acceptodds
Under review as a conference paper at ICLR 2027

When Are Non-Invasive Medical Digital Twins Identifiable? Tracing Clinical Identifiability via a Cascaded Lens

Abstract

Non-invasive medical digital twins seek to infer patient-specific physiology from indirect clinical observations. However, distinct latent physiological states and model parameters may produce indistinguishable measurements, leaving their interpretation ambiguous even when a twin accurately reproduces the observed data. We study clinical identifiability: whether physiological states, parameters, or clinically prescribed summaries are uniquely determined by the available observations. Specifically, we formulate non-invasive digital twins as cascaded dynamics that separate upstream physiology from downstream observation processes. This structure yields a low-order sufficient criterion for local observability and structural identifiability, certifying recoverability from a small collection of Lie-derivative blocks without constructing the full observability-identifiability matrix. We further combine admissible reduction with functional identifiability to assess when clinically meaningful physiological summaries remain recoverable despite ambiguity in the full latent state. Experiments on synthetic cascades and a cerebral hemodynamics digital twin constructed from perfusion MRI evaluate low-order certification, computational cost, and the stability of reduced physiological targets across observationally similar solutions. Together, these analyses provide a principled basis for matching model resolution, observation pathways, and clinical inferential targets to the information supported by non-invasive measurements.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Non-invasive medical digital twins, functional identifiability, cerebral hemodynamics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_qjQjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The goal of this paper is to investigate how to determine if non-invasive medical digital twins are identifiable. More specifically, the authors examine whether hidden physiological states, model parameters, and clinically relevant quantities can be uniquely determined based upon indirect measurements. To address this issue, the authors use a cascade of dynamical systems which separates the "upstream" physiological dynamics from the "downstream" measurement process.

The primary theoretical contributions are a low-order sufficient condition for local observability and structural identifiability that uses specific Lie derivative blocks, and a functional identifiability framework which integrates admissible model reduction with a row-space condition to recover clinical quantities without requiring full-state reconstruction.

The authors evaluate their proposed approach on two types of simulated systems, namely linear and nonlinear cascaded systems, and a digital twin of cerebral hemodynamics built from the ISLES2017 perfusion MRI dataset. The experiments assess computational efficiency, structural ambiguity, and the consistency of regional cerebral blood flow (CBF), cerebral blood volume (CBV), and mean transit time (MTT) among different observationally similar solutions.

Critical strengths

The authors address a significant issue: an accurate reproduction of clinical observations does not guarantee that the actual physiological states or parameters are uniquely identified. The cascaded formulation is a logical approach to separate the dynamic processes involved in physiology from those of measurement. It is beneficial to distinguish between full-state identifiability and functional identifiability because obtaining clinically relevant information can be more practical than obtaining all of the hidden state variables.

The proposed low-order block criterion provides an easily understood sufficient condition for local observability and structural identifiability. The accompanying synthetic experiments explore computational efficiency and demonstrate how individual parts of a cascaded system become observable as the order of their derivatives increases. A cerebral perfusion application demonstrates that the theoretical perspectives presented here may be relevant to physiological modeling.

Critical weaknesses

There is insufficient evidence to support the claim that cerebral blood flow and cerebral blood volume are functionally identifiable, and this claim also appears inconsistent with the model's scaling structure.

In Section 4.2 on the cerebral hemodynamics model, the authors present three different components: 1) a regional effective inflow; 2) an upstream outflow state; and 3) a downstream vascular volume scale. Both target functions CBF and CBV are defined using the effective flow amplitude and volume scale.

Equations (4.1) and (4.3), combined with the parameterization of Appendix E.3, seem to allow for a continuous scaling transformation. A continuous scaling transformation can be constructed by multiplying the effective inflow, upstream outflow, and vascular volume by the same positive constant, while inversely scaling the resistance parameter that determines the initial outflow and leaving the relaxation time and tracer input unchanged. This appears to result in no change in the predicted tracer concentration trajectory. While the individual CBF and CBV values will be different under such a transformation, the ratio between the two (and hence MTT) will remain constant.

The above appears to indicate that neither CBF nor CBV can be uniquely identified on its own through the reported observation model unless there is some additional means of fixing the scale or an independent measurement. The problem is even more significant since the low row-space residual values reported for both quantities in Table 2 are presented as numerical evidence for functional identifiability, yet the apparent scaling ambiguity makes it difficult to reconcile these results.

