acceptodds
Under review as a conference paper at ICLR 2027

Modeling Longitudinal Clinical Dynamics with Covariance-Aware Neural Kalman Filtering for Multimodal Survival Prediction

Abstract

Cancer prognosis reflects both baseline molecular characteristics and clinical evolution during follow-up. We propose a covariance-aware neural Kalman framework that integrates baseline RNA-seq and demographic covariates with irregular longitudinal clinical observations. A neural observation encoder maps clinical visits into a latent state space, where explicit covariance propagation and an analytically computed Kalman gain govern sequential state updates. Elapsed-time-dependent process covariance incorporates irregular visit intervals into the correction mechanism. Independently trained molecular and clinical predictors are combined through risk-level fusion selected by inner cross-validation. We evaluate the framework on 556 patients with 95 observed events from the MMRF CoMMpass multiple myeloma cohort using a 365-day landmark design and five repeated nested cross validation partitions. RNA + Covariance Kalman achieves the highest mean pooled out-of-fold C-index among the evaluated models, 0.716 ± 0.013, compared with 0.662 ± 0.028 for the static molecular model and 0.702 ± 0.012 for RNA + GRU. Its one- and two-year post-landmark AUCs are 0.743 ± 0.019 and 0.740 ± 0.021, respectively, with a two-year Brier score of 0.121 ± 0.004. Ablations support the contributions of covariance-derived correction and elapsed-time-dependent process covariance. These results demonstrate the value of covariance-aware state estimation for integrating molecular risk with longitudinal clinical trajectories in this cohort.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
multimodal survival analysis, RNA-seq, Kalman network, state-space model, MMRF CoMMpass

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Rc9UOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a neural Kalman framework that integrates RNA-seq data, demographic information, and longitudinal clinical records to predict survival risk. It uses a neural encoder to map irregular clinical visits into a state space, with explicit Kalman covariance propagation. This results in a clinical Cox risk, which is then combined with the molecular Cox risk learned from the RNA-seq data and static covariates to make the prediction. However, the paper has several issues, including insufficient experimental validation and the lack of innovation in its methods.

1.The model was validated using a cohort of only 556 patients, with just 95 recorded deaths; given the limited scale of the experiment and the lack of data diversity, the results are not particularly convincing.

  1. Recurrent Kalman Networks (RKN)(Becker et al., ICML 2019) is a general-purpose approach to temporal modeling; its mechanisms for neural observation encoding, latent state covariance propagation, and Kalman updates are highly relevant to the present work. However, it was not discussed in the paper.

  2. The performance comparisons focus primarily on Static Molecular and RNA+GRU. The comparison is rather limited. RNA+RKNs and more closely related methods, like pretrained sequence encoders, contrastive sequence learning methods, should also be included in the discussion and in the evaluation and comparison.

  3. The static molecular branch employs an ensemble of two MLP Cox predictors with differing regularization strengths; however, the motivation for this approach and its implementation details are not sufficiently addressed.

  4. Although the fusion coefficients \lambda and \alpha are chosen based on the inner-layer OOF predictions, the selection procedure, and more importantly, the effectiveness of using such a simple linear combination method, remain unclear.

Critical strengths

A neural encoder is used for the sequential state estimation, mapping irregular clinical visits into a low-dimensional clinical-state space with Kalman covariance propagation.

Critical weaknesses
  1. The motivation of using a Kalman state estimator (instead of the popular methods such as Transformers) is not clearly stated. It is also unclear why 'covariance-aware' is crucial for this case. 2.The molecular Cox risk and the clinical Cox risk are learned independently. It's not clear how each of them and their linear combination contributes to the prediction.
  2. Some key technical details are missing, including the two compact MLP Cox predictors, E(.) in Eq.(5), the static representation s_i and the observation mask m_it, etc.
Questions for authors

What's the point of being 'covariance-aware', and why can't they be learned by other state estimators?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Rc9U

We thank the reviewer for the insightful feedback and for highlighting the sequential state-estimation design of our clinical encoder. To clarify the scope of our contributions more precisely: our work centers on scenario-specific methodological design and systematic empirical validation for sparse, irregular clinical survival prediction, rather than proposing a revision to general Kalman filtering theory. Specifically, we couple learned clinical representations with explicit elapsed-time-dependent covariance propagation and analytical Kalman correction, and systematically verify its advantages over mainstream sequence models, state-space baselines, and learned-gate alternatives under a rigorous leakage-controlled evaluation framework. Below we respond to each specific concern in detail.

