acceptodds
Under review as a conference paper at ICLR 2027

Large Language Models for Multimodal Irregular Time Series with Informative Sampling

Abstract

Multimodal irregular time series (MITS) consist of asynchronous observations from heterogeneous numerical and textual channels, such as lab measurements and clinical notes in electronic health records (EHR). LLMs are natural candidates for processing such heterogeneous data, yet existing MITS models rely on specialized fusion mechanisms rather than leveraging LLMs for unified MITS processing, and LLM-based approaches for irregular time series address purely numerical observations. Moreover, no prior work has examined whether LLMs can learn to use informative sampling patterns. We introduce MILM (Multimodal Irregular time series Language Model), which serializes MITS into a time-ordered sequence of time-channel-value triplets and fine-tunes an LLM through a two-stage strategy for MITS classification. The first stage trains on value-redacted MITS to predict from sampling patterns alone, and the second stage trains on full MITS to jointly model sampling patterns and observed values. MILM achieves the best average rank of 1.1 against 15 baselines on 4 EHR datasets, improving average precision by up to 3.1 points over the strongest baseline. The two-stage strategy also outperforms directly fine-tuning on full MITS in a single stage (MILM-Direct), even when the single-stage baseline receives the same total number of training epochs. Value redaction evaluations confirm that sampling patterns carry predictive signal and that MILM learns to exploit them. Under the value pending evaluation we introduce, where some values are unavailable at prediction time, MILM remains ahead of MILM-Direct when the affected observations are dropped, and its lead widens when their time and channel are shown. Our code is available at https://anonymous.4open.science/r/milm-7750.

Primary area:
learning on time series and dynamical systems
Keywords:
Multimodal Irregular Time Series, Large Language Model

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_1ZTcOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces MILM, which classifies multimodal irregular time series. it first searializes every observation as an tine-ordered XML time–channel–value triplet, and fine-tunes Qwen3-4B with LoRA. Stage 1 of the process trains on value-redacted sequences so the model learns from sampling patterns. Stage 2 continues on full sequences. Predictions are read from logits scores after an <answer> tag in the output. Experiments are conducted on MIMIC-IV and eICU for in-hospital mortality and length-of-stay prediction. MILM obtains an average rank of 1.1 against baselines on 4 EHR datasets.. Several ablations are done. Value-redaction and a new "value pending" evaluation are also reported in the paper, as well as a Yelp restaurant-closure experiment in the appendix showing transferability of the approach to another domain.

Critical strengths

Simple method with well-designed control and useful eval. MILM-Direct-Eq and the channel- vs time-ordered ablation are the right methods to support the claims. Good reporting practices are employed: 5 runs with SDs, hyperparameter search based on validation AUROC, wall-clock and memory are reported for all methods along with prompts, code release. Data was used under DUA with local compute only.

Critical weaknesses

There is only weak evidence that the proposed two-stage "informative sampling" strategy is the key point here. Most of the advantage over baselines may come from LLM serialization itself. MILM-Direct already beats the best baseline on 6 of 8 metrics and MILM improves over MILM-Direct by only 0.17–1.46 points.

cohort construction and splits should be revisited and limit results generalizability. Splits are random at ICU-stay level and not by patient, however, MIMIC and eICU contain multiple stays per patient, so patient-level leakage is possible. It is not stated whether the five runs vary the split or only the seed. eICU spans 100s of hospitals, a random split lets a model learn site-specific practices. The conclusion lists cross-hospital shift only as future work.

The cohorts are heavily filtered.

Questions for authors

Are splits patient-disjoint? If not, please re-run all experiments with patient-grouped splits (and, for eICU, a hospital-held-out split).

Were eICU note types audited for treatment-limitation content?

Would appreciate Table 1 to also be reported under realistic value-pending setting for all methods (not only MILM vs MILM-Direct).

4 comments

We thank the reviewer for the detailed review and for recognizing our experiment design. We address each point below.

W1: Contribution of Stage 1

Time-ordered serialization is itself a component of MILM alongside two-stage fine-tuning. MILM-Direct is our ablation that removes Stage 1. The question is therefore whether Stage 1 adds a reliable gain on top of the serialization. To test this, we ran two paired analyses that pool all 20 dataset-seed pairs: a paired t-test over seeds and a paired bootstrap over test stays (10,000 resamples).