The fact that the authors used regularization and physiological knowledge to initialize their model does not imply that they have established that the parameters can be uniquely identified based upon the observations of the system alone. This distinction between identifying parameters uniquely from observations (structural identifiability) and using prior assumptions to select one of several possible sets of parameters that are consistent with the observed behavior of the system (prior-based parameter selection) is fundamental to the authors' research questions.

The experimental validation is inadequate to demonstrate that the theoretical criteria are applicable in practice.

Theorem 3.4 offers a low-order certificate under restrictive constraints on the measurement dimension and rank. More specifically, the constraint m >= max(n_D, n_U + q_D, q_U) for the number of observable channels may be difficult to satisfy when considering the non-invasive applications that motivate this work. The primary synthetic examples employ a scalar downstream observable and evaluate trajectory-stacked rank diagnostics, which are explicitly distinct from the pointwise sufficient conditions of Theorem 3.4. Thus, these experiments mainly illustrate cases where low-order certification is not achieved and demonstrate computational savings, rather than establishing how frequently the theoretical certificate applies to realistic medical digital twins.

A large part of the computational comparison is based on the use of selected low-order blocks or a staged procedure, as opposed to constructing a predetermined higher-order observability-identifiability matrix. If there were a direct comparison with an existing incremental rank-testing procedure, it would offer more compelling evidence of the practical computational advantage.

Additionally, the cerebral hemodynamics experiments were conducted with a compartmentally reduced version of the original spatially dynamic system. The fitting procedure utilized regularization priors and initializations derived, in part, from regional perfusion summary statistics. Because of the close association between the summary statistics and the target quantities of interest, the observed stability among the retained minimizers cannot independently demonstrate that the target quantities are functionally identifiable. The current experiments would have been much more useful if they had separated the contributions of observational data, model reduction, initialization, and regularization.

Questions for authors

Can the authors please explain what prevents the following scaling transformation in the cerebral hemodynamics model? The scaling transformation would involve multiplying the effective flow amplitude, upstream outflow, and vascular volume by a common positive factor and inversely scaling the resistance parameter which determines initial outflow, while leaving the relaxation time and tracer input unchanged.

How can the reported row-space residuals be reconciled with the claimed functional identifiability of CBF and CBV if the tracer concentration trajectory is preserved under the transformations described in Equations (4.1) and (4.3), and yet CBF and CBV are changed individually? Are there other observations or hard constraints placed on the parameters in the implementation of the model?

Are the authors able to explain how the regional CBF and CBV summaries used for both initialization and regularization are calculated, as well as whether these overlap with the "silver-standard" estimates presented in Figure 4? I think it would be very helpful if the authors repeated the functional stability analysis with significantly wider ranges of initial parameters and without target-informed regularization. This would enable us to separate structural identifiability from numerical stability provided by prior information.

In regard to the cerebral hemodynamics model, would the authors provide information regarding the dimensions of the observed outputs, known inputs, upstream/downstream states, and unknown parameters? Would the authors also state whether the conditions required by Theorem 3.4 are satisfied? Additionally, could the authors describe how the proposed methodology compares to a standard incremental observability-rank method where the algorithm stops when it has found sufficient rank, without constructing a higher-order matrix in advance?

Since regional compartmental dynamic models are based upon an approximate spatial aggregation of voxels, can the authors assess the effect of this approximation on the identifiability of the original physiological quantities? Specifically, what provides assurance that the clinical target functionals of the aggregated model retain their intended interpretation when applied to the original voxelwise system?

3 comments

Thank you for identifying the scaling symmetry and suggesting an early-stopping baseline. Please see our Global Response for an overview.

W1 and W2 / Q1–Q3. What does the tracer identify, and what role do the Figure 4 reference maps play?

Q1. We agree that absolute CBF and CBV can change while the tracer trajectory and MTT remain identical. In Eqs. (4.1) and (4.3), multiplying flow and volume by s>0s>0 and dividing resistance by ss scales the entire outflow by ss at fixed relaxation time and tracer input. To verify this, we doubled flow and volume and halved resistance in all 651 reference-free regional fits. Every tracer value and MTT was preserved exactly. The transformed parameters remained within the fitting bounds in 618 regions, and any interior fit admits sufficiently small rescalings.


