REVIEW 3 major objections 4 minor 11 references
TrajSurv: Learning Continuous Latent Trajectories from Electronic Health Records for Trustworthy Survival Prediction
T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read TrajSurv claims that continuous latent trajectories aligned to SOFA severity through time-aware contrastive learning make ICU survival prediction accurate and interpretable.
desk verdict Solid ML4H paper with a genuine new objective and useful transparency tooling; the load-bearing transparency claim is weakened by lookahead in the trend label, but the accuracy results stand on their own. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the learned vector field $f_\theta$ of the NCDE, shaped by the TACL loss. In derivative form, $dz/dt(t) = f_\theta(z(t))\, dX/dt(t)$, so each column of $f_\theta$ describes how a unit change in one clinical feature contributes to the magnitude and direction of latent motion; the paper uses column magnitudes as feature importance and column cosine similarities as feature relevance. TACL adapts Rank-N-Contrast to survival by contrasting anchor latent states against SOFA severity $s_t$ and trend $v_t = (s_{t+\Delta t}-s_{t-\Delta t})/(2\Delta t)$, with the time mask $\Phi$ penalizing contrasts between states far apart in time. The last latent state feeds a nonlinear Cox proportional-hazards head for the survival prediction itself, and DTW clustering of the full trajectories provides the trajectory-to-outcome link.
What would settle it
Train the same architecture with the trend defined from the past only, $v_t = (s_t - s_{t-\Delta t})/\Delta t$, keeping everything else fixed, and compare the Spearman correlation between latent-state distances and severity distances across the 36-hour window; if the causal variant keeps the alignment gains, the lookahead in $v_t$ is not doing the work, and if it does not, the future-dependent component of the label is responsible.
Extended reading notes
Core claim
TrajSurv's central claim is that the continuous latent trajectory $\tau = \{z_t, t \in [0,t_n]\}$, produced by a neural controlled differential equation $z_t = z_0 + \int_0^t f_\theta(z_s)\,dX_s$, can be made clinically meaningful by a time-aware contrastive learning (TACL) objective that pulls latent states together when their SOFA severity and severity-trend labels are close, while a time mask $\Phi(t_i,t_j;\kappa_2) = \exp(-|t_i-t_j|/\kappa_2)$ weakens contrasts between states far apart in time. Once the latent space is aligned, the paper claims the model becomes interpretable in two steps: the columns of the learned vector field $f_\theta$ rank feature importance and feature relevance (how similarly two features drive latent motion), and dynamic-time-warping clustering of trajectories recovers distinct clinical progression patterns whose Kaplan-Meier curves stratify survival. On MIMIC-III and eICU the reported accuracy is competitive (C-index 0.803 and 0.823) with transparency that existing deep survival models lack. The paper frames this as closing the gap left by continuous-time survival models that supervise only the final latent state, leaving intermediate trajectory evolution unaligned and uninterpretable.
Load-bearing premise
The load-bearing premise is that a severity-trend label computed from a short window centered on each time point can supervise a latent state that is supposed to summarize only what happened up to that time; if the future half of the window leaks information, the reported clinical alignment may not reflect a causal trajectory.
Editorial extensions
If this is right
- Clinicians can ask not only how risky a patient is, but which ongoing lab trends are currently moving the latent risk state and in which direction.
- The learned vector field is reusable across downstream tasks on longitudinal EHR because it encodes feature-change-to-state-change dynamics rather than just one outcome.
- Trajectory clustering yields a data-driven phenotyping in which patients in the same cluster share severity evolution and survival curves, supporting risk stratification for care planning.
- Because the time mask separates states with similar severity at different times, the model can distinguish 'severe but improving' from 'severe and deteriorating', matching clinical reasoning about trends.
Reading between the lines
- Editorial inference: the trend label $v_t$ is a central difference that uses severity at $t+\Delta t$, while $z_t$ is defined from history up to $t$; the reported clinical alignment could be inflated by this lookahead, and a causal variant using only past severity would be a sharper test.
- Editorial inference: replacing SOFA labels with other severity instruments such as APACHE or MELD, or with a data-driven state definition, would show whether TACL's alignment transfers beyond the ICU setting the paper evaluates.
- Editorial inference: the vector-field feature rankings are computed from an average field over patients and times, so population-level interpretations may smooth over heterogeneous individual dynamics; per-patient vector-field analysis is the natural next check.
