following treatment, while the reinforcement learning agent sequentially optimizes treatment decisions by interacting with simulated patient trajectories. This framework enables dynamic, patient-specific treatment planning by continuously adapting recommendations to evolving clinical states. It also generates interpretable treatment trajectories and detailed simulation logs to improve transparency and support clinical decision-making. The proposed framework was evaluated against existing reinforcement learning-based treatment planning approaches. It achieved a cumulative reward of 63,918.87, an average training loss per episode of 0.0056, and a policy improvement score of 6.62%, demonstrating effective sequential learning and robust treatment optimization in a simulated recurrent treatment environment. These findings highlight the potential of recurrent patient state-transition simulation with reinforcement learning as a flexible decision-support framework for personalized bladder cancer treatment planning and AI-assisted precision oncology.

Figures

Figures reproduced from arXiv: 2607.16916 by Anshu Garg, Divyansh Chawla, Isshaan Singh.

Figure 1
Figure 1. Figure 1: Proposed Framework for the Bladder Cancer [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Initial Tumor Categories by Treatment [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: depicts the total number of initial tumors across all treatment groups. This plot aggregates the tumor categories, showing how many patients fall into each tumor category regardless of the treatment received. It offers a broad overview of the tumor burden in the entire cohort at baseline, helping to identify how common or rare each tumor category is. This is significant for understanding the disease distri… view at source ↗
Figure 6
Figure 6. Figure 6: Deviations between initial and recurrent tumours [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 4
Figure 4. Figure 4: Treatment Distribution of the therapy [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: presents the Kaplan-Meier survival curve, which illustrates the survival probability of patients over time. The survival curve is based on the duration between the start and stop times for each patient, with the event status indicating whether the patient experienced recurrence, died from bladder cancer, or died from other causes. This plot is significant because it provides a visual representation of pati… view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of Feature Importance [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗

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

37 extracted references · 18 canonical work pages · 1 internal anchor

  1. [1]

    Tianlei, Xie., Shan, Peng., Shujun, Liu., Minghao, Zheng., Wenli, Diao., Meng, Ding., Yao, Fu., Hongqian, Guo., Wei, Zhao., Junlong, Zhuang. (2024). 1. Multi -cohort validation of Ascore: an anoikis -based prognostic signature for predicting disease prog ression and immunotherapy response in bladder cancer. Molecular Cancer, doi: 10.1186/s12943-024-01945-9

    work page doi:10.1186/s12943-024-01945-9 2024

  2. [2]

    Kit, Gallagher., Maximilian, Strobl., Derek, S., Park., Fabian, C., Spoendlin., Robert, A., Gatenby., Philip, K., Maini., Alexander, R.A., Anderson. (2024). Data from Mathematical Model -Driven Deep Learning Enables Personalized Adaptive Therapy. doi: 10.1158/0008 - 5472.c.7267974.v1

    doi:10.1158/0008 2024

  3. [3]

    Elena, M., Tosca., Alessandro, De, Carlo., Davide, Ronchi., Paolo, Magni. (2024). 1. Model‐Informed Reinforcement Learning for Enabling Precision Dosing Via Adaptive Dosing. Clinical Pharmacology & Therapeutics, doi: 10.1002/cpt.3356

    work page doi:10.1002/cpt.3356 2024

  4. [4]

    Nesrine, Lajmi., Sofia, Alves -Vasconcelos., Apostolos, Tsiachristas., Andrew, Haworth., Kerrie, Woods., Charles, Crichton., Theresa, Noble., Hizni, Salih., Kinga, A, Várnai., Harriet, Branford -White., Liam, Orrell., Andrew, Osman., Kevin, M., Bradley., Lara, Bonney., Daniel, R., McGowan., Jim, Davies., Matthew, S, Prime., Andrew, Bassim, Hassan. (2024)....

    doi:10.1017/pcm.2024.1 2024

  5. [5]

    (2023). 3. Case Studies for Overcoming Challenges in Using Big Data in Cancer. Cancer Research, doi: 10.1158/0008-5472.can-22-1277

    doi:10.1158/0008-5472.can-22-1277 2023

  6. [6]

    John, Adeoye., L., M., Christy, Hui., Yu -xiong, Su. (2023). 1. Data -centric artificial intelligence in oncology: a systematic review assessing data quality in machine learning models for head and neck cancer. Journal of Big Data, doi: 10.1186/s40537-023-00703-w

    work page doi:10.1186/s40537-023-00703-w 2023

  7. [7]

    Vitaly, V., Bozhenko., T., M., Tatarnikova. (2023). 3. Application of Data Preprocessing in Medical Research. doi: 10.1109/WECONF57201.2023.10148004

    arXiv 2023

  8. [8]

    Sophia, Yazzourh., Nicolas, Savy., Philippe, Pierre., Michael, R., Kosorok. (2024). 2. Medical Knowledge Integration into Reinforcement Learning Algorithms for Dynamic Treatment Regimes. doi: 10.48550/arxiv.2407.00364

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2407.00364 2024

  9. [9]