Q2. The low residuals in the submitted Table 2 (revised Table 3) are conditional on flow and volume information. We added a clarification in Section 4.2.2 explaining how the reference-derived priors anchor the fitted scale and how retaining flow and volume information in the diagnostic distinguishes points along the scaling family. To verify the latter, we changed only the observation information at the same reference-informed fits in 43 patients, keeping input amplitude fixed. The mean residuals were below 0.001% for both targets with flow and volume retained. With the tracer alone, they rose to 61.4% for CBF and 61.8% for CBV.


Q3. Yes. The fitting anchors came from the same silver-standard maps displayed in Figure 4. Within each ROI, the medians of positive CBF and CBV map values set the initialization centers, penalty centers, and parameter bounds. We have removed the caption’s description on these maps "used only for visualization” to avoid potential confusion.

We further test how much agreement persists without these anchors: we repeated the fits with tracer-derived centers and wider flow and volume bounds. We retained the original 20-start schedule (from one-third to three times each center at its widest setting). With identical starts and bounds, removing the four soft penalties increased mean pairwise discrepancy from 12.8% to 220.2% for CBF and from 27.2% to 122.9% for CBV. The median patient-level MTT discrepancy rose from 16.2% to 83.7%. A separate four-patient check with common starting centers for all ROIs and zero penalties also gave large CBF and CBV discrepancies (236.4% and 192.6%). Table 12 in revised Appendix H reports the cohort ablation. We derived the tracer-based centers following Tran et al. (2026).

The tracer-only analysis gives partial support for local MTT identifiability at the individual-region level. In the separate seven-parameter compartment analysis, all fitted parameters were free. Mean ROI residuals were 2.3% for MTT, 58.3% for CBF, and 58.7% for CBV (revised Table 4). The MTT residual fell below 1% in 96.9% of individual ROIs and 34.9% of joint patient evaluations.

W2 / Q4. Does Theorem 3.4 apply in the clinical model, and how does the cascade compare with incremental rank testing?

We tested Theorem 3.4 pointwise and compared it with the incremental rank test as suggested. On the original 18 systems, all five direct-channel systems passed the low-order certificate at order two. Five flat-coupling systems reached full rank at orders three through seven, and eight scalar-aggregate systems remained uncertified through order seven. Incremental OI adds one derivative order at a time and stops at full rank. Both methods used the same differentiation engine, caching policy, rank threshold, evaluation points, and order cap. Villaverde et al. (2019)

The cascade was 2.31× faster on the original suite against early stopping. The original 14.8× speedup used a fixed order-seven baseline. Revised Table 2 and Appendix G report median total times and median paired speedups over six runs.

SettingIncremental OI (s)Cascade reusing T1T_1 (s)Speedup
Original 18 systems0.3350.1482.31×
Polynomial cascades, all 2K2K downstream outputs0.5160.1284.04×
Same cascades and points, KK outputs1.5241.5860.99×

Both methods reached the same decision at the same order in all 30 evaluations. With all downstream outputs, both stopped at order two. At K=8K=8, reusing T1T_1 to compute U2U_2 reduced distinct partial derivatives from 160 to 56. With half the outputs, the dimension check triggered incremental testing, and all six points reached full rank at order four. These counts describe the tested readouts. Prevalence across clinical model families remains open.

We remark that, Theorem 3.4 can fail to certify full recovery in clinical models with many latent states or unknown parameters and few measured outputs. This is the case in our perfusion model, where each region provides a single tracer channel (m=1<nU+qDm=1<n_U+q_D). We therefore introduce functional identifiability to ask which clinical targets the observations can still determine when full recovery cannot be certified (Theorem 3.9). To test the effect of fewer outputs directly, we halved them from 2K2K to KK on the same polynomial cascades. All six points still reached full rank, now at order four instead of order two (Appendix G).


W2 / Q5. How well do the regional targets represent the original voxelwise quantities?

Fitting after spatial averaging changed CBF by 3.94%, CBV by 1.05%, and MTT by 8.25% relative to averaging the voxel fits. In four patients, we fitted the same 16 voxels in each of 16 regions per patient both ways (1,024 voxels), holding equations, input, optimizer, and prior centers fixed. For MTT, the voxelwise comparator was the mean of the individual transit times. The fitted tracer trajectories differed by 4.60%. We took medians over regions and then over patients. This pilot quantifies the approximation in the sampled regions. Extending the identifiability guarantee to voxelwise physiology requires compatible regional dynamics and a regional target that preserves the intended voxelwise quantity.