- Editorial inference: comparing DTW clusters against diagnosis codes, or substituting dynamic state warping for DTW, would test whether trajectory shape itself, not just endpoint severity, carries the prognostic signal.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes TrajSurv, a survival prediction model for irregularly sampled longitudinal EHR data. TrajSurv uses a neural controlled differential equation (NCDE) to map clinical time series into continuous latent trajectories, and a time-aware contrastive learning objective (TACL) that aligns latent states with SOFA severity scores and severity trends. For transparency, the model performs a two-step interpretation: analyzing the learned vector field to derive feature importance and feature relevance, and clustering latent trajectories with dynamic time warping to identify clinical progression patterns associated with survival outcomes. The method is evaluated on MIMIC-III and eICU for in-hospital mortality prediction, reporting a C-index of 0.803 on MIMIC-III and 0.823 on eICU, with ablations, cross-cohort generalization, and case studies supporting the claims of competitive accuracy and interpretability.
Significance. If the claims hold, TrajSurv would be a useful contribution to continuous-time survival analysis with an interpretable latent trajectory, addressing a real gap: existing deep survival models often do not explain how feature changes over time drive outcomes. The empirical part is substantial: results on two large ICU datasets, ablations, cross-cohort testing, and public code. The transparency framework, especially the vector-field-based feature importance and relevance, is novel and clinically motivated. However, the load-bearing transparency claim depends on the clinical alignment of the latent trajectories, and the current TACL supervision introduces a temporal lookahead that needs to be addressed before the alignment and interpretation results can be taken at face value.
major comments (3)
- [Section 3.1.2, Eq. (v_t definition), and Section 3.1.1] The trend label v_t = (s_{t+Delta t} - s_{t-Delta t})/(2 Delta t) depends on the future severity s_{t+Delta t}, with Delta t = 2h in the experiments. This means the TACL objective in Eq. (8) supervises the latent state z_t with information from up to 2 hours after t, while Section 3.1.1 explicitly defines z_t as summarizing only historical information up to t. Training z_t against a future-dependent label can make the latent states appear aligned with near-future clinical progression even though, at inference time, the NCDE computes z_t from the past only. This lookahead directly affects the central transparency claim, since the clinical alignment shown in Figure 2C/D and the trajectory interpretation in Section 5.3 are all trained and evaluated on the same future-dependent labels. The paper does not discuss this temporal mismatch. Please either use a causal, backward-only trend definition or provide an analysis demonstrating that the lookahead does not materially affect the alignment and interpretation results.
- [Section 4.4 and Figure 2C/D] The clinical alignment evaluation is partly circular. TACL is trained to make latent distances rank-correlate with SOFA and SOFA-trend label distances (Eq. 8), and Figure 2C/D then reports the Spearman correlation between exactly these same two quantities with and without TACL. This confirms that the contrastive objective achieved its training goal, but it does not independently establish that the latent trajectories reflect clinical progression. An independent validation would strengthen the claim, for example by showing that latent trajectory similarity predicts outcomes or clinical annotations not used in the TACL labels, or by qualitatively comparing trajectory clusters with physician-adjudicated progression patterns.
- [Abstract and Section 5] The abstract and Section 5 claim 'superior transparency' over existing deep learning methods, but the transparency evaluation is primarily qualitative and is not compared against baselines on any quantitative metric. Table S1 lists conceptual differences with SHAP/permutation importance and attention mechanisms, but no head-to-head measurement (e.g., interpretability fidelity, feature importance agreement with clinical knowledge, or clinician-rated usefulness) is provided. The claim of superiority should be either quantified against baseline interpretation methods or tempered to a claim of complementary transparency.
minor comments (4)
- [Throughout] There are formatting artifacts in the text, such as 'SOF A' for SOFA, 'T rajSurv' for TrajSurv, 'J un' and 'Y etisgen' in author names, and 'M¨ uller' in the references. These should be cleaned before publication.
- [Section 4.1] Please clarify whether the SOFA scores used to compute trends at time t are available in real time at t. Specifically, with Delta t = 2h, s_{t+2h} is not known at t in a deployed system; state explicitly whether the labels are only used for training and whether this is consistent with the definition of z_t.
- [Section 5.3.2] The number of clusters C is stated to be four, but the selection criterion for C is not described. Please report how C was chosen (e.g., elbow method, clinical interpretability, silhouette score).
- [Section 5.3.2] The Kaplan-Meier curves are described as showing 'statistically significant separation,' but no formal test statistic or p-value is reported. Add a log-rank test or equivalent across the clusters.
Circularity Check
TrajSurv's transparency sub-claim is partially circular: TACL trains latent distances on SOFA/trend label distances and then evaluates alignment with those same distances, while the trend label is computed from future SOFA values (t+2h), so the 'clinically aligned progression' evidence is partly manufactured by the training objective and by temporal leakage.
