{"id":"44fc434a-8aed-46cb-9c97-de6d8915615a","arxiv_id":"2505.09899","paper_version":1,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A proposal to combine neural ODEs, physics-informed networks, and reinforcement learning for personalized radiopharmaceutical dosing, presented without experimental validation.","lead":"This paper reviews computational tools for personalized cancer treatment and proposes a framework that uses neural networks to model drug distribution and reinforcement learning to plan dosing. It is a research proposal with no implementation or test results.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposed training loss in §IV is internally inconsistent: Eq. 6 contradicts the PBPK dynamics in Eqs. 2–4, and Eq. 5 does not supervise against simulator data.","rationale":"The reader correctly notes the unvalidated sim-to-real transfer and unsubstantiated MDP transition modeling. However, the more load-bearing issue is internal: the proposed physics-informed training objective cannot do what the framework requires of it. Eq. 5 only enforces consistency between two neural networks, and Eq. 6 imposes a conservation law that contradicts the stated PBPK model. Because this flawed objective is the interface between the PBPK simulator and the RL-based decision policy, the central claim—that this pipeline will enable optimized patient-specific dosing—is unsupported at the level of mathematical formulation, before any consideration of generalization. This is not a demand for empirical validation (which would leave the verdict UNVERDICTED) but a concrete correctness objection to the framework as written. The framework could be repaired by adding a supervised simulator-fidelity term and correcting the conservation constraint, so rejection outright is too strong; a conditional decision requiring correction and a demonstrative fit to the simulator is the appropriate verdict. My concern partially overlaps with the reader's weakest_assumption (both point to the model's fidelity to real or simulated pharmacokinetics), but the reader located the failure at sim-to-real transfer, whereas I find a more basic failure in the training loss itself.","tokens_in":8225,"tokens_out":6045,"duration_ms":61251,"concrete_test":"Take the three-compartment model in Eqs. 2–4 with fixed k_met>0, k_ex>0, and a known analytic solution. Implement the proposed loss as the sum of Eqs. 5–6, with C_NN as a neural ODE and f as an independent network, and train with no additional data-fitting term. Check whether the learned trajectories reproduce the analytic solution for all t. If the loss minimum is infeasible or trajectories deviate systematically (e.g., total drug mass remains constant when it should decay), the objective is incoherent. Then add a supervised term matching simulator outputs and correct L_phys to use remaining, not administered, drug mass; verify whether convergence to the analytic solution is restored, indicating the framework is salvageable only after revision.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section IV defines the training objective for the Neural ODE-enhanced PBPK model. Eq. 5 sets L_ODE = (1/N) Σ || dC_NN(t_i)/dt − f(t_i, C_NN(t_i), θ) ||^2, where f is itself a neural network parameterized by θ. This is a self-consistency constraint: it is minimized when the derivative of C_NN equals the f-network evaluated at C_NN, regardless of whether either expression matches the actual PBPK equations (2)–(4). No supervised term comparing C_NN to trajectories from the [14] simulator is presented, so nothing forces the learned dynamics to reflect true radiopharmacokinetics. Eq. 6 then imposes L_phys = || Σ_i C_NN,i(t) V_i − D ||^2, with D the total administered dose. But Eqs. 2–4 include liver metabolism (k_met L) and renal excretion (k_ex K), so total drug mass in the compartments decays over time; equating it to a constant D at all times contradicts those ODEs. Enforcing both losses simultaneously would force k_met = k_ex = 0 or require D to become time-dependent, which is not stated. Since §V.A routes the MDP transition probabilities through this same Neural ODE, the RL policy optimization inherits these biases. The framework's central promise therefore rests on a training objective that is not well-posed as written and has no demonstrated mechanism to fit PBPK data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reviews challenges and current practices in radiopharmaceutical therapy (RPT) and theranostics, and proposes a data-driven decision-support framework that combines Neural Ordinary Differential Equations (Neural ODEs) and Physics-Informed Neural Networks (PINNs) with Physiologically Based Pharmacokinetic (PBPK) models, integrated with reinforcement learning (RL) for personalized dosing. Sections II and III survey dosimetry methods and open problems; Section IV presents a PBPK model with three compartments and defines a training loss; Section V formalizes an MDP and outlines RL-based policy optimization. The manuscript contains no implementation, simulation