Q1. Cohort scale and result generalizability

A1.Thank you for your valuable comments. We would like to provide context for the cohort scale and additional evidence supporting the reliability of our findings. First, the MMRF CoMMpass study is the largest and most comprehensively profiled public cohort for multiple myeloma with matched longitudinal clinical records and RNA-seq data. The 95 observed death events in our 365-day landmark subset are consistent with the typical scale of deep learning survival studies using this cohort and landmark design. For Cox proportional hazards modeling with moderate-dimensional predictors, this event count provides sufficient statistical power for the within-cohort comparative analyses we present. Second, our evaluation design is constructed to minimize overfitting and maximize result reliability: All preprocessing, feature selection, and hyperparameter tuning are strictly confined to training folds, with no data leakage into evaluation partitions. We use five repeated nested cross-validation runs with stratified sampling to ensure stable performance estimates. All performance gains are consistent across repeated runs and validated through paired statistical tests and patient-level bootstrap analyses.

Q2.Recurrent Kalman Networks and expanded sequence comparisons

A2. We thank the reviewer for this constructive suggestion. In the original submission, our main comparisons focused primarily on Static Molecular and RNA+GRU. In the revised manuscript, we have expanded Section 2 (Related Work) to discuss Recurrent Kalman Networks (RKNs) and related sequence representation learning paradigms, and we have substantially broadened our empirical benchmark in Appendix E (Table 7) and Appendix H (Table 9) to evaluate RNA+RKN alongside a time-aware Transformer and two learned-gain neural filtering baselines under the exact same nested-CV protocol.

Core innovation in combining neural encoding with explicit Kalman filtering, and distinction from RKNs: RKN (Becker et al., ICML 2019) is indeed a strong and representative neural state-space model that combines deep observation encoders with factorized latent Kalman updates. While sharing the high-level spirit of marrying neural networks with state-space estimation, our core innovation lies in how we couple neural representation learning with explicit continuous-discrete Kalman covariance propagation specifically for irregular clinical survival prediction:

Neural clinical encoding with explicit time-aware covariance propagation: We use a neural observation encoder to map heterogeneous, partially observed clinical visits into a compact latent state space, where process uncertainty grows explicitly with the elapsed inter-visit interval rather than relying on locally linear transition bases learned for dense sequences as in RKN.

Sample-efficient inductive bias for sparse clinical trajectories: By delegating nonlinear feature extraction to the neural encoder while governing temporal state updates through analytical covariance recursion, our model achieves high sample efficiency on sparse clinical histories (median 4 visits per patient).

Empirical comparisons against RKN and additional sequence models: Across five matched nested-CV repeats under identical clinical inputs, observation masks, and late-fusion rules:

Comparison with RNA+RKN (Appendix E, Table 7): RNA+CovarianceKalman achieves a mean pooled OOF C-index of 0.7191, outperforming the time-aware RKN adaptation (0.6901) across all five repeats with a mean paired improvement of +0.0290.

Comparison with Time-Aware Transformer and learned-gain neural filters (Appendix E & Appendix H): RNA+CovarianceKalman also consistently outperforms RNA+Transformer (0.7160 vs. 0.6990, +0.0169 mean paired gain, winning all 5 repeats, Table 9) as well as learned constant-gain (0.7079, +0.0112 gain) and learned time-dependent-gain (0.7035, +0.0156 gain) variants (Table 7).

