acceptodds
Under review as a conference paper at ICLR 2027

Where Is the Patient Heading? Multi-Path Forecast-Aware Mortality Prediction in the ICU

Abstract

Mortality prediction in the intensive care unit (ICU) requires estimating risk from the patient history available at a given decision time. Although existing predictors can model longitudinal observations and learn associations with future outcomes, they need not explicitly represent anticipated physiological changes or incorporate the outcomes of similar patients. We investigate whether explicit physiological forecasts and retrieved patient outcomes improve prediction under matched observation windows. We propose TRACE (**T**rajectory-**R**etrieval **A**ugmented **C**linical **E**stimation), which forecasts quantiles of near-future vital signs and uses the predicted median and interval width as features for mortality classification. A complementary retrieval pathway estimates risk from the observed outcomes of similar training patients, and the two estimates are combined through entropy-conditioned weighting. Across seven forecasting backbones on MIMIC-IV and eICU-CRD, TRACE improves discrimination relative to matched observed-history and point-forecast controls. Component comparisons associate both quantile-derived features and outcome retrieval with further gains, and the observed benefit over matched controls extends across the evaluated decision times. On a matched MIMIC-IV task, TRACE also achieves the highest AUROC and AUPRC among the mortality predictors evaluated. These findings support explicit physiological forecasting and outcome retrieval as complementary components for ICU mortality prediction in the studied settings.

Primary area:
learning on time series and dynamical systems
Keywords:
ICU mortality prediction, clinical time-series modeling, physiological forecasting, quantile forecasting, retrieval-augmented prediction

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_6G5YOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper suggest a method that combines forecast derived vital features with mortality outcomes retrieved from similar training patients. The paper is clear and we'll written.

Critical strengths
  1. Good results
  2. Squeezing more signal from the data
Critical weaknesses

Poor evaluations, you didn't try the real SOTA or at least AutoML methods The hyperparameters of the comparison methods are missing and the delta between LR and xgoboost are strange

Questions for authors

Did you use early stopping on XGBoost? Did you search for the hyperparameters in the comparison methods?

1 comment

Thank you for requesting a broader comparison and clearer information about baseline training. We have expanded the evaluation to include TabICLv2, DuETT, GARLIC, TPC, and ALFIA, alongside Logistic Regression, XGBoost, GRU-D, and Medformer. The models are fitted or trained and evaluated separately on MIMIC-IV and eICU-CRD, using the matched 0–36-hour observation window, patient-level splits, and in-hospital mortality endpoint. The reported experiments use three seeds, with standard deviations shown where available.

1. Expanded comparison with relevant predictors.

ModelMIMIC-IV AUROCMIMIC-IV AUPRCeICU-CRD AUROCeICU-CRD AUPRC
Logistic Regression0.81440.45110.77520.2312
XGBoost0.8343 ± 0.01260.4781 ± 0.01180.7788 ± 0.00290.2348 ± 0.0042
GRU-D0.8229 ± 0.00440.4684 ± 0.00770.7807 ± 0.00370.2367 ± 0.0071
Medformer0.6963 ± 0.00680.2985 ± 0.00250.6540 ± 0.00290.1101 ± 0.0016
TabICLv20.8317 ± 0.00170.4747 ± 0.00330.8019 ± 0.00090.2579 ± 0.0006
DuETT0.8364 ± 0.00140.4729 ± 0.00300.7971 ± 0.00290.2531 ± 0.0075
GARLIC0.8153 ± 0.00580.4406 ± 0.00830.7897 ± 0.00460.2457 ± 0.0123
TPC adapted0.8352 ± 0.00760.4797 ± 0.01240.7952 ± 0.00410.2512 ± 0.0018
ALFIA0.8058 ± 0.00500.4237 ± 0.01310.7636 ± 0.01770.2208 ± 0.0214
TRACE–iTransformer0.8443 ± 0.00310.5127 ± 0.01020.7821 ± 0.00350.2371 ± 0.0093

On MIMIC-IV, TRACE has the highest mean AUROC and AUPRC among the evaluated models. Its AUPRC exceeds XGBoost by 0.0346 and the strongest baseline mean AUPRC, obtained by TPC, by 0.0330. On eICU-CRD, TRACE is close to XGBoost and GRU-D in mean discrimination, while TabICLv2, DuETT, TPC, and GARLIC achieve higher values. These results will be included in the revision to give a fuller account of the framework's relative performance across cohorts.