Table R1. MILM minus the comparison method in points, with 95% intervals.

ComparisonMetricPaired t-test over seedsPaired bootstrap over test stays
MILM vs. MILM-DirectAU-ROC+1.10 [0.39, 1.82], p = 0.004+1.11 [0.40, 1.83], p = 0.003
AP+0.51 [−0.27, 1.29], p = 0.19+0.52 [−0.64, 1.70], p = 0.36
MILM vs. MILM-Direct-EqAU-ROC+1.21 [0.24, 2.18], p = 0.017+1.22 [0.44, 2.00], p = 0.002
AP+0.41 [−0.56, 1.39], p = 0.39+0.42 [−0.74, 1.64], p = 0.47

The AU-ROC gain is significant under both analyses, including against the epoch-matched MILM-Direct-Eq. Although its intervals are wider, MILM's mean AP is also higher in both comparisons. Stage 1 also helps when some values are not yet available at prediction time (Table 4). MILM keeps its lead over MILM-Direct when pending observations are dropped, and the lead widens in all four dataset-metric combinations when their time and channel are shown.

Q1 & W2: Patient-level splits, the five runs, and hospitals

The splits are random at the ICU-stay level, so some test stays belong to patients who also have stays in training or validation. These account for 9.0%, 6.7%, 9.8%, and 9.1% of the test stays on MIMIC-IV-IHM, MIMIC-IV-LOS, eICU-IHM, and eICU-LOS. We re-ran the evaluation of all methods after removing these stays, so that every test patient is unseen during training.

Table R2a. AU-ROC on the patient-disjoint test sets. Best results in bold, second-best in italics.

MethodMIMIC-IV-IHMMIMIC-IV-LOSeICU-IHMeICU-LOS
GRU-D76.97 ± 0.9369.68 ± 1.0066.54 ± 1.1860.05 ± 1.68
GRU-D + Note75.51 ± 1.4370.25 ± 2.5166.93 ± 0.9557.66 ± 3.34
mTAND78.42 ± 0.8970.48 ± 1.1568.24 ± 1.4158.01 ± 1.56
mTAND + Note78.87 ± 1.4873.99 ± 0.9068.30 ± 1.4465.21 ± 0.91
KEDGN80.38 ± 1.4672.12 ± 1.7472.88 ± 1.8964.87 ± 1.79
KEDGN + Note73.84 ± 3.1062.81 ± 5.3371.60 ± 1.2262.73 ± 2.33
UTDE75.84 ± 3.3375.53 ± 2.9173.41 ± 1.0164.06 ± 3.15
FuseMoE76.88 ± 1.8776.26 ± 0.4871.52 ± 1.0565.64 ± 2.08
VITAL-stats78.41 ± 0.7673.35 ± 0.9971.07 ± 1.3058.99 ± 2.21
VITAL-stats + Note78.98 ± 0.6773.67 ± 1.0670.12 ± 1.2559.30 ± 0.81
VITAL-LLM67.42 ± 2.7768.29 ± 1.8262.19 ± 3.0358.94 ± 2.04
VITAL-LLM + Note67.89 ± 2.6768.62 ± 1.2861.29 ± 2.7359.11 ± 2.23
ISTS-PLM81.24 ± 1.0773.96 ± 1.2269.58 ± 1.1062.63 ± 1.62
ISTS-PLM + Note81.16 ± 0.6973.13 ± 1.7569.59 ± 1.0164.15 ± 1.04
Qwen3-4B (zero-shot)70.8967.9866.4456.69
MILM82.49 ± 0.6979.75 ± 0.5773.00 ± 1.2766.01 ± 1.80

Table R2b. AP on the patient-disjoint test sets. Best results in bold, second-best in italics.