results, or patient data; it is a framework proposal with a literature review.","tokens_in":8635,"tokens_out":3428,"duration_ms":35814,"significance":"The topic is timely and clinically relevant, and the review of RPT dosimetry and PBPK modeling is useful for readers entering the field. If the proposed framework were made technically sound and validated on synthetic or clinical data, it could contribute to precision oncology decision support. However, the paper currently offers only a conceptual outline with no demonstrated feasibility. The training objective in Section IV is internally inconsistent, and the MDP transition model is underspecified, so the central promise of optimized, patient-specific dosing is not supported as written. The survey component has merit, but the proposal's technical foundations need substantial revision.","major_comments":[{"comment":"","section":"Section IV, Eqs. (5) and (6)"},{"comment":"","section":"Section V.A and Algorithm 2"},{"comment":"","section":"General (Sections IV-V)"}],"minor_comments":[{"comment":"","section":"Algorithm 1, line 2"},{"comment":"","section":"Section II.B, Eq. (1)"},{"comment":"","section":"Section IV, Abstract and Introduction"},{"comment":"","section":"Section V.A"},{"comment":"","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is best viewed as a perspective/review with a proposal. The technical core (the training loss and MDP transition model) is currently ill-posed and needs fixing before the contribution can be assessed. I would advise the editor to request a major revision that either (a) focuses the paper as a pure review and clearly labels the proposal as speculative, or (b) adds a proof-of-concept experiment and repairs the loss-function inconsistency. The literature survey is competent, but the proposal's lack of validation and internal inconsistency prevent acceptance in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the first half of the paper is a competent survey of RPT dosimetry and RL evaluation, and the overall direction—Neural ODEs for PBPK plus RL for dosing—is a legitimate research question. The problem is in the details of the proposed loss: as written, the training objective in §IV is internally inconsistent, and the stress-test note is right.\n\nWhat the paper does well: it gives a useful snapshot of current RPT agents, dosimetry tools, PBPK models, and RL metrics, with relevant citations to both the clinical and computational literature. It correctly identifies real clinical pain points (one-size-fits-all dosing, dosimetry workflow burden, the gap between population models and individual patients) and points to the [14] simulator as a natural testbed. The review of RL evaluation—IQR, CVaR, performance profiles—is a solid checklist for anyone moving into this area.\n\nWhere it falls short: the core proposal is not yet mathematically well-posed. Eq. 5 only enforces that the derivative of the neural output matches the f-network at the same state; there is no supervised term comparing against simulator trajectories, so nothing ties the learned dynamics to the PBPK equations. Eq. 6 then asks that the sum of compartment doses equal a constant D at all times, but Eqs. 2–4 include liver metabolism and renal excretion, so total drug mass decays. Both losses can only be satisfied simultaneously if k_met = k_ex = 0, which is not stated and would make the PBPK model meaningless. Since the MDP transition probabilities are routed through this same Neural ODE, the RL stage inherits the error. The sim-to-real question (training on [14], deploying on patients) is also waved at rather than argued.\n\nThe reader's 'unverdictable' is appropriate—there is no measured result—but this becomes a question of whether the proposal is coherent enough to guide future work. Right now it isn't, but it's fixable in a revision with a corrected loss and arguably a demo on synthetic data.\n\nI'd send this to peer review at a venue that takes position papers, because the survey and the research agenda are valuable and the authors are clearly engaged with the literature. I'd expect major revision, not acceptance as-is.","headline":"A creditable survey of RPT decision support, but the proposed framework's training equations are internally inconsistent and need major revision before the research agenda can be taken seriously.","tokens_in":8994,"tokens_out":4248,"would_cite":false,"duration_ms":44144,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes that combining Neural ODEs with PBPK models gives reinforcement learning a simulated training ground for patient-specific radiopharmaceutical dosing.","keywords":["theranostics","radiopharmaceutical therapy","precision oncology","physiologically based pharmacokinetic models","neural ordinary differential equations","physics-informed neural networks","reinforcement learning","personalized dosimetry"],"falsifier":"Train the proposed Neural ODE on outputs of the PBPK simulator and test it against measured time-activity curves from a cohort of Lu-177-DOTATATE or Lu-177-PSMA