AuthorsOpenReview ↗

Response to Reviewer Rc9U

Q3.Motivation for covariance-aware state estimation and why it cannot be easily learned by other state estimators

A3. This is a central question of our work, and we thank the reviewer for pushing us to articulate this motivation more sharply. The necessity of explicit covariance-aware state estimation stems from the structural mismatch between longitudinal clinical follow-up records and general-purpose sequence encoders:Why general sequence models struggle to learn temporal weighting from scratch:Longitudinal clinical trajectories differ fundamentally from dense natural language or sensor streams: each patient has only a handful of visits (a median of 4 visits in our cohort), separated by highly irregular time gaps (14 to 218 days) with frequent missing measurements. Standard sequence models (such as Transformers and GRUs) or unconstrained learned-gate estimators attempt to infer temporal decay and measurement reliability implicitly through attention weights or gating parameters. Under moderate clinical sample sizes and right-censored survival supervision, learning these uncertainty-weighting dynamics purely from scratch is data-inefficient and prone to overfitting.What being "covariance-aware" provides mechanically:Our framework embeds a well-founded physical inductive bias directly into the latent state recursion rather than asking a network to rediscover it from sparse data:Time-aware prior uncertainty expansion: Process uncertainty grows monotonically with the elapsed interval between visits, which expands the predicted prior covariance after long follow-up gaps and formalizes the clinical reality that a patient's unobserved state becomes less certain over longer intervals.Analytical uncertainty-calibrated correction: The Kalman gain is computed analytically from the ratio of prior prediction covariance to measurement covariance R. Consequently, when a new clinical visit arrives after a long gap, the expanded prior uncertainty automatically assigns higher weight to the incoming observation; conversely, when visits are closely spaced, the filter smooths noisy measurements against the prior state without requiring extra learnable gating parameters.Empirical evidence that other state estimators cannot replicate this behavior as effectively:We verified this advantage through three sets of controlled comparisons under identical clinical inputs, observation masks, normalized visit times, and inter-visit intervals:Against attention and recurrent sequence encoders (Appendix A & Appendix H, Table 9): In the clinical-only setting, CovarianceKalman achieves a C-index of 0.6616 compared with 0.6238 for the time-aware Clinical Transformer and 0.6320 for Clinical GRU. In the multimodal setting, RNA+CovarianceKalman reaches 0.7160, outperforming both RNA+Transformer (0.6990, +0.0169 mean paired gain, winning all 5 repeats) and RNA+GRU (0.7022, +0.0137 mean paired gain, p = 0.006, winning all 5 repeats).Against RKN and unconstrained learned-gate estimators (Appendix E, Table 7 & Section 4.6): In a matched five-repeat evaluation, RNA+CovarianceKalman (0.7191) outperforms a time-aware Recurrent Kalman Network adaptation (0.6901, +0.0290 gain, winning all 5 repeats). Even when keeping the exact same state-prediction backbone and replacing only the covariance-derived gain with a learned constant gain (0.7079) or a learned time-dependent gain driven by inter-visit intervals (0.7035), performance drops by 0.0112 and 0.0156 C-index, respectively.Under test-time missing observations (Appendix D, Table 6): When randomly masking 10%, 30%, and 50% of observed clinical feature values at test time without retraining, RNA+CovarianceKalman retains a C-index of 0.7051 even at 50% missingness (compared with 0.6912 for RNA+GRU), demonstrating that explicit covariance propagation and observation masking provide superior robustness for sparse, irregular clinical survival trajectories.