MethodMIMIC-IV-IHMMIMIC-IV-LOSeICU-IHMeICU-LOS
GRU-D56.15 ± 0.5846.98 ± 1.5736.18 ± 3.0252.04 ± 2.81
GRU-D + Note54.36 ± 3.9647.44 ± 2.9137.23 ± 0.9247.02 ± 3.87
mTAND58.66 ± 1.5948.68 ± 1.8638.38 ± 3.4551.81 ± 1.32
mTAND + Note60.41 ± 1.2953.93 ± 2.0739.87 ± 2.4553.44 ± 0.67
KEDGN61.78 ± 3.4248.97 ± 2.1443.73 ± 1.3056.15 ± 3.28
KEDGN + Note49.39 ± 3.8440.36 ± 5.0342.01 ± 1.0652.40 ± 3.06
UTDE57.23 ± 3.6256.61 ± 3.6042.26 ± 1.2051.64 ± 4.21
FuseMoE57.11 ± 1.5057.45 ± 1.6242.01 ± 1.4053.98 ± 3.01
VITAL-stats56.02 ± 2.3452.01 ± 3.2440.46 ± 2.6948.65 ± 1.48
VITAL-stats + Note55.95 ± 1.7253.70 ± 3.5439.24 ± 2.0049.66 ± 1.54
VITAL-LLM40.95 ± 3.1246.88 ± 3.4329.29 ± 2.2850.46 ± 2.45
VITAL-LLM + Note40.37 ± 3.5348.27 ± 2.1328.76 ± 2.3050.27 ± 1.80
ISTS-PLM61.20 ± 1.4953.82 ± 1.2538.64 ± 0.4954.15 ± 2.13
ISTS-PLM + Note60.90 ± 0.8853.01 ± 1.9937.86 ± 0.6554.35 ± 1.62
Qwen3-4B (zero-shot)49.9345.2131.5647.48
MILM66.14 ± 1.3160.30 ± 1.5943.28 ± 1.8756.97 ± 1.05

MILM remains first in 6 of the 8 columns and second in the other 2.

We also re-ran the evaluation on larger test sets built from patients who appear in neither training nor validation using relaxed cohort filters. Each set is three times the size of the original test set. For eICU, we additionally built hospital-held-out test sets that exclude every hospital present in training or validation.

Table R3a. AU-ROC on the new test sets. Best results in bold, second-best in italics.

MethodMIMIC-IV-IHM unseen patientMIMIC-IV-LOS unseen patienteICU-IHM unseen patienteICU-LOS unseen patienteICU-IHM unseen hospitaleICU-LOS unseen hospital
GRU-D76.52 ± 0.3364.35 ± 0.3265.69 ± 0.8656.09 ± 1.7566.65 ± 1.1256.81 ± 1.79
GRU-D + Note76.51 ± 0.8164.87 ± 1.5567.48 ± 0.8856.41 ± 0.7467.57 ± 1.0454.58 ± 2.76
mTAND76.23 ± 1.3063.53 ± 0.6964.41 ± 0.2658.65 ± 1.5168.62 ± 1.1758.61 ± 0.78
mTAND + Note78.29 ± 0.8164.03 ± 0.7665.26 ± 0.5359.19 ± 0.5968.36 ± 0.9655.96 ± 1.46
KEDGN78.34 ± 0.5864.04 ± 1.9270.12 ± 0.6659.54 ± 1.4772.31 ± 1.6559.80 ± 1.59
KEDGN + Note74.32 ± 2.6556.97 ± 4.4870.01 ± 1.0060.20 ± 0.6373.49 ± 0.5960.13 ± 1.44
UTDE77.60 ± 2.8567.55 ± 1.8867.03 ± 0.6660.31 ± 0.8367.69 ± 0.8255.93 ± 1.24
FuseMoE80.11 ± 1.5268.24 ± 0.7067.85 ± 0.5060.69 ± 0.3067.99 ± 0.8056.49 ± 1.48
VITAL-stats78.24 ± 0.7565.47 ± 0.8268.61 ± 0.8259.11 ± 1.2070.87 ± 1.5058.67 ± 0.37
VITAL-stats + Note78.77 ± 0.5165.44 ± 1.2968.87 ± 1.1759.64 ± 1.4070.55 ± 0.7659.31 ± 1.00
VITAL-LLM71.42 ± 1.0059.31 ± 1.1857.77 ± 0.4958.95 ± 1.2157.82 ± 0.8356.02 ± 1.06
VITAL-LLM + Note70.83 ± 0.8960.53 ± 1.0658.16 ± 0.7259.64 ± 1.0356.76 ± 0.5654.06 ± 0.73
ISTS-PLM78.78 ± 0.6065.83 ± 0.9168.84 ± 0.1360.40 ± 0.4873.12 ± 1.5361.91 ± 0.85
ISTS-PLM + Note78.87 ± 0.5265.73 ± 0.5769.40 ± 0.5260.44 ± 0.7872.81 ± 0.8061.81 ± 0.41
Qwen3-4B (zero-shot)75.6762.0564.8753.1270.5659.17
MILM84.20 ± 0.2870.36 ± 0.4371.63 ± 0.5162.69 ± 0.6675.57 ± 0.7860.66 ± 0.79

Table R3b. AP on the new test sets. Best results in bold, second-best in italics.

