acceptodds
Withdrawn from ICLR 2027

TREDL-Spec: Temporal Replicate-Aware Disentangling for Longitudinal Clinical Spectroscopy

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Longitudinal spectroscopy can provide non-invasive monitoring of biological change, but repeated measurements introduce variability from both temporal state and measurement variation. Existing representation learning methods have separately explored static–temporal factorization, known relationships between observations, and measurement-specific variation, but these complementary sources of information have not been jointly leveraged to structure learned representations. This gap is particularly consequential for clinical spectroscopy, an increasingly explored but still comparatively underdeveloped area of biomedical AI, where substantial measurement variability can be difficult to distinguish from true biological change and can compromise the reliability and clinical trustworthiness of learned representations. We introduce TREDL-Spec (Temporal Replicate-Aware Disentangling for Longitudinal Spectroscopy), a replicate-aware framework that factorizes each spectrum into subject-stable, temporally evolving, and measurement-specific representations. Nested invariance constraints exploit repeated measurements (replicates) to explicitly associate these representations with their intended sources, while temporal dynamics are learned separately from measurement-specific nuisance variation. We evaluate the framework on controlled synthetic spectroscopy data and longitudinal human Raman spectroscopy data. TREDL-Spec achieved the highest SAP and DCI disentanglement on the synthetic dataset and the highest DCI on the human dataset, while its most predictive latent dimensions provided the strongest subject and temporal factor predictability on both datasets. TREDL-Spec also achieved the lowest future-spectrum prediction error across all evaluated metrics. These results demonstrate that explicitly structuring representations around biological and measurement-specific sources can support longitudinal spectral prediction while providing an explicit factorization of measurement variability.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
spectroscopy, Raman spectroscopy, longitudinal representation learning, disentangled representation learning, biomedical machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_d66uOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TREDl-Spec, which addresses the issue of disentangling patient-specific, temporal, and measurement variation in longitudinal spectroscopy acquisitions. The proposed method combines regularization constraints of each variation in the latent space and provides competitive performance against two baseline models.

Critical strengths
  • The paper addresses an important but underexplored topic: disentangling variation sources from longitudinal spectroscopy acquisitions.
  • The proposed method is experimented on real human longitudinal spectroscopy data, which is difficult to collect.
Critical weaknesses
  • The extensiveness of experiments can be improved.
  • The performance gain of the proposed method is not clearly convincing
  • There exists a gap between the background motivation and the implementation of TREDL-Spec. The method combines multiple predictions and regularization constraints on the latent space, but the claim of 'disentanglement' of variations is not well supported empirically or theoretically.
Questions for authors
  • Extensiveness of experiments: please consider adding more baseline methods for disentangled representation learning, and show that the proposed method has a clear advantage over them.
  • Please consider elaborating the discussion on the minimal performance gain against current baselines (CaRiNG and TopDis).
  • Figure 3b does not seem convincing that the latent space is well disentangled. Please further justify the qualitative result.
  • Please further justify, or tone-down, the motivation - implementation relationship claim. The target objectives are often naïve, and no empirical evaluation (e.g., ablation studies) supports the disentanglement claim.
  • It is unclear what 'nested' means in the context. Please consider adding an explanation or definition of the term.
Reviewer_6naZOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces TREDL-Spec, a representation learning framework for longitudinal clinical Raman spectroscopy that uses paired same-visit technical replicates across two timepoints to factorize spectral measurements into subject-stable (z^S), temporally evolving (z^T), and measurement nuisance (z^N) latent subspaces via replicate invariance losses, gradient-reversal adversarial training, and future clean-spectrum prediction.

Critical strengths

The paper tackles a novel and practically compelling problem in clinical spectroscopy: exploiting repeated technical replicates across longitudinal patient visits to disentangle true temporal biological state from static inter-subject variability and measurement nuisance. Because real-world clinical measurements are heavily confounded by patient baselines and probe-level acquisition noise, explicitly modeling this hierarchical replicate structure is a creative and well-motivated formulation with clear relevance to longitudinal biomedical monitoring.