AuthorsOpenReview ↗

Response to Reviewer Rc9U

Q4. Multimodal fusion mechanism and branch contributions

A4.We have added full details of the molecular ensemble and multimodal fusion design in Appendix C and Appendix I, with ablation studies quantifying their contributions. Two-MLP molecular ensemble: The design is motivated by overfitting risk for high-dimensional RNA-seq features in a moderate-sized cohort. We use two single-hidden-layer MLP Cox predictors with distinct capacities and regularization strengths (RNA Strong: width 32, dropout 0.28, weight decay 0.015; RNA Light: width 16, dropout 0.35, weight decay 0.025). Their blend weight λ is selected by grid search exclusively on inner-fold out-of-fold predictions, with no access to test data. The ensemble achieves a mean C-index of 0.6623, outperforming both individual components (0.6556 and 0.6348), as shown in Appendix C (Table 5). Multimodal fusion mechanism: Both branches are trained independently with Cox partial likelihood, and their outputs represent standardized log-hazard risk scores. Linear fusion therefore corresponds to an additive log-hazard model, which is statistically interpretable and has minimal parameters to avoid overfitting. The fusion weight α is selected by grid search over {0, 0.05, ..., 0.70} strictly on inner-fold OOF predictions within each training partition, and fixed before evaluation on the outer validation fold. This procedure ensures no test leakage. As shown in Appendix I (Table 10): Used alone, the molecular branch achieves 0.6623 and the clinical branch achieves 0.6616. Linear fusion raises the C-index to 0.7160, demonstrating strong complementary prognostic information between the two modalities. The performance gain is robust across a range of fixed fusion weights and even a nonlinear max-risk rule, confirming the gains come from complementary signals rather than weight tuning.

Q5. Missing technical definitions and implementation details

A5.We apologize for the missing technical definitions in the original manuscript. We have expanded Appendix B and Appendix J to provide complete, self-contained specifications of all architectural components and symbols. For clarity, we summarize the key definitions below: Static representation s_{i}backslash (Eq. 2): Concatenation of 32 training-fold PCA components (from the 3,000 most variable RNA-seq genes), age, an age-missingness indicator, and one-hot encoded sex variables. Observation mask m_{i,t}(Eq. 2): A binary feature-level vector indicating which clinical variables are observed at visit t for patient i. A separate sequence-validity mask marks padded sequence positions, which do not trigger state updates. Observation encoder E_{\phi }\left(\cdot \right) (Eq. 5): Maps the concatenated input vector u_{i,t} (clinical values, mask, normalized visit time, normalized inter-visit interval) into the 16-dimensional latent state space. The architecture is: linear projection (24 units) → LayerNorm → ReLU → dropout (rate 0.05) → final linear projection (16 units). Molecular ensemble: Two single-hidden-layer MLP Cox predictors with the specifications detailed above, combined via inner-OOF-selected weight λ.

Thank for your effort in the rebuttal. Considering the full discussion, including your responses to other reviewers, I will keep my score as is.

Reviewer_BPzuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a multimodal pipeline for establishing patient risk. The first branch is static, using RNA-sequences and "static covariates" to predict risk. For this branch, the authors train two MLP Cox models, one strongly and one lightly regularized, and they are combined in a linear combination. The second clinical branch uses embedded clinical information in a two-step Kalman filter process involving a time-aware uncertainty step (where longer gaps result in greater uncertainty), and an update step, which incorporates noise to update the latent state with uncertainty/covariance. The two outputs from the static and clinical branches are then united in another linear combination, which is used to predict patient risk.

Critical strengths
  1. The work is important in its application domain, and useful methodology and predictions in this field can be used to intervene in a risky patient's care in a timely manner.
  2. Results seem to support the benefits of multi-modality when comparing fused to static performance alone.
Critical weaknesses
  1. The contribution lies primarily in putting together this pipeline to solve a particular problem, which should not be understated, but nonetheless means the novelty is low. No fundamentally novel ideas are implemented or tested.
  2. A number of design choices and items are unclear or poorly defined. For example, two MLP Cox models are used for the static branch, yet no explanation is given for this requirement, nor does ablation testing reveal why two models are required. The static covariates are also not fully explained, etc.
  3. Reproducibility is not well-documented. Many hyperparameters and details of variables, including how they were chosen or initialized, are not available in the paper.
  4. Comparisons against alternatives are lacking. The approach is compared to a GRU alternative, but significance is not established. Furthermore, the paper alludes to transformers and state-space models, yet no comparison is given. KalmanNet is referenced as well but not compared against.