AuthorsOpenReview ↗

Gentle reminder

Dear Reviewer,

As the rebuttal period draws to a close, we would welcome any further feedback on our responses and revisions. In response to your concerns, we added verification of the scaling symmetry, reference-free diagnostics and regularization ablations, comparisons against incremental OI, and an analysis of the effect of spatial averaging.

We hope these additions help address your concerns and would be happy to discuss any remaining questions. Thank you again for your time and careful review!

Reviewer_upyeOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines the identifiability of non-invasive medical digital twins from indirect clinical measurements. The authors model digital twins as cascaded dynamical systems and derive low-order rank conditions for assessing structural identifiability. They also introduce an admissible-reduction framework to characterize when clinically relevant quantities can be identified even if the underlying physiological states and model parameters are not uniquely determined. Experiments on synthetic dynamical systems and cerebral perfusion MRI show that the proposed approach reduces computational costs relative to constructing full observability matrices and improves the stability of regional estimates of cerebral blood flow (CBF), cerebral blood volume (CBV), and mean transit time (MTT).

Critical strengths
  • solves a useful problem related to the relationship between reconstruction and states.
  • appears to be efficient
Critical weaknesses
  • the paper should more precisely distinguish its theoretical contributions from existing results on structured observability, functional observability, and exact lumping
  • theorem 3.4 may be unrealistic for non-invasive medical systems with relatively few observation channels and many latent variables. The authors should characterize how frequently this condition holds in practical settings and clarify the applicability of the proposed criterion when it fails.
  • synthetic experiments do not establish demonstrating successful certification in identifiable systems
  • The connection between admissible reduction and practical identifiability is not fully established. The empirical evaluation relies on spatial averaging and compartmental closure, which are approximate rather than exact reductions. Since Theorem 3.9 requires an admissible reduction with compatible dynamics, it remains unclear whether the theoretical guarantees extend to the original voxelwise physiological system.
  • ablations needed to confirm the benefit of the proposed reduction
  • no results on noisy real world data scenarios
Questions for authors

see weaknesses for follow up experiments

7 comments

Thank you for your comments and questions. Please see our Global Response for an overview. We address the six weaknesses in the order listed in your review.

W1. How do the theoretical contributions differ from structured observability, functional observability, and exact lumping?

  • Difference from original structured observability. The general observability–identifiability (OI) rank condition tells us when the full matrix certifies local identifiability, but leaves open which derivative blocks suffice for our nonlinear cascade. Theorem 3.4 exploits the fact that the upstream dynamics evolve independently, the downstream dynamics receive upstream information, and the output is y=g(xD)y=g(x_D). Under these assumptions, full rank of three blocks through order two is sufficient. We also added a comparison with a stronger baseline, incremental OI, which stops as soon as it reaches full rank. With eight upstream states and 16 outputs, both methods stopped at order two, yet our test evaluated only 56 distinct partial derivatives compared with 160 for incremental OI (Appendix G).
  • Difference from original functional observability. The classical row-space criterion tests whether a given observation determines a target locally. Here, we consider a clinical problem in which the observation is invariant to scale, yet some quantities of clinical interest depend on that scale. In our perfusion model, flow and volume can scale together with a compensating change in resistance, leaving the tracer curve unchanged. A flow or volume reference can resolve this ambiguity in absolute CBF and CBV. MTT is preserved under the same rescaling. To test whether the tracer still constrains MTT when all seven fitted parameters can vary, we evaluated its tracer-only row-space residual (Table 4).
  • Difference from original exact lumping. Exact lumping asks whether the reduced coordinates have closed dynamics. We further ask whether the reduced observation trajectory uniquely determines a clinical target (Theorem 3.9). In our clinical model, spatial averaging is the step that replaces voxelwise dynamics with a regional approximation. To test how much this changes the targets themselves, we compare regional estimates of CBF, CBV, and MTT with the averages of voxelwise fits in 43 patients (W4), where small approximation errors support using the regional targets as proxies for voxelwise averages when assessing functional identifiability at the regional level.

W2. When does Theorem 3.4 apply with few observation channels, and what happens when it fails?

We tested the effect of fewer measurements by changing only the readout on three polynomial cascades with K=2,4,8K=2,4,8 upstream states. Each system has 2K2K downstream states, 5K5K augmented coordinates, and a lower bound of m=2Km=2K measured outputs for Theorem 3.4. We evaluated both readouts at the same two points per system.

