acceptodds
Withdrawn from ICLR 2027

Real-Time Causal Dynamics on Geometric Latent Spaces for Precision Therapeutic Monitoring

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

18 views

Abstract

Despite impact in early-stage drug discovery, the exploration of artificial intelligence in therapeutic drug monitoring, a key bottleneck in patient care, has remained limited. We present Real-Time Causal Geometric Flows (RT-CGF), a geometry-aware, causal, and adaptive framework for therapeutic drug monitoring of narrow-therapeutic-index drugs, addressing a longstanding gap in translating from discovery to deployment. Unlike traditional models that rely on fixed, population-level dynamics, RT-CGF redefines TDM as a dynamical system on learned Finsler manifolds, where pharmacokinetic trajectories evolve as geodesics under patient-specific geometric constraints. This previously unexplored formulation enables real-time adaptation to physiological shifts (e.g, acute kidney injury) while preserving causal invariance. Extending stochastic intervention calculus for continuous-time counterfactuals and persistent homology regularization for topological consistency, RT-CGF achieves fewer dosing errors and improved AUC in adverse event detection on representative clinical datasets. Beyond its methodological novelty, RT-CGF introduces an interesting possibility for precision medicine by harmonizing Finsler-geometric dynamical systems, counterfactual reasoning, and persistent homology constraints, a convergence that transcends the limitations of static, population-level models to enable adaptive, patient-specific policy frameworks with implications for safety, assurance and regulation efforts for digital drug development.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Therapeutic Monitoring, geometric flows, Finsler geometry, pharmacology

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_7eaHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper propose RT-CGF as real time causal geometric flows. It is a framework for therapeutic drug monitoring. It combines transformer Finsler dynamic with orthogonality penalty and regularization for persistent homology. The experiments showed improved prediction and adverser-event detection on 8k patients cohort.

Critical strengths
  1. RT-CGF is a dynamic, patient-specific and geometrically grounded model. It models Finsler geometry and stochastic intervention calculus.
  2. The experiments evaluated vancomycin and tacrolimus as two clinically relevant drugts with different clearance mechanism.
Critical weaknesses
  1. The presentation of this paper is hard to follow. An overview figure is missing and key mathematical claims such as the strong convexity of the learned metric and the causal invariance induced by the orthogonality constraint, also lack sufficient proofs. Equations are not numbered, making it hard to refer to specific equation during review.

2, The overall loss function (Page 5, Line 243) has three important hyper-parameters as weights. But there is no sensitivity experiments for these hyper-parameters.

  1. The construction (Page 3, Line 129) doesn't guarantee a valid Finsler metric: because (A) depends on (v), this expression is not quadratic. So its velocity Hessian need not be positive definite. PLease provide relevant proof.
Questions for authors

Please see Weaknesses section and consider each point as a question.

Reviewer_h7S6OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose RT-CGF, a framework that models drug dynamics by learning a dynamic geometric law with topological constraints to preserve causal invariance of intervention. The method learns dynamics on a Finsler manifold instead of traditional Euclidean space to take into account specific behaviors of asymmetrical dynamics in drug monitoring. They use a Temporal Geometric Transformer to model the evolution, and take into account irregular time intervals and longe range interactions. The TGT models the current pharmacy kinetic state, then a CMIT models the effects of endogenous interventions like drug adjustment on the system dynamics. It factorises state and causal intervention on dynamics. Four losses enforces correct dynamics fit, topological, causal and biological constraints. RT-CFG is evaluated on two datasets, one synthetic and one real-world. The method outperforms considered baselines on trajectory prediction, and show increased robustness to noisy measurements and counterfactual accuracy.

Critical strengths
  • The set of baseline is well chosen and covers all the aspects of the methodology
  • The method adopt interesting approaches for drug kinetic predictions and provide natural constraints on the learned dynamics
  • The framework is outperforming existing baselines, showing a potential gap in what literature provides
Critical weaknesses
  • The causal claim is too strong and not backed up correctly. The orthogonality is a relaxed structural constraint, implied through a loss without guarantees of convergence. For example, the claim “do-calculus-inspired regularizer enables first-order causal reasoning under continuous-time” is thus too strong without mathematical proofs and guarantees.
  • Several properties of the models relies on losses, but the work would benefit from more careful analysis of the sensitivity of the method to losses. The authors mostly test presence or absence of losses, but not the impact of the lambda parameters scaling them.
  • Figures are overlooked and not very informative. Figure 3 lacks confidence interval or uncertainty estimation metrics. None of the figures are commented or even mentioned in the text. The work would strongly benefit from visual description of the pipeline and analysis of the figures to help the reader understand the relevance of results.
  • The work would strongly benefit from external validation, or out-of-distribution test in synthetic data for example.
  • Mentioned drawbacks on Bayesian methods was that they rely on priors, but RT-CFG also rely on prior and expected behavior or the modeled trajectories in the PHR
Questions for authors
  • How sensitive is the model to loss parameters tweak, as you rely on them for your claim on causality, biological plausibility and topology?
  • What takes the remaining 70% of training time if not PH or Transformer blocks?
  • It is not clear if the resulting smoothness of shapes is due to Finsler geometry or PH. Could the monotonic decay in RT-CGF be induced by the geometrical constraint rather than topological? Maybe comparing the quality of curves (qualitatively and quantitatively) with the same model but without the topological loss would be informative.