Minor Issues Figures are not well formatted. For example, the Figure 1 title looks like it uses two different fonts/bolds, and text in some locations is compressed horizontally.

Questions for authors
  1. Why use two MLP Cox models? Why not more, or an ensemble?
  2. Have you considered or attempted an order-free approach to the clinical branch? Either shuffled or perhaps with partial observations?
3 comments
AuthorsOpenReview ↗

Response to Reviewer BPzu

We thank the reviewer for constructive feedback and for highlighting the clinical value of multimodal risk prediction. We have expanded the Appendix with component ablations, partial-observation stress tests, additional sequence baselines, and complete implementation details, and we have revised Figure 1 to ensure clean visual formatting. Q1.Novelty and contribution positioning A1. We appreciate this candid feedback and recognize that our framework builds on established neural and Kalman filtering components. Our core contribution lies in bridging neural representation learning with explicit continuous-discrete Kalman covariance propagation for right-censored survival prediction over irregular clinical visits. Unlike black-box recurrent networks or learned-gate filters such as KalmanNet, our model scales process covariance explicitly with elapsed inter-visit time and computes an analytical Kalman gain from predicted and measurement uncertainties. It addresses specific, under-explored challenges of multimodal survival prediction with sparse, irregular clinical trajectories. Specifically, our work makes the following non-trivial design contributions: We integrate explicit elapsed-time-dependent process covariance into a neural Kalman encoder for right-censored survival prediction, providing a structured inductive bias for irregular visit timing that is not present in standard GRU or Transformer sequence models. We establish a rigorous, leakage-controlled landmark evaluation pipeline with matched multimodal late fusion, enabling fair head-to-head comparison of longitudinal clinical encoders under identical molecular and clinical inputs. We provide systematic ablations and baselines isolating the value of covariance-derived correction, time-aware process noise, and multimodal fusion, offering empirical insights for future clinical survival modeling work.

Q2. Motivation and ablation of the two-MLP Cox molecular ensemble A2. Thank you for your valuable comments. The two-MLP design is motivated by the core challenge of high-dimensional transcriptomic data in a moderate-sized clinical cohort: single neural predictors are prone to overfitting and unstable risk estimates when trained on limited patient samples with censored survival supervision. Instead of a single network, we pair two compact single-hidden-layer MLP Cox predictors with distinct capacities and regularization strengths: RNA Strong: wider hidden layer (32 units), lighter regularization (dropout 0.28, weight decay 0.015) RNA Light: narrower hidden layer (16 units), stronger regularization (dropout 0.35, weight decay 0.025) The two models capture complementary patterns: the stronger model fits finer-grained molecular patterns, while the lighter model provides a more conservative, generalizable baseline. Their combination is weighted by a coefficient λ selected exclusively on inner-fold out-of-fold predictions, with no test leakage. As shown in Appendix C (Table 5), across five repeated nested-CV runs, the ensemble achieves a mean C-index of 0.6623, outperforming both RNA Strong (0.6556) and RNA Light (0.6348) alone. We chose a two-model ensemble rather than a larger ensemble for parsimony: adding more models introduces minimal marginal gain while increasing computational complexity and hyperparameter tuning burden. This lightweight design is held identical across all multimodal baselines to ensure fair comparison. Static covariates consist of age, an age-missingness indicator, and one-hot sex encodings, as documented in Appendix B. We have added this motivation and ablation to the revised manuscript in Section 3.2 and Appendix C.