Table I. Certification with full and partial downstream observation.

Measured outputsEvaluationsCertified by low-order blocksFirst full-rank order
All downstream states, m=2Km=2K662
Half the downstream states, m=Km=K604

With half the outputs, the dimension check sent all six evaluations directly to incremental rank testing. Each reached full rank at order four, the earliest possible order because derivatives through order rr supply (r+1)K(r+1)K rows for 5K5K coordinates.

For the clinical application, we can answer the applicability question directly. Each region has one measured tracer channel, one upstream outflow state, and an unknown downstream volume. Thus m=1<nU+qDm=1<n_U+q_D in every region. Section 4.2 accordingly evaluates individual clinical targets using the functional criterion of Theorem 3.9. Estimating how often the low-order condition holds across different clinical model families would require a broader sample.

W3. Do the synthetic experiments demonstrate successful certification in identifiable systems?

Yes. We evaluated the pointwise rank conditions on all 18 systems from Appendix F.2. All five direct-channel systems passed the low-order test at order two. Five flat-coupling systems reached full rank only at orders three through seven. The eight scalar-aggregate systems remained rank deficient through the order-seven cap.

The polynomial examples in Table I above add positive cases with unknown parameters in both subsystems. With all downstream outputs observed, the cascade and an incremental baseline both certified full rank at order two, and the cascade was 4.04× faster. Against the same early-stopping baseline, the speedup over the original 18-system suite was 2.31×.


W4. The connection between admissible reduction and practical identifiability is not fully established.

Our functional identifiability derivation assumes an exactly admissible reduction, in practice, approximation errors may still exist. Here, we have conducted an additional experiment to quantify the level of identifiability practically, achieved through our proposed admissible reduction.

Specifically, on all 43 of ISLES stroke patients, we compute the error introduced when a single regional fit approximates the average of the voxelwise fits. We further tested how much mean voxel MTT could vary within the measured tracer error.

Table II. Regional approximation error.

Regional approximation error relative to voxelwise fitsTracerCBFCBVMTT
Median within patients, then across patients3.41%4.84%0.52%7.25%

These small differences support using the regional fit to estimate mean voxel CBF, CBV, and MTT under our fitting assumptions. In 78.8% of 642 evaluable ROIs, a 20% change in mean voxel MTT stayed within the measured tracer discrepancy and increased the observed tracer loss by at most 2%.


W5. Ablations needed to confirm the benefit of the proposed reduction

We sampled 16 voxels from each of 16 regions in each of four patients (1,024 voxels). In one arm, we fitted the individual voxel curves and averaged their predictions and target estimates. In the other, we fitted the mean of exactly those curves. Both arms used the same compartment equations, input, regional prior centers, and optimizer. Each gap is relative to the voxel-first result. We took the median across regions within each patient, then across patients.

Table III. Relative gap between fitting the regional mean and averaging voxel fits.

Fitted tracer trajectoryCBFCBVMTT
4.60%3.94%1.05%8.25%

Volume was least sensitive to the order of operations. For MTT, the regional estimate differed by 8.25% from the mean of the individual voxel transit times. The two procedures had similar relative errors against the measured mean tracer curve (32.6% for the regional fit and 33.0% for the voxel-first fit). This pilot quantifies the approximation for the sampled regions. Establishing a gain in physiological accuracy would require an independent target measurement.


W6. How do the estimates behave on noisy real world observations?

We clarify that, in Section 4.2 of the original submission, we evaluated our model on measured tracer curves from ISLES 2017, a real-world dataset of clinical MRI scans from patients with acute ischemic stroke (Winzeck et al., 2018).

Reviewer_upyeOpenReview ↗

Follow up Q's

Thank you for the rebuttal and additional experiments. The revisions address some of my original concerns, particularly the incremental rank-testing baseline and positive certification examples. However, a few questions remain despite the rebttal.

  • The authors now acknowledge that its sufficient conditions are not satisfied in the clinical application. Can they identify a realistic physiological setting where the proposed low-order certificate applies and provides a computational advantage over incremental rank testing?

  • Why does MTT stability deteriorate substantially without regularization if MTT is functionally identifiable?

  • How can identifiability guarantees transfer to the original voxelwise system when the reduction is only approximate?