MethodMIMIC-IV-IHM unseen patientMIMIC-IV-LOS unseen patienteICU-IHM unseen patienteICU-LOS unseen patienteICU-IHM unseen hospitaleICU-LOS unseen hospital
GRU-D42.86 ± 0.9650.61 ± 0.9124.54 ± 1.0551.49 ± 0.8322.13 ± 1.5959.14 ± 1.24
GRU-D + Note40.82 ± 3.4050.74 ± 2.5526.15 ± 1.0550.98 ± 0.5722.07 ± 1.1257.28 ± 2.70
mTAND45.87 ± 1.3649.30 ± 0.8223.06 ± 0.8553.59 ± 1.0420.30 ± 1.5060.85 ± 0.42
mTAND + Note46.73 ± 1.5749.03 ± 1.2823.35 ± 0.8752.65 ± 1.0020.01 ± 1.0858.41 ± 1.26
KEDGN43.87 ± 1.9648.37 ± 2.4930.44 ± 1.3152.58 ± 1.3224.21 ± 3.0261.50 ± 1.29
KEDGN + Note36.09 ± 5.2942.94 ± 3.6030.37 ± 1.7753.97 ± 0.8725.44 ± 1.4061.72 ± 1.51
UTDE42.15 ± 6.8053.57 ± 2.2524.65 ± 0.4253.39 ± 1.4319.70 ± 0.5958.35 ± 1.25
FuseMoE47.04 ± 3.8754.39 ± 0.4825.67 ± 0.9254.17 ± 0.8618.31 ± 1.5458.58 ± 1.86
VITAL-stats45.25 ± 1.7752.53 ± 0.8726.98 ± 1.0952.41 ± 1.2521.48 ± 2.7860.19 ± 0.70
VITAL-stats + Note47.55 ± 0.7851.45 ± 1.2027.79 ± 1.0353.47 ± 1.7422.63 ± 1.4260.71 ± 0.97
VITAL-LLM28.44 ± 0.9944.75 ± 1.9020.18 ± 0.5052.94 ± 1.7116.17 ± 2.1158.42 ± 1.47
VITAL-LLM + Note27.09 ± 1.0045.53 ± 1.3221.19 ± 1.1553.32 ± 2.0016.59 ± 0.7656.45 ± 1.34
ISTS-PLM46.25 ± 1.2551.94 ± 0.5028.34 ± 0.9353.64 ± 0.6125.62 ± 0.8763.01 ± 0.75
ISTS-PLM + Note47.17 ± 0.6352.17 ± 1.0628.45 ± 0.5753.58 ± 0.8724.96 ± 0.8462.61 ± 0.61
Qwen3-4B (zero-shot)38.0548.2122.7848.7217.9460.61
MILM53.17 ± 0.6657.47 ± 1.0332.50 ± 0.9555.94 ± 0.8230.00 ± 0.7661.56 ± 0.70

MILM ranks first in both metrics on 5 of the 6 new test sets.

The five runs use the same split and differ in the training seed. All methods share this split and these seeds, so the comparisons are matched. We have updated Section 5.1 of the revised paper to state this more explicitly.

Q2: Treatment-limitation content in eICU notes

Our eICU extraction uses only the patient, lab, and note tables. Code status and end-of-life care plans are recorded in the carePlanGeneral and carePlanEOL tables, which no method in our study reads. The sentence in Appendix B.1 that describes the care plan as covering end-of-life planning refers to these unused tables, and we have corrected it in the revised paper. We also audited the notes we use with a keyword search. A line noting an end-of-life or care-limitation discussion appears in 5 of the 551 eICU-IHM test stays and 2 of the 408 eICU-LOS test stays. Excluding these stays changes AU-ROC by less than 0.7 points for every method. The ranking of all methods is unchanged on eICU-LOS, and the top four are unchanged on eICU-IHM.