AuthorsOpenReview ↗

Response to Reviewer BPzu

Q3.Statistical significance and comparisons with Transformers and Kalman variants A3.We thank the reviewer for this suggestion. We have substantially expanded our baseline comparisons and statistical validation, with full results in Appendix A, Appendix E, and Appendix H. First, we have added formal paired statistical tests across the five repeated nested-CV runs. The improvement of RNA+CovarianceKalman over RNA+GRU is statistically significant for both overall C-index (+0.0137, paired two-sided t-test, p = 0.006) and 365-day AUC (+0.0230, p = 0.007). Second, we have added two categories of additional baselines, all trained and evaluated under identical input and fusion protocols:

  1. Time-aware Transformer-Cox baseline: Using the same clinical inputs, timing channels, and late-fusion pipeline, RNA+CovarianceKalman (0.7160) outperforms RNA+Transformer (0.6990) across all five repeats with a mean paired gain of +0.0169 (Appendix H, Table 9).
  2. Kalman-filtering related state-space baselines: We include a time-aware adaptation of Recurrent Kalman Networks (RKN) — a representative state-space model for sequential inference — and two learned-gain variants (learned constant gain, learned time-dependent gain). Our model consistently outperforms all three, with gains of +0.0290 over RKN and +0.0112/+0.0156 over the two learned-gain baselines (Appendix E, Table 7). Regarding KalmanNet: KalmanNet is designed for scenarios where partial analytical dynamics are known and replaces the Kalman gain with a learned recurrent network. In our clinical setting, no prior dynamical model of disease progression is available, so learned-gain baselines serve as a more direct and fair comparison for the value of explicit covariance-derived gain.

Q4.Order-free / non-recurrent clinical modeling and partial observations We thank the reviewer for this thought-provoking question. We address it from two perspectives: the necessity of sequential ordering, and robustness to partial observations. On order-free / non-recurrent modeling Clinical follow-up trajectories carry temporal information about disease progression: later visits reflect more advanced clinical state. For this reason, ordered sequential modeling is clinically meaningful. To quantify the value of recursive state estimation over order-insensitive aggregation, we compared our model against a clinical summary baseline, which condenses each patient’s trajectory into static summary statistics (mean, trend, recency) without recursive updates. As shown in Appendix A (Table 2), RNA+CovarianceKalman (0.7160) outperforms RNA+ClinicalSummary (0.6843) by +0.0317 (p = 0.020), confirming that ordered sequential state capture provides substantial prognostic value. Robustness to partial observations To evaluate performance under missing measurements, we conducted a partial-observation stress test by randomly masking 10%, 30%, and 50% of clinical feature entries at test time without retraining. As shown in Appendix D (Table 6): RNA+CovarianceKalman maintains higher discrimination than RNA+GRU across all missingness levels. Even with 50% of features masked, our model retains a C-index of 0.7051, which is comparable to the unperturbed RNA+GRU score of 0.7022. This robustness comes from explicit observation masking and covariance-weighted updates: the filter naturally down-weights noisy or missing observations during state correction, without additional training. We have added these results and discussion to Appendix D.

Q5. Reproducibility, hyperparameters, and Figure 1 formatting We apologize for the missing implementation details and formatting issues in the original submission. We have addressed both points thoroughly in the revision:

  1. Reproducibility details: We have documented complete variable definitions, fold-contained feature selection criteria, network architectures, optimization hyperparameters, and NCI GDC data provenance (including file UUIDs and MD5 checksums) in Appendix B, Appendix G, and Appendix J. All preprocessing and model selection rules are described in sufficient detail to be fully replicated.
  2. Figure formatting: We have regenerated Figure 1 and all other figures with unified typography, consistent font weights and sizes, standard aspect ratios, and properly spaced text elements to ensure readability and compliance with ICLR formatting guidelines.

I appreciate the authors’ effort in the rebuttal. Some points were clarified, but taking the full discussion into account, I believe the core gaps remain, so I will keep my score as is.

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

The paper predicts residual overall survival at a 365-day landmark using baseline RNA, demographic information, and first-year clinical histories from 556 CoMMpass patients, including 95 deaths. Using a neural observation encoder feeds a latent Kalman filter with elapsed-time-dependent process covariance; its clinical Cox risk is combined with a separately trained molecular predictor through late fusion. Five repeated nested cross-validation runs yield favorable results for static prediction, mean AUC, and Brier scores.