Thank you again to the authors for their additional experiments, but would appreciate any follow up between now and the discussion deadline

Reviewer_CMygOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the identification conditions for digital twin models where the authors assume that some analytical and latent physiological model is responsible for indirect clinical measurements, and the authors focus particularly on "non-invasive" measurements.

The main contribution of the paper is a formulation of the digital twin as a "cascaded" dynamical system and deriving a low-order rank based criteria for evaluating the standard conditions of local observability and structural identifiability used in other dynamic systems. They also further introduce the notions of admissible model reduction and functional identifiability to determine whether clinically relevant information can be identifiable even when full-state recovery is impossible.

To demonstrate these ideas, the authors conducted experiments on synthetic dynamical systems and cerebral perfusion MRI data, and the experiments demonstrate computational savings from the proposed low-order analysis. In terms of the practical implications of this work, I think that the main point is that these identifiability conditions can be used to guide the selection of model complexity for digital twins since digital twin models can vary in their level of abstraction and number of parameters.

Critical strengths

The problem of whether digital twins are identifiable is very relevant because there is currently a lot of excitement about applying them in medicine but a recurring issue with the validity of these models is whether they are identifiable.The paper addresses this problem in a rigorous way and makes some interesting connections to ideas in dynamic system modeling.

Critical weaknesses

(1) One of the main claimed contributions of the paper is viewing non-invasive medical digital twins as "cascaded systems". However, this does not seem to be any different from the composite functions formulation in (Kuang et al., 2024) which the authors cite. The paper also seems to blur the distinction between cascaded systems and state space models. I am not a control theorist but my understanding is that cascaded systems are a more restrictive type of state-space model, yet the authors' proposed model assumes a more general state space model (see lines 130-133). My concern is that the main theoretical contributions seem to leverage results on cascaded systems and I am not sure if the authors are overgeneralizing these results to the more general state space models.

Another (related) conceptual issue is that the paper puts a lot of emphasis on the idea of "non-invasive" digital twins being something special that needs a new model (See lines 40, 47 and 101), and this is used as a motivation for the proposed mathematical model. However, I am not sure why the measurements being non-invasive or invasive has any relevance to the proposed model. The authors seems to be just using a dynamic state space model which is applicable to any noisy measurement regardless of whether it is obtained through invasive or non-invasive means.

(2) Given that the digital twin is viewed as a classical dynamic state space model, it is not clear to what extend are the theoretical results new, and to what extend is there anything special about the digital twin problem compared to the standard dynamic system identification problem. For example, Theorem 3.9 seems to restate restate the standard idea that a function of the parameters is locally identifiable if its gradient lies in the span of the observable gradients. There is a huge body of work that cover similar conditions (e.g. see N Meshkat, "An algorithm for finding globally identifiable parameter combinations of nonlinear ODE models using Gröbner Bases", 2009.) and also the idea of an admissible reduction overlaps heavily with existing work on identifiable reparametrization.

The overlap with prior work is not surprising because the authors reduce digital twins to a dynamic system so this is equivalent to the general question of identifying that system. But I think the more important gap in our knowledge about medical digital twins (that is also peculiar to medicine) is whether we can identify the structural causal model underlying this system from observational data when there are some interventional variables involved. This is a technically different question from identifying a dynamic system and is the real gap in our knowledge because the main use case of digital twins models is to conduct virtual simulations of how the system will change under interventions (e.g. the LVAD example in (Kuang et al., 2024)). Because the paper overlooked this question, I believe the two main theoretical results are restating known results about dynamical systems.

(3) There are also some possible technical errors in the current submission:

  • Please check if the perfusion model is actually structurally non-identifiable in CBF and CBV. I have not been able to verify that for Eqs. (4.1) and (4.3).

  • Theorem 3.4 dpoes not seem to be applicable in the non-invasive regime of the paper because it requires m ≥ max{n_D, n_U + q_D, q_U}, i.e. at least as many outputs as upstream states plus downstream parameters, and as many as upstream parameters. But aren't the non-invasive settings characterized by few outputs as in the synthetic cascade example and the examples in Appendix F?

Questions for authors

I encourage the authors to respond to the two weaknesses raised above.

3 comments

Thank you for your comments and questions. Please see our Global Response for an overview. We address the numbered weaknesses below.

W1(a). Distinction between the paper's cascaded systems and general state space models.

The key distinction between our cascaded system and the general state-space models lies in the constraints imposed by the cascaded structures: Theorem 3.4 relies on one-way coupling and a readout that depends only on the downstream state.