The added evaluation also helps distinguish overall performance from the contribution of the proposed pipeline. With the same retrieval module added to a directly trained iTransformer classifier, TRACE still improves AUPRC from 0.4909 to 0.5127 on MIMIC-IV and from 0.2224 to 0.2371 on eICU-CRD. We will present these controlled comparisons alongside the expanded baseline table.

2. Logistic Regression versus XGBoost.

The repeated evaluation gives the following results.

DatasetModelAUROCAUPRCF1
MIMIC-IVLogistic Regression0.81440.45110.4652
MIMIC-IVXGBoost0.8343 ± 0.01260.4781 ± 0.01180.4840 ± 0.0142
eICU-CRDLogistic Regression0.77520.23120.2928
eICU-CRDXGBoost0.7788 ± 0.00290.2348 ± 0.00420.2962 ± 0.0069

XGBoost's mean advantages over Logistic Regression are 0.0199 AUROC and 0.0270 AUPRC on MIMIC-IV, compared with 0.0036 and 0.0036 on eICU-CRD. Thus, the size of the gap varies across the two cohorts. A plausible modeling explanation is that Logistic Regression imposes a linear log-odds relationship over its input features, while XGBoost can capture nonlinear relationships and feature interactions. This interpretation concerns the model families. The training and model-selection procedure for the expanded evaluation is detailed below.

3. Hyperparameter search and XGBoost early stopping.

For the expanded XGBoost evaluation, we performed a 60-trial random hyperparameter search and selected the configuration by validation AUPRC. Within each trial, early stopping monitored validation log-loss with a patience of 100 rounds and a maximum of 5,000 boosting rounds. Thus, validation log-loss determined when to stop boosting, while validation AUPRC determined which configuration to select.

Logistic Regression used standardized features with the L-BFGS solver, C = 1.0, balanced class weights, and a maximum of 2,000 iterations, following the fixed configuration used in Table 2 without additional hyperparameter search.

The expanded comparison and repeated evaluation provide a broader empirical basis for assessing TRACE. We will use the cohort-specific results above to state its strengths and limitations precisely, and provide the training details needed to make the comparison reproducible.

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

This paper proposed the TRACE framework consisting of three stages: (1) forecasting the physiology variables, (2) comparing with the nearest observed patient according to the similarity in the embedding, (3) using interpolation from neighbor patients to calibrate the final mortality outcome. The experiments are evaluated on the benchmark of MIMIC-IV and eICU-CRD showing improved performance compared with existing backbone models.

Critical strengths

The overall framework can be helpful in a clinical setting: considering the near future vital sign and the similar observed patients as a reference.

The experiment and evaluation is thourough: using 2 ICU datasets and applied to multiple existing backbones to validate the improvement.

Critical weaknesses

The motivation is unclear: why estimate near future physiology or to form a query-specific risk estimate is essential in the mortality prediction. The necessity of aggregating the near future forecasting and the patient retrievals is not fully explained. What kind of failure mode from existing time series model can be resolved, what prediction errors can benefit from explicit forecasting, or patient retrieval, or both mechanisms?

There lacks a comparison between the auxiliary forecasting supervision and explicit forecast features on the physiology. The auxiliary task of physiological forecasting could obtain the similar benefits during the representation learning, the advantage of explicit forecasting should be addressed as an item in ablation study.

For the retrieval mechanism, it always select the top-k from the observed training patients. This may raise the concern that if there lacks the sufficiently relevant observed patients in the rare subgroup, and the risk can be shifted.

Questions for authors

Why the combination of explicit physiological forecasts and retrieval are beneficial, any explanations and patient/subgroup patterns to further support this claim?

What is the advantage of explicitly forecasting the physiological variables over the auxiliary training into the model representations? Using the comparison on the same full historical inputs, model capacity and identical retrieval design?

How does TRACE deal with patients who lack sufficient similar examples from the observed reference cohort? For the retrieval mechanism, if we reduce the reference bank or remove the relevant patients, will it harm the model performance?