Critical weaknesses

My concerns of this paper focus around the weak experimentation validating the claims present within the paper, as well as the lack of proof, methodology or experimental, that the core disentanglement claims are supported.

  1. No ablation study

There is no ablation study within the paper, which is a major weakness. The architecture is fairly involved and there seems to be 5 different loss functions (reconstruction, future clean prediction, subject invariance, temporal invariance, and adversarial subject cross-entropy) and yet there is no ablation used to demonstrate the importance of them, and the writing makes this fairly unclear that there are that many loss functions. On line 292, it is stated that “Hence, the spectral autoencoder is trained using a weighted combination of clean-prediction loss and reconstruction loss” but the weights are never explained or ablated. In addition to this, the ablations would have been necessary to understand the effects of the other components, such as latent factorization, dilated convolution blocks, the adversarial classifier, setting z^N = 0 during prediction, etc etc…

  1. No or Unclear validation of the actual factorization

The paper proposes to factorize the latent space into subject-stable characteristics, temporally evolving biological state, and measurement-specific nuisance variation. It is argued that disentanglement was measured with MIG, SAP, DCI were metrics used to calculate the disengagement. However, they are all concatenated together for the calculation…. So it is unclear how they are disentangled across groups. In addition to this, Table 1 and 2 do not demonstrate which subspace the coordinate corresponds to. To validate the paper's core claim, the authors must report a full 33 subspace probing matrix showing the predictability of (Subject, Time, Measurement) from zS alone, zT alone, and zN alone.

  1. The methodology does not seem to enforce information leakage across specific latent states.

The methodology does not prevent information leakage across the specific latent subspaces. While the paper claims to factorize spectral variation into three distinct semantic subspaces—subject-stable (z^S), temporally evolving (z^T), and measurement nuisance (z^N)—the training losses in Sections 3.3–3.6 only enforce one-way invariance and leave multiple pathways for cross-subspace leakage completely unconstrained. First, nothing prevents static subject identity from leaking into the temporal subspace z^T: because the temporal invariance loss L_T = ‖z^T_(1,a) - z^T_(1,b)‖² + ‖z^T_(2,a) - z^T_(2,b)‖² (Section 3.3, line 257) only compares same-timepoint replicates from the same subject, any static subject-specific feature is already identical across replicates (1,a) and (1,b), meaning z^T can freely encode subject identity with zero penalty in L_T and still minimize the future prediction loss D(z^S₁, ẑ^T₂, 0) (line 305). Second, nothing prevents temporally evolving biological state from leaking into the nuisance subspace z^N: the gradient-reversal adversarial classifier on z^N (Section 3.5, lines 281–287) only penalizes predictability of subject identity, not timepoint t, so z^N can absorb temporal variation to help minimize the individual reconstruction loss D(z^S_(1,a), z^T_(1,a), z^N_(1,a)) (line 314). Third, the predicted future temporal latent ẑ^T₂ = F(z^T₁, t₁, t₂, Δt) (Section 3.4, line 272) is only supervised through the decoder output in spectral space, with no latent-space consistency loss (such as ‖ẑ^T₂ - z^T₂‖²) to guarantee that ẑ^T₂ even remains aligned with the encoder's temporal subspace Z^T. Because the evaluation in Tables 1–2 only reports the single most predictive scalar coordinate across the entire 96-dimensional concatenated vector [z^S; z^T; z^N] rather than probing each subspace separately, there is no theoretical or empirical guarantee that the three subspaces actually isolate their intended factors.