Q3: Value pending results for all methods

Appendix E (Table 15) reports all baselines under drop observation. Show presence requires representing an observation without its value, which these baselines cannot do because they need a numeric value for every observed entry. In the revised paper, Section 5.5 now refers to Table 15 by name for better visibility.

We have added Tables R1 to R3 to the appendix in the revised paper.

We thank the reviewer again for suggesting the patient- and hospital-held-out evaluations, which strengthened the paper. We hope these results and clarifications address the reviewer's questions.

Thank you for the additional analysis. The patient-disjoint evaluation addresses my concern about leakage. The eICU note audit and Table 15 largely address my other questions. I remain confident in my rating.

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

This paper proposes MILM, a framework for classifying multimodal irregular time series (MITS) using large language models. The key motivation is that both observed values and sampling patterns, including observation timing and channel identities, may carry predictive information.

MILM represents numerical measurements and textual observations as chronologically ordered time-channel-value triplets and employs a two-stage fine-tuning strategy. The first stage trains the model using only timestamps and channel identities, encouraging it to learn informative sampling patterns. The second stage fine-tunes the model on complete observations to incorporate both sampling patterns and values.

The authors evaluate MILM on four clinical prediction datasets derived from MIMIC-IV and eICU, covering in-hospital mortality and length-of-stay prediction, as well as an additional Yelp restaurant closure prediction dataset. MILM achieves the best average rank among 16 methods on the clinical benchmarks. Ablation studies demonstrate the benefits of two-stage fine-tuning and chronological serialization. Additional value-redaction and value-pending evaluations investigate the model's ability to exploit sampling information when observation values are unavailable.

Critical strengths

1. A simple and effective approach to leveraging informative sampling patterns.

The two-stage fine-tuning strategy is intuitive and well motivated. By first training on value-redacted observations, the method explicitly encourages the model to extract predictive information from observation timing and channel identities. The comparison against direct fine-tuning, including an epoch-matched baseline, provides useful evidence that the improvement is not simply due to additional training.

2. Comprehensive empirical evaluation with informative diagnostic experiments.

The paper compares MILM against a broad set of conventional, LLM-based, and multimodal time-series baselines. Beyond the main results, the value-redaction and value-pending evaluations provide useful evidence about what the model learns and how it behaves when observation values are unavailable. The additional experiment on Yelp data further supports the potential applicability of the method beyond healthcare.

Critical weaknesses

1. The improvements from the proposed two-stage training strategy are relatively modest, and their statistical significance is unclear.

While MILM consistently outperforms MILM-Direct, the gains are often small relative to the reported standard deviations. For example, the improvement in MIMIC-IV-IHM AU-ROC is only 0.48 points (81.30 vs. 80.82), and the corresponding AP improvement is 0.17 points (63.86 vs. 63.69). This raises questions about whether the additional training stage provides a sufficiently meaningful improvement under standard evaluation conditions. A paired statistical significance analysis would help establish the reliability of these gains.

2. The clinical evaluation protocol and computational trade-offs require stronger justification.

The datasets are randomly partitioned at the ICU-stay level rather than explicitly at the patient level, potentially allowing multiple stays from the same patient to appear across training and testing. Furthermore, the standard MIMIC-IV evaluation uses observations recorded within the first 24 hours, although some values may not have been available by the prediction time, as acknowledged in the value-pending experiments. These choices may lead to optimistic estimates of real-world performance.

In addition, MILM requires substantially more computational resources than most conventional baselines. Although the appendix reports training time and GPU memory usage, the paper does not fully establish whether the predictive gains justify this additional computational cost.

Questions for authors

Q1. Statistical significance of the two-stage improvement: The improvements over MILM-Direct are consistent but often modest relative to the reported standard deviations. Could the authors provide paired statistical significance tests across random seeds or data splits? This would help clarify whether the proposed training strategy offers reliable improvements beyond experimental variability.

Q2. Patient-level data splitting and temporal leakage: Appendix B indicates that train/validation/test splits are performed at the ICU-stay level. Can the authors clarify whether multiple stays from the same patient can appear in different splits? Additionally, since the standard MIMIC-IV evaluation uses charttime rather than result availability time, how should the main results be interpreted in the context of real-time clinical prediction?