-
self definitional
[Section 5.2, Eq. (8)]
"We further evaluated the clinical alignment between latent states and patient states by computing Spearman’s correlation between latent distance and SOF A or SOF A trend distance at each time point."
The TACL objective in Eq. (8) is a Rank-N-Contrast loss whose pairwise structure is defined by the same label distance |s_i - s_j| + δ|v_i - v_j| that Section 5.2 measures after training: TACL explicitly reshapes latent distances to follow the ordering of SOFA/trend distances. Computing Spearman correlation between latent distance and SOFA/trend distance therefore checks the training objective rather than an independent clinical ground truth. Figure 2C/D's 'clinical alignment' result is the same quantity the loss was designed to maximize, so this part of the transparency claim reduces, by construction, to the definition of the loss.
-
fitted input called prediction
[Sections 3.1.1, 3.1.2, and 4.1]
"we define the ongoing trend of severity as vt = (st+∆t − st−∆t)/2∆t"
Section 3.1.1 defines z_t as summarizing the patient's historical information up to t without knowing the clinical features after t, while Section 3.1.2 defines v_t using s_{t+Δt}. With Δt=2h in the experiments, every anchor latent state z_t is supervised in LTACL by a trend label that includes severity two hours in the future. The latent trajectory can therefore become 'clinically aligned' by peeking ahead, and the cluster patterns and KM separation in Section 5.3 are partly manufactured projections of the future-containing SOFA label rather than evidence that past feature changes causally drive the trajectory. The Limitations section acknowledges reliance on SOFA as supervision, but it does not address the temporal direction of that supervision signal.
full rationale
Survival accuracy (C-index, Brier, dynamic AUC) is evaluated on held-out MIMIC-III and eICU test splits against external baselines, so the accuracy sub-claim has independent empirical grounding and is not circular. The circularity is confined to the paper's central transparency sub-claim. First, the clinical-alignment evaluation in Section 5.2 measures Spearman correlation between latent distance and SOFA/trend distance, which is precisely the pairwise label-distance structure optimized by the TACL loss in Eq. (8); reporting this as evidence that the latent trajectory 'reflects clinical progression' is largely a check of the training objective, not an external clinical benchmark. Second, the trend definition v_t = (s_{t+Δt} − s_{t−Δt})/(2Δt) uses future severity when Δt=2h, contradicting the history-only definition of z_t in Section 3.1.1; TACL therefore can manufacture the appearance of clinically aligned trajectories by allowing z_t to be shaped by information after t. The two-step interpretation (vector-field importance and DTW clustering) inherits this leakage, since the clusters are subsequently shown to have distinct average SOFA trajectories that were part of the supervision signal. Because the survival-prediction claim is externally validated while the transparency claim substantially reduces to the TACL objective and a future-looking label, the paper is partially circular rather than wholly so. Score 6 reflects that the 'prediction' of clinical alignment is partly forced by construction, while independent survival performance keeps the paper from a higher score.
Assumptions & free parameters
free parameters (7)
- alpha (TACL weight) =
tuned on validation set, final value in Table S2
- kappa1 (contrast temperature) =
tuned on validation set, final value in Table S2
- kappa2 (time mask scale) =
tuned on validation set, final value in Table S2
- delta (severity/trend balance) =
tuned on validation set, final value in Table S2
- dz (latent dimension) =
tuned on validation set, final value in Table S2
- C (number of trajectory clusters) =
4
- Delta t (trend window half-width) =
2 hours
assumptions (6)
- standard math NCDE integral representation and Riemann-Stieltjes integration are valid for the interpolated EHR control signal.
- domain assumption SOFA score is a valid proxy for patient clinical state.
- domain assumption Cubic Hermite spline interpolation of the input with backward differences is an unbiased representation of the true clinical process.
- domain assumption The last latent state is a sufficient predictor of survival, so that the Cox model on z_tn captures all relevant risk information.
- ad hoc to paper The time mask Phi(ti, tj; kappa2) is an appropriate way to encode time dependency in the contrastive objective.