Critical strengths

The paper strengths in my opinion are as follows,

  • The explicit landmark restricts inputs to information available before prediction, while fold-contained preprocessing and inner-fold model selection reduce opportunities for leakage and selection bias.
  • The model incorporates visit timing and missingness masks, and its positive covariance parameterization and Joseph update provide a mathematically consistent filtering mechanism.
  • The GRU comparison uses the same visit representation and fusion protocol, while clinical-only results, multiple survival metrics, and structural ablations provide complementary evidence beyond a single performance number.
Critical weaknesses

However, I also noted some weaknesses which I'll outline below,

  • The paired tests across five repeated partitions of the same 556 patients characterize split variability rather than patient-sampling uncertainty. Consequently, the approximately 0.014 C-index improvement over GRU needs stronger uncertainty estimation or independent validation before supporting broad superiority claims.
  • Recurrent Kalman Networks already combine learned observation representations with latent Kalman updates, making their relationship to this contribution important to establish. The fixed-gain ablation also provides a weaker control than a learned constant or time-dependent gain for isolating the value of covariance propagation.
  • The manuscript omits the exact clinical/static feature lists, clinical-variable selection rules, cohort release, and several baseline architecture and tuning details. These omissions prevent confident reconstruction of the task and assessment of comparison fairness.
Questions for authors

I would appreciate if the authors could provide answsers to the following questions:

  • Can you provide patient-level confidence intervals for the paired performance improvement that account for model refitting and the evaluation design? Please distinguish this uncertainty from variability across random partitions and identify the primary statistical comparison.
  • Does the advantage persist against a matched Recurrent Kalman Network or learned time-dependent gain?
  • Would it be possible to provide the exact cohort release, clinical/static inputs and selection rules, complete baseline configurations, and evaluation code? Visit-frequency, missingness, and numbers-at-risk summaries would also help assess the setting in which the gains arise.
3 comments
AuthorsOpenReview ↗

Response to Reviewer eYHB

We thank the reviewer for recognizing our leakage-controlled landmark design, the mathematical consistency of our Joseph-form covariance updates, and our matched evaluation protocol. We have expanded the Appendix with additional statistical analyses, architectural baselines, and full cohort documentation to address each of your questions. Q1. Statistical uncertainty and primary comparison A1. Thank you to the reviewer for raising an important distinction between partition-induced variability and patient-sampling uncertainty. We have clarified this distinction in the revised paper and supplemented two additional tiers of patient-level analyses to more rigorously characterize the robustness of the observed performance gains. Our primary comparison remains the matched paired performance of RNA+CovarianceKalman versus RNA+GRU across five repeated nested-CV runs under identical inputs, fusion protocols, and random seeds. Across all five matched repeats, RNA+CovarianceKalman consistently outperforms RNA+GRU with a mean paired C-index improvement of +0.0137 (paired two-sided t-test, p = 0.006) and a one-year AUC gain of +0.0230 (p = 0.007), as detailed in Appendix A (Table 2). This analysis characterizes the stability of the performance ranking across data partitions. To separate partition sensitivity from patient-sampling variability and estimate variability arising from patient sampling, we added two independent analyses in Appendix F (Table 8):

  1. Patient-level bootstrap confidence intervals: Using patient-averaged out-of-fold risk scores across all five repeats, we performed 10,000 paired bootstrap resamples at the patient level. The mean C-index gain remains positive at +0.0113, with a 95% bootstrap interval of [-0.0029, 0.0261].
  2. Full pipeline refitting under cohort resampling: We ran 10 complete bootstrap replicates of the full nested-CV pipeline, including fold-local feature selection, preprocessing, and inner-fold hyperparameter tuning. RNA+CovarianceKalman outperforms RNA+GRU in 9 out of 10 replicates, with a mean paired C-index improvement of +0.0182 (median +0.0134). All results demonstrate consistent advantages of the proposed method within the MMRF CoMMpass 365-day landmark cohort.