Questions for authors

see weaknesses above

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

This paper proposes TREDL-Spec, a repeat-measurement-aware disentangled representation learning framework for longitudinal clinical spectral data, aiming to disentangle subject-stable features, time-varying biological states, and measurement-specific variations. The framework uses a shared encoder to decompose spectra into three latent representations and introduces nested invariance constraints based on repeated measurements at the same time point and observations across different time points, along with adversarial learning to reduce information leakage between latent spaces. The model also employs a Temporal Dynamics Network to predict future biological state representations and jointly optimizes spectral reconstruction and future spectral prediction. Experiments on synthetic and real-world longitudinal Raman spectra demonstrate that TREDL-Spec generally outperforms CaRiNG and TopDis in disentanglement metrics and future spectral prediction accuracy.

Critical strengths
  1. The paper leverages repeated measurements from the same subject at the same and different time points to construct nested invariance constraints that guide the learning of distinct latent representations. This design effectively exploits the structure of longitudinal spectral data to disentangle subject-stable features, temporal variations, and measurement-specific variations.

  2. The paper evaluates the proposed framework on both synthetic and real-world spectral data, comparing it with CaRiNG and TopDis. The high DCI disentanglement scores on both datasets and the lowest future spectral prediction errors provide preliminary evidence of the framework's effectiveness.

Critical weaknesses
  1. Unclear connection between the research motivation and downstream tasks. The paper aims to disentangle different sources of variation in longitudinal spectra through decoupled representation learning, but the specific scientific or clinical problem this approach is intended to address remains unclear. The model is primarily evaluated on future spectral prediction, yet the authors do not adequately explain the practical value of predicting complete future spectra or why this task requires decoupled representations. Moreover, since future spectral prediction is included in the training objectives, prediction performance does not provide independent evidence of the effectiveness of the learned representations.
  2. Insufficient validation of the decoupling claim and limited generalizability. The paper divides spectral representations into three semantically distinct latent spaces, but the experiments do not sufficiently demonstrate that each captures its intended factors of variation. The metrics in Table 1 are computed from the concatenated representation and therefore cannot directly validate the semantic roles of individual subspaces. Similarly, the t-SNE visualization in Figure 3 shows clustering only for the subject-stable representation, without directly validating the other two subspaces. Furthermore, the experiments are limited to synthetic Raman data and a single real-world clinical Raman spectral setting, leaving the generalizability of the framework to other spectral scenarios unclear.
  3. Unclear organization and insufficient experimental analysis. The paper presents its main experimental results and analysis in Section 3.7 of the "Methods" section rather than in separate "Experiments" or "Results" sections, mixing methodological descriptions, experimental setup, and interpretation of results. Moreover, the analysis focuses primarily on performance comparisons, with limited systematic discussion of baseline selection, the contributions of key modules, or how the findings support the paper's central claims. This weakens the clarity and persuasiveness of the overall argument.
Questions for authors

In addition to the weaknesses above, I have the following questions regarding the model design and experimental validation:

  1. How do the existing constraints ensure semantic separation among the three latent spaces? The proposed nested invariance constraints encourage consistency across related observations, while the adversarial classifier discourages subject identity information in the nuisance representation. However, these mechanisms do not appear to explicitly prevent information leakage across the three latent subspaces. For example, (z^T) may still encode subject-specific information, while (z^N) may retain temporal information. Could the authors clarify how the current objective function prevents such information leakage?
  2. Have ablation studies been conducted on individual latent components and constraint terms? Have the authors evaluated the necessity of each latent component and its associated constraints? For example, how do model performance and disentanglement change when a particular latent component or its corresponding loss term is removed? Could controlled experiments also verify that each component responds to its intended factors of variation, rather than merely showing that removing a module degrades overall performance?
  3. Does the Temporal Dynamics Network actually rely on the learned temporal representation? Since the Temporal Dynamics Network directly incorporates measurement time, the model may primarily learn population-level spectral changes from the time variables rather than relying on the learned temporal representation (z^T). The authors should provide further evidence to establish that (z^T) makes a meaningful contribution to future-spectrum prediction.

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

Related papers

Discussion (0)

Sign in to comment.