- ad hoc to paper Feature importance defined by column magnitude of the average vector field is a valid proxy for clinical importance.
invented entities (1)
-
Clinically aligned continuous latent trajectory
Cite this review
Pith. "Pith review of TrajSurv: Learning Continuous Latent Trajectories from Electronic Health Records for Trustworthy Survival Prediction." pith.science (2026). https://pith.science/paper/23OYBQKX
@misc{pith2026250800657,
author = {Pith},
title = {Pith review of: TrajSurv: Learning Continuous Latent Trajectories from Electronic Health Records for Trustworthy Survival Prediction},
year = {2026},
howpublished = {\url{https://pith.science/paper/23OYBQKX}},
note = {Machine review of arXiv:2508.00657}
}
read the original abstract
Trustworthy survival prediction is essential for clinical decision making. Longitudinal electronic health records (EHRs) provide a uniquely powerful opportunity for the prediction. However, it is challenging to accurately model the continuous clinical progression of patients underlying the irregularly sampled clinical features and to transparently link the progression to survival outcomes. To address these challenges, we develop TrajSurv, a model that learns continuous latent trajectories from longitudinal EHR data for trustworthy survival prediction. TrajSurv employs a neural controlled differential equation (NCDE) to extract continuous-time latent states from the irregularly sampled data, forming continuous latent trajectories. To ensure the latent trajectories reflect the clinical progression, TrajSurv aligns the latent state space with patient state space through a time-aware contrastive learning approach. To transparently link clinical progression to the survival outcome, TrajSurv uses latent trajectories in a two-step divide-and-conquer interpretation process. First, it explains how the changes in clinical features translate into the latent trajectory's evolution using a learned vector field. Second, it clusters these latent trajectories to identify key clinical progression patterns associated with different survival outcomes. Evaluations on two real-world medical datasets, MIMIC-III and eICU, show TrajSurv's competitive accuracy and superior transparency over existing deep learning methods.
Figures
Figures from the paper (2 more)
Reference graph
Works this paper leans on
-
[5]
to prevent overfitting. B.4. Optimization We used AdamW Loshchilov and Hutter (2019) to optimize TrajSurv. For the NCDE module, we used the torchcde package with torchdiffeq backend. We trained TrajSurv for 100 epochs, with early stopping based on the C-index on the validation set, with patience
work page 2019
-
[6]
Chirag Nagpal, Willa Potosnak, and Artur Dubrawski. auton-survival: an open-source pack- age for regression, counterfactual estimation, evaluation and phenotyping with censored time-to-event data. arXiv preprint arXiv:2204.07276 ,
-
[7]
Continuous-time modeling of counterfactual outcomes using neural controlled differential equations
Nabeel Seedat, Fergus Imrie, Alexis Bellot, Zhaozhi Qian, and Mihaela van der Schaar. Continuous-time modeling of counterfactual outcomes using neural controlled differential equations. arXiv preprint arXiv:2206.08311 ,
-
[11]
TrajSurv’s training and evaluation were performed on a single NVIDIA Tesla T4 or RTX 2080 Ti GPU. B.5. Comparison Methods Implementation All baseline models underwent hyperparameter tuning using grid search on the validation set. For CoxPH, RSF, and Boosting, we used the implementations in the scikit-survival library P¨ olsterl (2020). For RDSM, we used t...
work page 2020
-
[2009]
Lu Wang, Yan Li, and Mark Chignell
URL https://api.semanticscholar.org/CorpusID:902086. Lu Wang, Yan Li, and Mark Chignell. Combining ranking and point-wise losses for training deep survival analysis models. In 2021 IEEE international conference on data mining (ICDM), pages 689–698. IEEE,
work page 2021
-
[2012]
Dynamical Survival Analysis with Controlled Latent States
Linus Bleistein, Van-Tuan Nguyen, Adeline Fermanian, and Agathe Guilloux. Dynamical survival analysis with controlled latent states. arXiv preprint arXiv:2401.17077 ,
-
[2017]
URL https://arxiv. org/abs/1703.01141. Adrian O Hosten. Bun and creatinine. Clinical Methods: The History, Physical, and Laboratory Examinations. 3rd edition,
-
[2019]
Scott M Lundberg and Su-In Lee
URL https://arxiv.org/abs/1711.05101. Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems , 30,
Show all 11 references
-
[2021]
Patrick Kidger, James Morrill, James Foster, and Terry Lyons
URL https://arxiv.org/abs/2004.11362. Patrick Kidger, James Morrill, James Foster, and Terry Lyons. Neural controlled differential equations for irregular time series. Advances in Neural Information Processing Systems , 33:6696–6707,
2004 arXiv
-
[2022]
Neural controlled differential equations for online prediction tasks
James Morrill, Patrick Kidger, Lingyi Yang, and Terry Lyons. Neural controlled differential equations for online prediction tasks. arXiv preprint arXiv:2106.11028 ,
-
[2024]
Additional Results A.1
19 TrajSurv Appendix A. Additional Results A.1. Model Calibration To assess TrajSurv’s calibration, we generated calibration plots across quartiles of follow-up times on MIMIC-III data. Figure S1 illustrates that TrajSurv’s predicted risk scores align well with observed outcom...
2017
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.