2 comments

Thank you for highlighting the motivation, representation choice, and dependence on relevant reference patients. We have added a comparison against Direct Classification with the same retrieval module, an outcome-stratified analysis of probability errors, a matched forecast-representation comparison, and reference-bank stress tests on both datasets.

1. Why use both explicit forecasts and retrieved outcomes, and which errors do they help correct?

The two pathways provide different representations for the same decision-time task. The forecast pathway maps the available history into predicted medians and interval widths of near-future heart rate, respiratory rate, and oxygen saturation. The retrieval pathway estimates risk from the observed outcomes of training patients with similar observed histories. At inference, both pathways use observations available within the first 36 hours. Subsequent physiological measurements are used as forecasting targets during training, while prediction-time features are generated forecasts.

To test their respective contributions, we added the missing Direct + retrieval control. Direct adapts iTransformer with a classification head for mortality prediction from observed history. For the retrieval-enabled comparison, Direct and TRACE use the same retrieval configuration and reference evidence. Thus, the comparison holds the retrieval estimate and its weighting rule fixed while changing the model pathway. Results are means ± standard deviations across three training seeds on the same patient-level split.

DatasetConfigurationAUROCAUPRCF1
MIMIC-IVDirect0.7422 ± 0.00510.3005 ± 0.00260.3573 ± 0.0052
MIMIC-IVDirect + retrieval0.8338 ± 0.00590.4909 ± 0.00790.5026 ± 0.0191
MIMIC-IVQuantile-only0.7906 ± 0.00190.4011 ± 0.00380.4341 ± 0.0030
MIMIC-IVTRACE0.8443 ± 0.00310.5127 ± 0.01020.5193 ± 0.0105
eICU-CRDDirect0.7227 ± 0.00330.1577 ± 0.00080.1707 ± 0.0108
eICU-CRDDirect + retrieval0.7711 ± 0.00410.2224 ± 0.00730.2693 ± 0.0075
eICU-CRDQuantile-only0.7512 ± 0.00100.1958 ± 0.00410.2288 ± 0.0062
eICU-CRDTRACE0.7821 ± 0.00350.2371 ± 0.00930.2917 ± 0.0102

Retrieval improves both model pathways. With retrieval held fixed, the forecast-based pipeline further improves AUPRC over Direct + retrieval by 0.0218 on MIMIC-IV and 0.0147 on eICU-CRD, with AUROC improvements of 0.0105 and 0.0110. Without retrieval, Quantile-only also improves over Direct on both datasets. Thus, the added control provides evidence for the forecast-based pipeline beyond the contribution of the retrieval module.

We additionally examined probability errors separately for patients who died and survived. In this analysis, A1 denotes Direct and A3 denotes Quantile-only. Delta Brier is the score after the change minus the score before it, so negative values indicate improvement.

DatasetConfigurationAUROCAUPRCF1
MIMIC-IVDirect0.7422 ± 0.00510.3005 ± 0.00260.3573 ± 0.0052
MIMIC-IVDirect + retrieval0.8338 ± 0.00590.4909 ± 0.00790.5026 ± 0.0191
MIMIC-IVQuantile-only0.7906 ± 0.00190.4011 ± 0.00380.4341 ± 0.0030
MIMIC-IVTRACE0.8443 ± 0.00310.5127 ± 0.01020.5193 ± 0.0105
eICU-CRDDirect0.7227 ± 0.00330.1577 ± 0.00080.1707 ± 0.0108
eICU-CRDDirect + retrieval0.7711 ± 0.00410.2224 ± 0.00730.2693 ± 0.0075
eICU-CRDQuantile-only0.7512 ± 0.00100.1958 ± 0.00410.2288 ± 0.0062
eICU-CRDTRACE0.7821 ± 0.00350.2371 ± 0.00930.2917 ± 0.0102

For a patient who died, moving toward the observed outcome means increasing the predicted mortality probability. For a survivor, it means decreasing that probability. This percentage measures the direction of individual changes, while Brier score also accounts for their magnitudes.