Q3. Contribution of sampling information versus simpler sampling statistics: Could the authors compare MILM against a baseline using simple sampling-derived features, such as per-channel observation counts, measurement frequencies, or time since the last observation? Such a comparison would help determine whether the gains arise from learning complex temporal sampling patterns or from exploiting simpler statistical signals already present in the observation process.

3 comments

We thank the reviewer for the thoughtful review and for finding the two-stage design well motivated and the diagnostic experiments informative. We address each point below.

Q1 & W1: Statistical significance of the two-stage improvement

To test whether the gain is reliable, we ran two paired tests that pool all 20 dataset-seed pairs: the paired t-test over seeds that the reviewer suggested and a paired bootstrap over test stays (10,000 resamples).

Table R1. MILM minus the comparison method in points, with 95% intervals.

ComparisonMetricPaired t-test over seedsPaired bootstrap over test stays
MILM vs. MILM-DirectAU-ROC+1.10 [0.39, 1.82], p = 0.004+1.11 [0.40, 1.83], p = 0.003
AP+0.51 [−0.27, 1.29], p = 0.19+0.52 [−0.64, 1.70], p = 0.36
MILM vs. MILM-Direct-EqAU-ROC+1.21 [0.24, 2.18], p = 0.017+1.22 [0.44, 2.00], p = 0.002
AP+0.41 [−0.56, 1.39], p = 0.39+0.42 [−0.74, 1.64], p = 0.47

The AU-ROC gain is significant under both analyses, including against MILM-Direct-Eq, so it does not come from additional training. Although its intervals are wider, MILM's mean AP is also higher in both comparisons. The value pending evaluation (Table 4) shows a further benefit of Stage 1. MILM keeps its lead over MILM-Direct when pending observations are dropped, and the lead widens in all four dataset-metric combinations when their time and channel are shown.

Q2 & W2: Patient-level splitting

The splits are random at the ICU-stay level, so stays from the same patient can appear in different splits. These test stays account for 9.0%, 6.7%, 9.8%, and 9.1% of the test sets on MIMIC-IV-IHM, MIMIC-IV-LOS, eICU-IHM, and eICU-LOS. To measure their effect, we removed them and re-evaluated all methods on the remaining patient-disjoint test stays.

Table R2a. AU-ROC on the patient-disjoint test sets. Best results in bold, second-best in italics.

MethodMIMIC-IV-IHMMIMIC-IV-LOSeICU-IHMeICU-LOS
GRU-D76.97 ± 0.9369.68 ± 1.0066.54 ± 1.1860.05 ± 1.68
GRU-D + Note75.51 ± 1.4370.25 ± 2.5166.93 ± 0.9557.66 ± 3.34
mTAND78.42 ± 0.8970.48 ± 1.1568.24 ± 1.4158.01 ± 1.56
mTAND + Note78.87 ± 1.4873.99 ± 0.9068.30 ± 1.4465.21 ± 0.91
KEDGN80.38 ± 1.4672.12 ± 1.7472.88 ± 1.8964.87 ± 1.79
KEDGN + Note73.84 ± 3.1062.81 ± 5.3371.60 ± 1.2262.73 ± 2.33
UTDE75.84 ± 3.3375.53 ± 2.9173.41 ± 1.0164.06 ± 3.15
FuseMoE76.88 ± 1.8776.26 ± 0.4871.52 ± 1.0565.64 ± 2.08
VITAL-stats78.41 ± 0.7673.35 ± 0.9971.07 ± 1.3058.99 ± 2.21
VITAL-stats + Note78.98 ± 0.6773.67 ± 1.0670.12 ± 1.2559.30 ± 0.81
VITAL-LLM67.42 ± 2.7768.29 ± 1.8262.19 ± 3.0358.94 ± 2.04
VITAL-LLM + Note67.89 ± 2.6768.62 ± 1.2861.29 ± 2.7359.11 ± 2.23
ISTS-PLM81.24 ± 1.0773.96 ± 1.2269.58 ± 1.1062.63 ± 1.62
ISTS-PLM + Note81.16 ± 0.6973.13 ± 1.7569.59 ± 1.0164.15 ± 1.04
Qwen3-4B (zero-shot)70.8967.9866.4456.69
MILM82.49 ± 0.6979.75 ± 0.5773.00 ± 1.2766.01 ± 1.80

Table R2b. AP on the patient-disjoint test sets. Best results in bold, second-best in italics.