    G., Saranya., K., Sivaraman. (2024). 4. Applying Reinforcement Learning to Optimize Treatment Plans for Chronic Disease Management. doi: 10.1109/accai61061.2024.10601732

    arXiv 2024

  10. [10]

    Dr.Vinod, Vegesna. (2024). AI -Driven Personalized Medicine: A Frame Work for Tailored Cancer Treatment. International journal of innovative research in advanced engineering, doi: 10.26562/ijirae.2024.v1106.06

    work page doi:10.26562/ijirae.2024.v1106.06 2024

  11. [11]

    Sumana, Srivatsa., David, Walsh., Mayada, Aljehani., Jason, Weinreb., Nathan, Becker., Benjamin, H., Ellis., Renee, George., Naim, Matasci., Reva, Basho., Mitchell, E., Gross., Christine, L., Swisher. (2024). Abstract 6467: Causal personalized treatme nt estimation framework using real - world electronic health record (EHR) data to inform cancer care deci...

    work page doi:10.1158/1538 2024

  12. [12]

    Judith, A.C., Rietjens., Ingeborg, P., M., Griffioen., Jorge, Sierra -Pérez., Gaby, Sroczynski., Uwe, Siebert., Alena, Buyx., Barbara, Peric., Inge, Marie, Svane., Jasper, B., P., Brands., Karina, Dahl, Steffensen., Carlos, Romero, Piqueras., Elham, He dayati., Maria, Karsten., Norbert, Couespel., Canan, Akoglu., Roberto, Pazo -Cid., Paul, Rayson., Hester...

    2024

  13. [13]

    Maria, Torrente., Ernestina, Menasalvas., Mariano, Provencio. (2023). 4. The CLARIFY digital decision support platform: An artificial intelligence tool for exploring multidimensional cancer data.. Journal of Clinical Oncology, doi: 10.1200/jco.2023.41.16_suppl.e13638

    work page doi:10.1200/jco.2023.41.16_suppl.e13638 2023

  14. [14]

    Rohit, P., Ojha., Yan, Lu., Kalyani, Narra., Rachel, J, Meadows., Aaron, W., Gehr., Emmanuel, Mantilla., Bassam, Ghabach. (2023). 5. Survival After Implementation of a Decision Support Tool to Facilitate Evidence -Based Cancer Treatment. JCO clinical c ancer informatics, doi: 10.1200/CCI.23.00001

    work page doi:10.1200/cci.23.00001 2023

  15. [15]

    Jonathan, Masih., Rajiv, Philip. (2023). Compartmental and Probabilistic Modelling for Research into the Development of Cancer and its Biological History. doi: 10.59828/ijsrmst.v2i4.91

    work page doi:10.59828/ijsrmst.v2i4.91 2023

  16. [16]

    Oncology Simulation Model: A Comprehensive and Innovative Approach to Estimate and Project Prevalence and Survival in Oncology

    (2022). Oncology Simulation Model: A Comprehensive and Innovative Approach to Estimate and Project Prevalence and Survival in Oncology. Clinical Epidemiology, doi: 10.2147/clep.s377093

    work page doi:10.2147/clep.s377093 2022

  17. [17]

    Prediction of Cancer Treatment Effectiveness and Patient Outcomes using Machine Learning Classification Approaches - A Review

    (2023). Prediction of Cancer Treatment Effectiveness and Patient Outcomes using Machine Learning Classification Approaches - A Review. International Journal of Scientific Research in Science and Technology, doi: 10.32628/ijsrst523103148

    work page doi:10.32628/ijsrst523103148 2023

  18. [18]

    Mohammad, Mahjoob. (2023). 4. Multiscale modeling of the dynamic growth of cancerous tumors under the influence of chemotherapy drugs. Computer Methods in Biomechanics and Biomedical Engineering, doi: 10.1080/10255842.2023.2215368

    arXiv 2023

  19. [19]

    Artificial Intelligence in Oncology: Current Capabilities, Future Opportunities, and Ethical Considerations

    (2022). Artificial Intelligence in Oncology: Current Capabilities, Future Opportunities, and Ethical Considerations. American Society of Clinical Oncology educational book, doi: 10.1200/edbk_350652

    work page doi:10.1200/edbk_350652 2022

  20. [20]

    Lin, Huo., Yue, Tang. (2022). Multi -Objective Deep Reinforcement Learning for Personalized Dose Optimization Based on Multi -Indicator Experience Replay. Applied Sciences, doi: 10.3390/app13010325

    work page doi:10.3390/app13010325 2022

  21. [21]

    Fawaz, E., Alsaadi., Amirreza, Yasami., Christos, Volos., Stelios, Bekiros., Hadi, Jahanshahi. (2023). A New Fuzzy Reinforcement Learning Method for Effective Chemotherapy. Mathematics, doi: 10.3390/math11020477

    work page doi:10.3390/math11020477 2023

  22. [22]

    Vahid, Reza, Niazmand., Mohammad, Ali, Raheb., Navid, Eqra., Ramin, Vatankhah., Amirmohammad, Farrokhi. (2024). 1. Deep reinforcement learning control of combined chemotherapy and anti -angiogenic drug delivery for cancerous tumor treatment. Computers in Biology and Medicine, doi: 10.1016/j.compbiomed.2024.109041

    arXiv 2024

  23. [23]