On MIMIC-IV, adding forecast features substantially reduces probability error in the deceased group, and adding retrieval yields its larger improvement in survivors. On eICU-CRD, the reductions are concentrated in survivors, with small increases in deceased-group Brier scores. These results identify cohort-dependent patterns. In particular, retrieval reduces excessive mortality probability assignments among many survivors, while the MIMIC-IV forecast comparison also shows improved probability assignment to observed deaths. We will reflect this subgroup trade-off in the discussion. These are analyses of probability errors, and their interpretation concerns the reported prediction task.

2. What do explicit forecast outputs add beyond a forecasting-pretrained representation?

We added a matched comparison that shares the full observed-history inputs, the forecasting checkpoint, and the downstream classifier design. The additional representation has 72 dimensions in both cases. The latent control uses a PCA-compressed encoder representation, and the explicit representation uses the predicted medians and interval widths. The retrieval-enabled conditions share the retrieval design.

DatasetAdditional representationRetrievalAUROCAUPRCF1
MIMIC-IVEncoder latent, PCA 72Off0.81650.46180.4494
MIMIC-IVForecast median + width, 72Off0.82390.47570.4679
MIMIC-IVEncoder latent, PCA 72On0.83980.50490.5066
MIMIC-IVForecast median + width, 72On0.84770.51950.5260
eICU-CRDEncoder latent, PCA 72Off0.74230.21600.1933
eICU-CRDForecast median + width, 72Off0.75010.21950.2114
eICU-CRDEncoder latent, PCA 72On0.76860.22760.2544
eICU-CRDForecast median + width, 72On0.78240.23370.2865

Explicit forecasts improve AUPRC by 0.0139 without retrieval and 0.0146 with retrieval on MIMIC-IV. The corresponding improvements on eICU-CRD are 0.0035 and 0.0061. The evidence supports explicitly exposing anticipated values and forecast spread to the classifier within this matched representation setting. Its scope is the evaluated 72-dimensional forecasting-pretrained latent control.

This experiment uses a shared full-history branch to control the observed information available to the classifier. Its forecast-plus-retrieval condition is reported separately from the main TRACE pipeline above. The two tables therefore address pipeline-level prediction and the choice of additional representation, respectively.

3. What happens when the reference bank contains fewer relevant patients?

We evaluated sensitivity to reference availability by reducing the reference bank and removing the nearest reference patients, with removal of the same number of randomly selected references as a control. The table reports overall-cohort AUPRC and the change in Brier score relative to retrieval Off within the low-similarity subgroup.

Reference conditionMIMIC-IV AUPRCeICU-CRD AUPRCMIMIC-IV low-similarity ΔBrier vs retrieval OffeICU-CRD low-similarity ΔBrier vs retrieval Off
Retrieval Off0.39410.2167——
Full bank0.51670.2331−0.0189−0.0223
Bank 50%0.50540.2296−0.0143−0.0214
Bank 25%0.50300.2227−0.0120−0.0159
Bank 10%0.46400.2136−0.0078−0.0081
Remove nearest 100.46680.2184−0.0081−0.0134
Remove random 100.49910.2269−0.0116−0.0193
Remove nearest 300.44750.2045−0.0060−0.0063
Remove random 300.48650.2258−0.0102−0.0184

Performance deteriorates as reference support decreases. Removing nearby references is more harmful than removing the same number at random. For example, removing the nearest 30 reduces AUPRC from 0.5167 to 0.4475 on MIMIC-IV and from 0.2331 to 0.2045 on eICU-CRD. Random removal of 30 yields 0.4865 and 0.2258, respectively. Average top-30 cosine similarity also falls from 0.9659 to 0.9011 on MIMIC-IV and from 0.9796 to 0.9479 on eICU-CRD when the nearest 30 are removed.

On MIMIC-IV, all displayed stressed conditions retain higher overall AUPRC than retrieval Off. On eICU-CRD, the 10% bank and nearest-30 removal conditions fall below retrieval Off, demonstrating a limitation under severe reference scarcity. The reported low-similarity subgroup has lower Brier error than retrieval Off across these conditions. That subgroup probability-error result and overall-cohort AUPRC describe different aspects of performance.

TRACE limits the retrieval contribution through the maximum interpolation weight and reduces it when the retrieved outcome distribution has high binary entropy. This rule responds to agreement in retrieved outcomes. Reference relevance requires separate assessment, as demonstrated by the bank stress tests. We will add this limitation concerning reference relevance to the discussion.