patients. If predicted organ and tumor absorbed doses deviate from clinically measured dosimetry by more than the tolerance used for treatment planning (typically 10–20%), the sim-to-real premise fails and the framework's policies cannot be trusted.","tokens_in":8062,"feed_emoji":"⚛️","tokens_out":6759,"duration_ms":64817,"temperature":0.7,"pith_summary":"This paper argues that the path to personalized radiopharmaceutical therapy runs through combining data-driven modeling with reinforcement learning. It proposes a framework in which Neural Ordinary Differential Equations and physics-informed constraints upgrade Physiologically Based Pharmacokinetic models into patient-specific simulators, and a reinforcement learning agent uses those simulators to learn dosing policies that balance tumor control against radiation exposure to healthy organs. The contribution is a prospective architecture and a research agenda, not a validated system: no experiments, clinical data, or trained policies are reported. A sympathetic reader would take the paper as setting out what must be built and which design choices are worth testing.","feed_headline":"Neural ODEs could power patient-specific radiopharmaceutical dosing","feed_subtitle":"A proposed pipeline trains an RL agent on PBPK-simulated bodies, then tunes it to each patient's imaging data.","key_machinery":"The load-bearing mechanism is the Neural ODE-augmented PBPK model used as the transition model of a Markov Decision Process. A Neural ODE is a neural network that represents the time derivative of a state vector, here the drug concentrations $P$, $L$, $K$ in plasma, liver, and kidneys, trained against the ODE system of compartmental mass transport and a physics-based loss that enforces conservation of total administered dose. The paper discretizes time, randomly initializes the network parameters, and optimizes them until the predicted concentrations match the PBPK dynamics; this learned dynamics then supplies the transition probabilities for a reinforcement learning agent whose actions are dosing decisions. The MDP's state and reward structure (time-integrated activity, absorbed doses, tumor control reward, organ-at-risk penalty) is what turns dosimetry into a sequential decision problem.","core_discovery":"On the paper's own terms, the central claim is that a Markov Decision Process whose transition probabilities are supplied by a Neural ODE-enhanced PBPK model of plasma, liver, and kidney compartments can support reinforcement learning of optimal radiopharmaceutical dosing policies. The state is defined by time-integrated activity and absorbed doses in tumor and organs at risk, actions are dosing regimens, and the reward function encodes both tumor control and organ-at-risk safety. The authors propose to train this system offline on data generated by a reaction-graph PBPK simulator, then refine it online as patient imaging data arrive. The paper is explicit that this is a proposal; the discovery, if it is one, is that the components exist and can be assembled this way.","pith_inferences":["The paper does not test sim-to-real transfer; the natural next experiment is to fit the proposed Neural ODE on simulator outputs and compare its absorbed-dose predictions against measured time-activity curves from Lu-177-DOTATATE or Lu-177-PSMA patients before any reinforcement learning training.","The same Neural-ODE-as-transition-model pattern could be transplanted to other sequential oncology decisions, such as chemotherapy scheduling or adaptive radiotherapy, wherever a differentiable compartment model exists.","If the learned transition model is imperfect, offline reinforcement learning methods with conservative policy constraints would be a safer training choice than unconstrained DQN or DDPG, since the agent would avoid overestimating states it has never seen.","A quantitative test of the framework's value would be an in silico comparison between the reinforcement-learning-optimized policy and standard fixed dosing on the simulator's ground-truth outcomes, measuring the difference in tumor control probability per unit organ-at-risk dose."],"forward_implications":["If Neural ODEs reproduce PBPK dynamics from simulated data, reinforcement learning policies can be trained entirely in silico before any patient exposure, so the first clinical use is a fine-tuning step rather than an exploration step.","The MDP formulation converts clinical imaging measurements (time-integrated activity and absorbed doses) directly into decision variables, giving dosimetry a formal role in treatment policy rather than an after-the-fact check.","Because the reward function separates tumor control from organ-at-risk dose, the approach makes the dose-safety trade-off explicit and tunable for each patient.","The physics-informed loss term could reduce the amount of patient-specific data needed to fit the dynamics, because the model is constrained to conserve administered dose instead of learning that from scratch.","Offline training on historical and simulated cases