We revised Section 3.1 to distinguish these assumptions from the general state-space formulation in Section 2.2, which allows feedback between state components and a readout that depends on the full state. In the cascade of Definition 3.1, upstream physiology drives the downstream dynamics without feedback. An upstream change therefore reaches the measured signal only through the downstream state. This cascaded structure is what makes the low-order rank criterion possible.

W1(b). How does the cascade differ from (Kuang et al., 2024), and what role does non-invasive measurement play?

(Kuang et al., 2024) learns to estimate patient-specific parameters from indirect measurements. Whilst in this work, we ask when those measurements uniquely determine the parameters locally. Specifically, Kuang et al. (2024) formulate estimation as a composite inverse problem linking a physiological simulator to a measurement map. We build on that formulation by giving the measurement process its own evolving state, as clarified in Section 3.1. Its dynamics let us trace how changes in physiological parameters propagate into the observed trajectory.

W2. Which theoretical results go beyond existing identifiability and reparameterization results, and what do they imply about interventions?

(a) What does Theorem 3.9 add to existing work on identifiable parameter combinations? We agree that the row-space criterion for local functional identifiability is classical. We address a complementary question: given a clinical target and a proposed reduction, when do the reduced observations locally determine that target? Under the admissible reduction, Theorem 3.9 applies the classical criterion to the specified target.

(b) Can observational data determine how a twin will respond to an intervention? The LVAD example in Kuang et al. (2024) fixes a pump model and compares the same cardiac twin with and without pump support. We added an analogous perfusion experiment with a further identifiability check: can two twins agree on the entire pretreatment observation yet predict different responses to the same intervention?

In Appendix I, our added recanalization experiment gives a counterexample: two perfusion twins reproduce the same entire pretreatment tracer curve and baseline flow, yet predict different responses to the same intervention. The twins differ only in the fraction of total resistance attributable to the clot (25% versus 50%). We simulate removal of 0%, 50%, or 100% of each twin’s clot resistance. Complete removal increases predicted steady regional outflow by 33% in one twin and 100% in the other, relative to their common baseline.

This establishes that the recanalization response is not identifiable from pretreatment tracer concentration and baseline flow in this model. The missing quantity is specific: baseline flow depends on total resistance, whereas the intervention acts on the share attributable to the clot. More generally, when models consistent with the pretreatment data predict different responses to the same intervention, identifying the response would require additional evidence or independently justified assumptions that resolve the ambiguity.

W3(a). Verify if the perfusion model is actually structurally non-identifiable in CBF and CBV.

Through forward simulation, we verified analytically that absolute CBF and CBV are structurally non-identifiable from the tracer curve alone. In Eqs. (4.1) and (4.3), scaling inflow, outflow, and vascular volume together preserve the tracer trajectory. Absolute CBF and CBV therefore vary across observationally equivalent solutions, while MTT has the same value along this scaling family. The low CBF and CBV residuals in Table 2 are consistent with this result.

As specified in the added analysis in Section 4.2.2, that diagnostic includes static flow and volume readouts, which distinguish solutions at different scales. In the original fits, the silver-standard maps also guide initialization and regularization, so agreement across fits can reflect the scale favored by those maps.

To further test what the tracer supports without using the silver-standard maps to select a scale, we repeated the fits in 43 patients and 651 ROIs using tracer-derived flow and volume estimates for initialization, following Tran et al. (2026). We used the same initialization and bounds in both reference-free conditions and switched the soft penalties on or off. Our new main-text table reports the tracer-only row-space residual and the discrepancy between retained fits:

Table I. Effect of regularization on reference-free perfusion estimates. Values are mean ± SD across patients (%).

TargetResidual, penalties onResidual, penalties offDiscrepancy, penalties onDiscrepancy, penalties off
CBF58.9 ± 1.671.1 ± 6.512.8 ± 20.7220.2 ± 516.4
CBV60.3 ± 2.779.6 ± 8.827.2 ± 41.7122.9 ± 55.6
MTT29.9 ± 33.573.8 ± 25.5346.6 ± 2047.3677.8 ± 3701.5

Removing the penalties substantially increased disagreement in absolute CBF and CBV. The patient-level MTT residual also rose from 29.9% to 73.8%, so invariance to the common scaling alone does not establish MTT identifiability. We therefore tested MTT separately in the reference-free compartment analysis in the Appendix H, allowing all seven fitted parameters, including input amplitude, to vary in the tracer Jacobian. The MTT residual was below 1% in 96.9% of individual ROIs and in 34.9% of patients when their ROIs were evaluated jointly. These results provide support for local functional identifiability of MTT, strongest when each region is evaluated individually.