Overall interpretation. The additional experiments support two contributions under the evaluated conditions. A forecast-based representation improves prediction with the retrieval configuration held fixed, and retrieved outcomes improve the forecast-aware model. The representation comparison further supports the evaluated choice of explicit medians and widths. The outcome-stratified and reference-bank analyses identify where those improvements arise and where they weaken, giving a more precise account of when this combination is useful.

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

This paper proposes TRACE, a framework for ICU in-hospital mortality prediction. It combines a time series forecasting module for physiological signals and a similar patients retrieval module. Given a clinical history, TRACE forecasts quantiles of future heart rate, respiratory rate, and oxygen saturation, then uses the predicted medians and interval widths alongside historical and static features for mortality classification. Another retrieval module estimates risk from the outcomes of similar training patients, and the two estimates are combined for final prediction. The experiments are conducted to show the improvements of proposed modules and the overall performance on the matched tasks.

Critical strengths

S1. The problem studied in this paper is important for some real-world applications.

S2. A well-motivated integration of physiological forecasting and patient-outcome retrieval: The proposed method combines explicit predictions of near-future physiology with outcome evidence from similar patients.

Critical weaknesses

W1. Limited methodological novelty: The proposed method primarily combines established techniques; the overall contribution is incremental. At the same time, the experiments do not sufficiently establish why such combination is necessary. In the ablation, retrieval produces the largest incremental gain, yet a history-based classifier combined with the same retrieval module, without forecasting, is missing.

W2. The evaluation result presentation is poor: a) All experiment results seem to be one single run without reported variability across random seeds. b) On MIMIC-IV, TRACE exceeds XGBoost by only 0.0057 AUROC and 0.0052 AUPRC, and the result on eICU-CRD is absent. Also, the selected baselines are outdated and insufficiently comprehensive, omitting more recent relevant methods. Together, TRACE ‘gains remain unclear. c) In Table 1, the Direct Classification baseline yields different scores across rows, even though it does not depend on the forecasting backbone

Questions for authors

Q1. What differs between the separately fitted Direct Classification controls in Table 1? Are their inputs, architectures, and hyperparameters identical, with only random seeds varying?

Q2. Does forecasting improve performance when retrieval is held fixed? A comparison against a history-based classifier combined with the same retrieval module would address this question.

Q3. Can you provide more results to address Weakness 2a and 2b?

2 comments

Thank you for recognizing the importance of the problem and the motivation for combining physiological forecasting with patient-outcome evidence. Your review identified an important missing comparison and several points requiring clearer evaluation. We have added Direct Classification with the same retrieval module, repeated experiments across three seeds, and expanded the baseline comparison to both datasets.

W1 and Q2. Does the forecast-based pipeline help when the same retrieval module is already available?

We agree that the original stepwise ablation left this comparison unresolved. We now directly compare a history-based iTransformer classifier with and without retrieval against the forecast-based pipeline with and without the same retrieval module.

Direct adapts iTransformer with a classification head and trains it to predict mortality from the observed history. The forecast-based pathway first predicts physiological medians and interval widths and supplies those features to the mortality classifier. The retrieval-enabled conditions use the same reference evidence and interpolation configuration. Thus, the comparison holds the retrieval estimate and its weighting rule fixed while changing the model pathway. All configurations are evaluated at the same 36-hour decision time on the same test cohort.

DatasetConfigurationAUROCAUPRCF1
MIMIC-IVDirect0.7422 ± 0.00510.3005 ± 0.00260.3573 ± 0.0052
MIMIC-IVDirect + retrieval0.8338 ± 0.00590.4909 ± 0.00790.5026 ± 0.0191
MIMIC-IVForecast-based, retrieval Off0.7906 ± 0.00190.4011 ± 0.00380.4341 ± 0.0030
MIMIC-IVTRACE0.8443 ± 0.00310.5127 ± 0.01020.5193 ± 0.0105
eICU-CRDDirect0.7227 ± 0.00330.1577 ± 0.00080.1707 ± 0.0108
eICU-CRDDirect + retrieval0.7711 ± 0.00410.2224 ± 0.00730.2693 ± 0.0075
eICU-CRDForecast-based, retrieval Off0.7512 ± 0.00100.1958 ± 0.00410.2288 ± 0.0062
eICU-CRDTRACE0.7821 ± 0.00350.2371 ± 0.00930.2917 ± 0.0102