Q2. Comparison with Recurrent Kalman Networks and learned-gain alternatives A2. Thank you for this insightful question. We have expanded Section 2 to formalize our differences from RKNs, and added matched empirical comparisons against a time-aware RKN survival adaptation and two stronger learned-gain baselines in Appendix E (Table 7). All baselines use the exact same clinical input features, observation masks, timing channels, late-fusion protocol, and inner-fold hyperparameter tuning to ensure comparison fairness. Core differences from Recurrent Kalman Networks (Becker et al., ICML 2019) RKNs are designed for dense, general-purpose sequence modeling with factorized covariance updates and locally linear transition functions for high-dimensional latent states. In contrast, our estimator is tailored explicitly for sparse, highly irregular clinical survival trajectories: We drive process covariance directly by the elapsed inter-visit time through a closed-form time-dependent scaling, rather than learning transition dynamics from data. We compute an analytical Kalman gain in the encoded clinical state space from propagated prior covariance and learned measurement covariance, rather than using factorized approximate updates. Our design prioritizes sample efficiency and inductive bias for the low visit-count, high-missingness clinical setting, rather than capacity for dense sequence modeling. Empirical comparison results Across five matched nested-CV repeats: RNA+CovarianceKalman achieves a mean pooled OOF C-index of 0.7191, outperforming the time-aware RKN adaptation (0.6901) across all five repeats with a mean paired gain of +0.0290. Compared to the stronger learned-gain controls: our model outperforms a learned constant gain (0.7079) by +0.0112 and a learned time-dependent gain driven by inter-visit intervals (0.7035) by +0.0156, winning 4 of 5 repeats against each. These results confirm that explicit covariance propagation and analytically derived gain provide a stronger inductive bias for irregular clinical survival trajectories than unconstrained learned gates or general-purpose RKN architectures.

AuthorsOpenReview ↗

Response to Reviewer eYHB

Q3. Dataset details, feature selection, and reproducibility A3. We apologize for the missing implementation details in the original submission. We have added full documentation of cohort provenance, feature selection rules, baseline configurations, and evaluation protocols in Appendix B, Appendix G, and Appendix J. All preprocessing, feature selection, and hyperparameter tuning are strictly confined to training folds, and all baselines use the identical input features, normalization statistics, and fusion pipeline to ensure fair comparison. Key details are summarized below: Data provenance: Data were retrieved from the MMRF CoMMpass study via the NCI GDC API (STAR-Counts workflow), starting from 859 RNA-seq profiles and 995 clinical records. GDC file UUIDs and MD5 checksums are preserved in our download manifest for full verification. Cohort characteristics: The 365-day landmark eligibility yields 556 patients (95 observed deaths, 461 censored) with 2,183 pre-landmark visits. Median visits per patient is 4 (IQR 4–4, range 1–6), with a median inter-visit interval of 89 days (IQR 78–98, range 14–218 days). Exact numbers at risk from day 0 to day 1095 are reported in Appendix B (Table 4). Feature selection: Clinical feature selection is executed strictly inside each training fold: candidate variables must appear in ≥3% of training visits with ≥2 distinct numeric values, ranked by coverage and value diversity after excluding outcome- and treatment-related fields. This procedure consistently selects 6 clinical input columns (ECOG status, height, weight across standardized and raw tables) across all 25 outer training folds. Static covariates include age, an age-missingness indicator, and one-hot sex encodings. Baseline configurations: Full architecture specifications, hyperparameter search grids, and optimization settings for all baselines are provided in Appendix B. To further support reproducibility, we will release our full evaluation code, data preprocessing pipeline, and trained model checkpoints upon acceptance of the paper.

Reviewer_eYHBOpenReview ↗

read your comments

Thank you for answering. Taking into consideration everything (including the rest of the discussion) - I feel that the gaps remain; therefore I will maintain my score as-is.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.