    George, A., Vouros. (2022). Explainable Deep Reinforcement Learning: State of the Art and Challenges. ACM Computing Surveys, doi: 10.1145/3527448

    doi:10.1145/3527448 2022

  24. [24]

    Andreas, D., Meid., Alexander, Gerharz., Andreas, Groll. (2022). Machine learning for tumor growth inhibition: Interpretable predictive models for transparency and reproducibility. CPT: pharmacometrics & systems pharmacology, doi: 10.1002/psp4.12761

    work page doi:10.1002/psp4.12761 2022

  25. [25]

    Kyle, Rose., Aram, Vosoughi., Gustavo, Borjas., Heather, L., Huelster., Philippe, E., Spiess., Anders, Berglund., Wade, J., Sexton., Anirudh, Joshi., Nagi, B., Kumar., RogerW, ., Li. (2023). Complimentary genomic, pathologic, and artificial intelligen ce analysis on low -grade noninvasive bladder cancer to predict downstream recurrence.. Journal of Clinic...

    work page doi:10.1200/jco.2023.41.6_suppl.553 2023

  26. [26]

    A review on modeling tumor dynamics and agent reward functions in reinforcement learning based therapy optimization

    (2022). A review on modeling tumor dynamics and agent reward functions in reinforcement learning based therapy optimization. Journal of Intelligent and Fuzzy Systems, doi: 10.3233/jifs-212351

    work page doi:10.3233/jifs-212351 2022

  27. [27]

    https://www.kaggle.com/datasets/utkarshx27/bladder - cancer-recurrences

  28. [28]

    Regression analysis of multivariate incomplete failure time data by modeling marginal distributions

    Wei, Lee -Jen, Danyu Y. Lin, and Lisa Weissfeld. "Regression analysis of multivariate incomplete failure time data by modeling marginal distributions." Journal of the American statistical association 84.408 (1989): 1065-1073

    1989

  29. [29]

    Automating the optimization of proton PBS treatment planning for head and neck cancers using policy gradient‐based deep reinforcement learning

    Wang, Qingqing, and Chang Chang. "Automating the optimization of proton PBS treatment planning for head and neck cancers using policy gradient‐based deep reinforcement learning." Medical Physics

  30. [30]

    Breast radiation therapy fluence painting with multi‐agent deep reinforcement learning

    Dongrong, Yang, et al. "Breast radiation therapy fluence painting with multi‐agent deep reinforcement learning." Medical Physics

  31. [31]

    Clinical VMAT machine parameter optimization for localized prostate cancer using deep reinforcement learning

    Hrinivich, William T., et al. "Clinical VMAT machine parameter optimization for localized prostate cancer using deep reinforcement learning." Medical physics 51.6 (2024): 3972-3984

    2024

  32. [32]

    Human -like intelligent automatic treatment planning of head and neck cancer radiation therapy

    Gao, Yin, Yang Kyun Park, and Xun Jia. "Human -like intelligent automatic treatment planning of head and neck cancer radiation therapy." Physics in Medicine & Biology 69.11 (2024): 115049

    2024

  33. [33]

    Alessandro, Ottaiano., Francesca, Grassi., Roberto, Sirica., Emanuela, Genito., Giovanni, Ciani., Vittorio, Patanè., Riccardo, Monti., Maria, Paola, Belfiore., Fabrizio, Urraro., Mariachiara, Santorsola., Alfonso, Maria, Ponsiglione., Marco, Montella., S alvatore, Cappabianca., Alfonso, Reginelli., Carlo, Sansone., Giovanni, Savarese., Roberta, Grassi. (2...

    2024

  34. [34]

    Hua -Jun, Zhou., Fengtao, Zhou., Zhao, Chenyu., Yingxue, Xu., Luyang, Luo., Hao, Chen. (2024). Multimodal Data Integration for Precision Oncology: Challenges and Future Directions. doi: 10.48550/arxiv.2406.19611

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.19611 2024

  35. [35]

    Sayan, Ghosal., Qiang, Chen., Giulio, Pergola., Aaron, L., Goldman., William, S, Ulrich., Karen, Faith, Berman., Giuseppe, Blasi., Leonardo, Fazio., Antonio, Rampino., Alessandro, Bertolino., Daniel, R., Weinberger., Venkata, S., Mattay., Archana, Venkat araman. (2021). G -MIND: An End-to-End Multimodal Imaging -Genetics Framework for Biomarker Identifica...

    2021

  36. [36]

    Pascal, Zille., Yu -Ping, Wang. (2017). Coupled Dimensionality-Reduction Model for Imaging Genomics. doi: 10.1007/978-3-319-67675-3_22

    work page doi:10.1007/978-3-319-67675-3_22 2017

  37. [37]

    Ming, Fan., Jiangning, Song., Zhaowen, Qiu. (2022). Editorial: Biomedical image or genomic data characterization and radiogenomic/image -omics. Frontiers in Genetics, doi: 10.3389/fgene.2022.994880

    arXiv 2022