Results are means ± standard deviations across three training seeds on the same patient-level split. Retrieval produces a substantial gain for Direct, as the review anticipated. With retrieval held fixed, TRACE provides further mean improvements of 0.0105 AUROC, 0.0218 AUPRC, and 0.0167 F1 on MIMIC-IV. The corresponding improvements on eICU-CRD are 0.0110, 0.0147, and 0.0224. Thus, the added comparison supports the forecast-based pipeline even when the history-based classifier receives the same retrieval module.

W1. Methodological contribution and the role of explicit forecasts.

TRACE combines established forecasting and retrieval components around a specific decision-time representation. The mortality classifier receives fixed-horizon medians and interval widths of continuous ICU physiology, while the retrieval pathway forms a query-specific risk estimate from training-patient outcomes. The explicit forecast is a transformation learned from physiological forecasting supervision and the observed history. The retrieved outcomes provide a second estimate derived from reference patients.

The new Direct + retrieval comparison tests whether this forecast-based pipeline adds value after the retrieval contribution is controlled. We additionally examined the choice of forecast representation by sharing the full-history inputs, forecasting checkpoint, and downstream classifier design, and comparing 72-dimensional PCA encoder latents with 72-dimensional explicit forecast medians and widths.

DatasetRetrievalLatent AUPRCExplicit forecast AUPRCDifference
MIMIC-IVOff0.46180.4757+0.0139
MIMIC-IVOn0.50490.5195+0.0146
eICU-CRDOff0.21600.2195+0.0035
eICU-CRDOn0.22760.2337+0.0061

These results provide evidence for exposing the forecast outputs explicitly within the evaluated representation setting. This full-history representation experiment is distinct from the main TRACE pipeline table. Together, the experiments substantiate the pipeline comparison and the selection of explicit forecast features. We will make this scope of the methodological contribution clearer and connect each design choice to its corresponding control in the revision.

W2 (c) and Q1. Why do Direct Classification scores differ across backbone rows?

Each Direct Classification control uses the corresponding backbone adapted with a mortality classification head. Accordingly, the rows differ in their encoders and learned history representations. The classification head predicts mortality directly from the observed history. The backbone is used as a classifier in this condition.

Our description as an observed-history MLP control did not adequately specify the backbone-specific encoders. We will correct the experimental description and table caption to state this implementation explicitly and identify the classification adaptation for each backbone. The new Direct + retrieval experiment above uses the iTransformer-based Direct control.

W2 (a) and Q3. Variability across training seeds.

The new main-pipeline ablations and expanded baseline experiments are evaluated across three seeds. We report means and standard deviations explicitly. For example, TRACE achieves 0.8443 ± 0.0031 AUROC and 0.5127 ± 0.0102 AUPRC on MIMIC-IV, and 0.7821 ± 0.0035 and 0.2371 ± 0.0093 on eICU-CRD. We will distinguish seed variability from the aggregation across forecasting-backbone configurations used in the original probability-quality table.

W2 (b) and Q3. Small XGBoost gap, missing eICU comparison, and additional baselines.

We expanded the matched comparison with TabICLv2, DuETT, GARLIC, TPC, and ALFIA, and evaluated the baselines separately on both cohorts. The observation window is 0–36 hours and the endpoint is in-hospital mortality.

ModelMIMIC-IV AUROCMIMIC-IV AUPRCeICU-CRD AUROCeICU-CRD AUPRC
Logistic Regression0.81440.45110.77520.2312
XGBoost0.8343 ± 0.01260.4781 ± 0.01180.7788 ± 0.00290.2348 ± 0.0042
GRU-D0.8229 ± 0.00440.4684 ± 0.00770.7807 ± 0.00370.2367 ± 0.0071
Medformer0.6963 ± 0.00680.2985 ± 0.00250.6540 ± 0.00290.1101 ± 0.0016
TabICLv20.8317 ± 0.00170.4747 ± 0.00330.8019 ± 0.00090.2579 ± 0.0006
DuETT0.8364 ± 0.00140.4729 ± 0.00300.7971 ± 0.00290.2531 ± 0.0075
GARLIC0.8153 ± 0.00580.4406 ± 0.00830.7897 ± 0.00460.2457 ± 0.0123
TPC adapted0.8352 ± 0.00760.4797 ± 0.01240.7952 ± 0.00410.2512 ± 0.0018
ALFIA0.8058 ± 0.00500.4237 ± 0.01310.7636 ± 0.01770.2208 ± 0.0214
TRACE–iTransformer0.8443 ± 0.00310.5127 ± 0.01020.7821 ± 0.00350.2371 ± 0.0093