Suggestions for improvement:

  • Reposition the paper on Finsler manifolds for pharmaco kinetics rather than causality aspects
  • Add experiments around losses that are critical in this setup
  • Improve figures
  • Provide a deeper introduction to the methods covered, from pharmaco kinetics to Finsler geometry. Readers would really benefit from it, especially in conferences like ICLR where the audience may not be familiar with these concepts
Reviewer_Bv4jOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses therapeutic drug monitoring for drugs with a narrow therapeutic window, specifically vancomycin and tacrolimus, where dosing must adapt to sparse blood-level measurements and to sudden physiological changes such as acute kidney injury. The authors propose RT-CGF, a continuous-time model with three components. A transformer parameterizes a patient-specific Finsler metric so that concentration trajectories follow geodesic flows on a learned manifold, intended to capture clearance that depends on the direction and speed of change. A cross-modal intervention transformer models dose changes inside a stochastic differential equation, regularized by a constraint that the intervention act orthogonally to the column space of the drift’s Jacobian, which the authors present as a continuous-time form of causal invariance. A persistent homology loss compares each predicted trajectory’s Vietoris–Rips persistence diagram with a prior built from expert-annotated trajectories, intended to enforce plausible shapes such as monotonic decay and to flag adverse events. Counterfactual dosing is posed as minimizing a Finsler distance to a target trajectory. The method is evaluated on a synthetic dataset and on a clinical cohort of 5,000 ICU vancomycin patients and 3,000 transplant tacrolimus patients, against eleven baselines.

Critical strengths

The paper targets a clinically important problem that has received little attention in machine learning. Dosing narrow-therapeutic-index drugs under acute physiological change is a real gap in current practice, and the experimental design includes a capacity-controlled Transformer–Neural ODE backbone and a Riemannian variant, which is the right way to isolate each component’s contribution. These points do not change my assessment, because the components themselves do not hold as defined.

Critical weaknesses

Each of the three components fails at the role it was introduced to fill. The metric is defined as F2=z˙⊤A(z,z˙)⊤A(z,z˙)z˙F^2 = \dot{z}^\top A(z, \dot{z})^\top A(z, \dot{z})\dot{z}, with positive definiteness of gijg_{ij} said to hold by construction and convexity said to follow “for this quadratic form,” but once AA depends on z˙\dot{z} the expression is not quadratic in z˙\dot{z}, and its Hessian includes derivatives of AA that can make gijg_{ij} indefinite or singular, while the ODE requires g−1g^{-1}. Positive homogeneity of degree one also needs AA to be unaffected by rescaling z˙\dot{z}, for example homogeneous of degree zero, which the described bias-free pathway does not establish. In the causal constraint, the Jacobian of the nn-dimensional drift with respect to the 2n2n-dimensional state generically has rank nn, so its column space is Rn\mathbb{R}^n, and the only vector orthogonal to it is zero. The penalty therefore drives ∂h/∂π\partial h/\partial \pi toward zero and removes the dose’s first-order effect, consistent with the paper’s own description (lines 179 to 181) that the intervention may act only in directions to which the drift is insensitive. The topology loss builds a Vietoris–Rips filtration on trajectory points, which depends only on the set of points and not on their time order, so a decaying trajectory and its time reversal have the same persistence diagram, and the loss cannot enforce monotonic decay, the property the paper says it enforces. Since each component is presented as a contribution, these issues remove the basis for the paper’s main claims.

The empirical evidence cannot be verified and contradicts itself. The clinical cohort has no named institution, database or citation, and the ethics statement does not identify the approving review board. In Table 4, the Transformer–Neural ODE without geometry or the intervention term reaches an MSE of 0.34, while “No Finsler + No SIC” reaches 0.71, although the first removes at least as much, and removing persistent homology raises counterfactual accuracy from 0.91 to 0.92. In Figure 1, the normal-clearance trajectory falls only from about 15.0 to 14.1 mg/L over 24 hours, a half-life of more than a week, and the clearance axis ends at 0.04 L/h, whereas vancomycin and tacrolimus have half-lives of hours and clearances on the order of liters per hour. Together with the issues above, the reported gains cannot be attributed to the proposed components.

Questions for authors

Clarifications:

  1. How is AA constrained so that FF is a valid Finsler metric, and how is gijg_{ij} guaranteed to be invertible during integration?
  2. Which nonzero ∂h/∂π\partial h/\partial \pi satisfies the orthogonality constraint when the drift Jacobian has full row rank?
  3. How does a Vietoris–Rips filtration on trajectory points penalize non-monotonic concentration curves?
  4. What is the source of the clinical cohort, which review board approved its use, and how do the two rows of Table 4 differ in configuration?

Suggestions for future work:

A comparison against MAP Bayesian forecasting with a population PK model would establish whether the method improves on clinical practice. Releasing the code and the synthetic data generator would allow the results to be verified.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.