W3(b). What happens when there are too few outputs for Theorem 3.4?

We tested your concern by halving the number of measured outputs on the same three polynomial cascades (K=2,4,8K=2,4,8 upstream states), keeping the dynamics and evaluation points fixed. With all 2K2K downstream states measured, the low-order blocks certified full rank at order two at all six points. With only KK measured, the dimension requirement failed at every point, yet higher-order testing recovered full rank at order four at all six. This is the earliest possible order for KK outputs and 5K5K augmented coordinates. The clinical model likewise falls below the low-order threshold (m=1<nU+qDm=1<n_U+q_D), which is why Section 4.2.2 evaluates individual targets using Theorem 3.9.

We also strengthened the runtime comparison. Table 4 in Appendix F.2 reported a 14.8× speedup on the original 18 systems against a baseline that always constructs the matrix through order seven. We reran those same systems against Incremental OI (Villaverde, Tsiantis, and Banga, 2019), which adds one derivative order at a time and stops as soon as it reaches full rank. Against this early-stopping baseline, the cascade was 2.31× faster on the original suite and 4.04× faster on the polynomial cascades with all outputs.

SettingIncremental OICascade blocks reusing T1T_1Paired speedup
Original Appendix F.2 suite, 18 systems0.335 s0.148 s2.31×
Polynomial cascades, all 2K2K outputs0.516 s0.128 s4.04×
Same cascades and points, only KK outputs1.524 s1.586 s0.99×

Times are median totals over six paired runs. Speedups are medians of the paired ratios. Both methods reached the same decision at the same order in all 30 evaluations. With all outputs, the advantage came from reusing the first-order block T1T_1 to compute U2U_2: at K=8K=8, this reduced the number of distinct partial derivatives evaluated from 160 to 56, even though both methods stopped at order two. With half the outputs, the dimension check sent the cascade directly to incremental testing, and runtimes were comparable.

AuthorsOpenReview ↗

Gentle reminder

Dear Reviewer,

As the rebuttal period draws to a close, we would welcome any further feedback on our responses and revisions. In response to your concerns, we added clarifications on the scope of our cascade formulation and the CBF/CBV scaling ambiguity, along with experiments with fewer outputs and comparisons against incremental rank testing.

We hope these additions help address your concerns and would be happy to discuss any remaining questions. Thank you again for your time and careful review!

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your careful reviews. Your questions prompted us to test when the cascade saves computation and which perfusion quantities the tracer measurements can determine. We have marked all the revisions in blue.

Major Revisions

  • New comparison against a stronger rank-testing baseline (Table 2, Appendix G). Our original baseline always built the observability–identifiability matrix through order seven. We added incremental rank testing, which builds the matrix one derivative order at a time and stops as soon as it reaches full rank. Both methods reached the same decision at the same order in all 30 evaluations. The cascade was 2.31× faster on the original 18 systems and 4.04× faster with all downstream outputs observed, where reusing derivative blocks saved work even though both methods stopped at order two. With half the outputs, both reached full rank at order four with comparable runtimes. [CMyg, upye, qjQj]

  • Reference-free cerebral perfusion analysis (Section 4.2.2, Table 4, Appendix H). Across 43 patients and 651 regions, we verified that absolute CBF and CBV can change while the tracer curve remains identical. The tracer-only MTT residual fell below 1% in 96.9% of individual regions and 34.9% of joint patient evaluations, providing partial support for local MTT identifiability at the regional level. [CMyg, upye, qjQj]

  • Intervention counterexample (Appendix I). Two twins matched the entire pretreatment tracer curve and baseline flow. Complete clot removal nevertheless increased predicted steady outflow by 33% in one twin and 100% in the other because the clot accounted for different shares of total resistance. [CMyg]

Additional Clarifications

  • Section 3.1 clarifies the one-way coupling assumed by the low-order test. Corollary D.4 specifies when adding prior information lowers the row-space residual. [CMyg, upye]
  • The individual responses report results on spatial averaging and sensitivity to added noise in measured perfusion curves. [upye, qjQj]

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.