In the expanded repeated evaluation, TRACE exceeds XGBoost on MIMIC-IV by 0.0100 mean AUROC and 0.0346 mean AUPRC. Relative to TPC, which has the highest mean baseline AUPRC in that cohort, the AUPRC difference is 0.0330. These are new repeated evaluations, reported separately from the original single-run comparison.

The eICU-CRD results give a more mixed picture. TRACE is close to XGBoost and GRU-D in mean discrimination, and several additional methods obtain higher scores. TabICLv2, for example, achieves 0.8019 AUROC and 0.2579 AUPRC, compared with TRACE's 0.7821 and 0.2371. Its F1 is also higher, 0.3181 versus 0.2917. We will include these results and qualify the performance discussion accordingly.

The strongest support for the proposed combination is therefore the controlled result on both cohorts. After applying the same retrieval module to the directly trained classifier, the forecast-based pipeline retains an additional mean benefit. The expanded comparison separately establishes its relative performance against alternative predictors, including the methods that perform better on eICU-CRD. Together with the matched representation experiment, these results give a more precise basis for assessing TRACE's design and empirical contribution.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Remarks on the Main Reviewer Concerns

We thank the Area Chair for arranging the additional reviews and all reviewers for their time and feedback. We summarize below the additional evidence addressing the main concerns raised in the reviews.

1. Does forecasting contribute beyond retrieval, and what supports the proposed design? (Reviewers EKSJ, EphD)

A central concern was that retrieval accounts for the largest incremental gain, leaving the contribution of forecasting unresolved without a Direct + retrieval control.

We added this comparison using a directly trained iTransformer mortality classifier with the same retrieval estimate and weighting rule as TRACE. Both methods use the same patient-level split and 36-hour observation window. Results are mean AUPRC ± standard deviation across three training seeds.

DatasetDirect + retrievalTRACE
MIMIC-IV0.4909 ± 0.00790.5127 ± 0.0102
eICU-CRD0.2224 ± 0.00730.2371 ± 0.0093

Mean AUROC and F1 also improve on both datasets. These results directly address the missing comparison and support an additional contribution from the forecast-based pathway with retrieval held fixed.

To further examine the representation choice, we compared explicit forecast medians and interval widths against a 72-dimensional forecasting-pretrained encoder representation, sharing the forecasting checkpoint, full-history inputs, and downstream classifier design. Explicit forecasts achieve higher AUPRC with and without retrieval on both datasets.

Together, these controls support the contribution of explicitly representing anticipated physiology and forecast spread within the proposed framework.

2. Are the performance gains supported by repeated evaluation and sufficiently broad comparisons? (Reviewers EKSJ, 6G5Y)

We evaluated the main pipeline comparisons and expanded baselines across three seeds, reporting means and standard deviations. We added TabICLv2, DuETT, GARLIC, TPC, and ALFIA and evaluated the models separately on both datasets using matched observation windows, patient-level splits, and mortality endpoints.

On MIMIC-IV, TRACE achieves the highest mean AUROC and AUPRC among the evaluated models. Its mean AUPRC exceeds XGBoost by 0.0346 and TPC, the strongest baseline by mean AUPRC, by 0.0330. On eICU-CRD, several additional baselines outperform TRACE. The expanded evaluation therefore identifies cohort-dependent relative performance, while the matched Direct + retrieval comparison demonstrates an additional mean benefit from the forecast-based pathway on both cohorts.

These additions strengthen the empirical support for TRACE and clarify the scope of its contribution. We will incorporate the additional results, clarified experimental descriptions and training procedures, and the observed limitations into the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.