followed by online refinement maps naturally onto the way theranostic cycles accumulate imaging data, so policies can improve cycle by cycle."],"supporting_citations":[{"why":"Supplies the PBPK simulator used to generate the high-fidelity datasets for training and evaluating the Neural ODE framework.","marker":"[14]"},{"why":"Provides the Neural ODE and machine-learning-in-PBPK background that the proposal builds on.","marker":"[31]"},{"why":"Defines the Theranostic Digital Twin concept that motivates patient-specific simulation and continuous treatment refinement.","marker":"[18]"},{"why":"Establishes radiopharmaceutical therapy dosimetry and its clinical challenges, motivating the need for personalized dosing.","marker":"[9]"},{"why":"Gives the dosimetry equations (absorbed dose from time-integrated activity and S-factors) used to define the MDP state.","marker":"[10]"},{"why":"Supports the treatment-policy-learning formulation for multiobjective outcomes, the basis of the proposed reward structure.","marker":"[32]"},{"why":"Lists real-world reinforcement learning challenges that frame the evaluation metrics and the practical implementation discussion.","marker":"[34]"},{"why":"Provides action-elimination for continuous action spaces, referenced for handling the high-dimensional dosing action space.","marker":"[36]"}],"fun_headline_variants":["RL-trained radiopharmaceutical dosing via neural ODEs","Patient-specific dosing plans from RL and neural ODEs","Neural ODEs + RL for optimal theranostic dosing","Sim-to-real RL for radiopharmaceutical therapy dosing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"A neural network trained on computer-simulated pharmacokinetic data is assumed to predict how real patients will distribute a radiopharmaceutical closely enough that dosing policies learned from the simulator are safe to use in the clinic.","fun_headline_variants_meta":{"raw":{"variants":["RL-trained radiopharmaceutical dosing via neural ODEs","Patient-specific dosing plans from RL and neural ODEs","Neural ODEs + RL for optimal theranostic dosing","Sim-to-real RL for radiopharmaceutical therapy dosing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1144,"prompt_tokens":832,"completion_tokens":312,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":448,"completion_tokens_details":{"reasoning_tokens":244}},"tokens_in":448,"tokens_out":312,"duration_ms":3742,"temperature":1.0,"reasoning_tokens":244,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:20:29.303621+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the proposed Neural ODE on outputs of the PBPK simulator and test it against measured time-activity curves from a cohort of Lu-177-DOTATATE or Lu-177-PSMA patients. If predicted organ and tumor absorbed doses deviate from clinically measured dosimetry by more than the tolerance used for treatment planning (typically 10–20%), the sim-to-real premise fails and the framework's policies cannot be trusted.","supporting_citations":[{"cited_title":"Fele-Paranj, B","cited_arxiv_id":null,"evidence_quote":"Supplies the PBPK simulator used to generate the high-fidelity datasets for training and evaluating the Neural ODE framework."},{"cited_title":"Machine learning and artificial intelligence in physiologically based pharmacokinetic modeling,","cited_arxiv_id":null,"evidence_quote":"Provides the Neural ODE and machine-learning-in-PBPK background that the proposal builds on."},{"cited_title":"Theranostic digital twins for personalized radiopharmaceutical therapies: Reimagining theranostics via computational nuclear oncology,","cited_arxiv_id":null,"evidence_quote":"Defines the Theranostic Digital Twin concept that motivates patient-specific simulation and continuous treatment refinement."},{"cited_title":"Radiopharma- ceutical therapy in cancer: clinical advances and challenges,","cited_arxiv_id":null,"evidence_quote":"Establishes radiopharmaceutical therapy dosimetry and its clinical challenges, motivating the need for personalized dosing."},{"cited_title":"Radiopharmaceutical dosimetry in targeted radionuclide therapy,","cited_arxiv_id":null,"evidence_quote":"Gives the dosimetry equations (absorbed dose from time-integrated activity and S-factors) used to define the MDP state."},{"cited_title":"Treatment policy learning in multiobjective settings with fully observed outcomes,","cited_arxiv_id":null,"evidence_quote":"Supports the treatment-policy-learning formulation for multiobjective outcomes, the basis of the proposed reward structure."},{"cited_title":"Challenges of real-world reinforcement learning: definitions, benchmarks and analysis,","cited_arxiv_id":null,"evidence_quote":"Lists real-world reinforcement learning challenges that frame the evaluation metrics and the practical implementation discussion."},{"cited_title":"Learn what not to learn: Action elimination with deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Provides action-elimination for continuous action spaces, referenced for handling the high-dimensional dosing action space."}],"review_version":1}