MethodMIMIC-IV-IHMMIMIC-IV-LOSeICU-IHMeICU-LOS
GRU-D56.15 ± 0.5846.98 ± 1.5736.18 ± 3.0252.04 ± 2.81
GRU-D + Note54.36 ± 3.9647.44 ± 2.9137.23 ± 0.9247.02 ± 3.87
mTAND58.66 ± 1.5948.68 ± 1.8638.38 ± 3.4551.81 ± 1.32
mTAND + Note60.41 ± 1.2953.93 ± 2.0739.87 ± 2.4553.44 ± 0.67
KEDGN61.78 ± 3.4248.97 ± 2.1443.73 ± 1.3056.15 ± 3.28
KEDGN + Note49.39 ± 3.8440.36 ± 5.0342.01 ± 1.0652.40 ± 3.06
UTDE57.23 ± 3.6256.61 ± 3.6042.26 ± 1.2051.64 ± 4.21
FuseMoE57.11 ± 1.5057.45 ± 1.6242.01 ± 1.4053.98 ± 3.01
VITAL-stats56.02 ± 2.3452.01 ± 3.2440.46 ± 2.6948.65 ± 1.48
VITAL-stats + Note55.95 ± 1.7253.70 ± 3.5439.24 ± 2.0049.66 ± 1.54
VITAL-LLM40.95 ± 3.1246.88 ± 3.4329.29 ± 2.2850.46 ± 2.45
VITAL-LLM + Note40.37 ± 3.5348.27 ± 2.1328.76 ± 2.3050.27 ± 1.80
ISTS-PLM61.20 ± 1.4953.82 ± 1.2538.64 ± 0.4954.15 ± 2.13
ISTS-PLM + Note60.90 ± 0.8853.01 ± 1.9937.86 ± 0.6554.35 ± 1.62
Qwen3-4B (zero-shot)49.9345.2131.5647.48
MILM66.14 ± 1.3160.30 ± 1.5943.28 ± 1.8756.97 ± 1.05

MILM remains first in 6 of the 8 columns and second in the other 2.

We also built larger test sets from patients who appear in neither training nor validation using relaxed cohort filters. Each set is three times the size of the original test set.

Table R3a. AU-ROC on the larger unseen-patient test sets. Best results in bold, second-best in italics.

MethodMIMIC-IV-IHMMIMIC-IV-LOSeICU-IHMeICU-LOS
GRU-D76.52 ± 0.3364.35 ± 0.3265.69 ± 0.8656.09 ± 1.75
GRU-D + Note76.51 ± 0.8164.87 ± 1.5567.48 ± 0.8856.41 ± 0.74
mTAND76.23 ± 1.3063.53 ± 0.6964.41 ± 0.2658.65 ± 1.51
mTAND + Note78.29 ± 0.8164.03 ± 0.7665.26 ± 0.5359.19 ± 0.59
KEDGN78.34 ± 0.5864.04 ± 1.9270.12 ± 0.6659.54 ± 1.47
KEDGN + Note74.32 ± 2.6556.97 ± 4.4870.01 ± 1.0060.20 ± 0.63
UTDE77.60 ± 2.8567.55 ± 1.8867.03 ± 0.6660.31 ± 0.83
FuseMoE80.11 ± 1.5268.24 ± 0.7067.85 ± 0.5060.69 ± 0.30
VITAL-stats78.24 ± 0.7565.47 ± 0.8268.61 ± 0.8259.11 ± 1.20
VITAL-stats + Note78.77 ± 0.5165.44 ± 1.2968.87 ± 1.1759.64 ± 1.40
VITAL-LLM71.42 ± 1.0059.31 ± 1.1857.77 ± 0.4958.95 ± 1.21
VITAL-LLM + Note70.83 ± 0.8960.53 ± 1.0658.16 ± 0.7259.64 ± 1.03
ISTS-PLM78.78 ± 0.6065.83 ± 0.9168.84 ± 0.1360.40 ± 0.48
ISTS-PLM + Note78.87 ± 0.5265.73 ± 0.5769.40 ± 0.5260.44 ± 0.78
Qwen3-4B (zero-shot)75.6762.0564.8753.12
MILM84.20 ± 0.2870.36 ± 0.4371.63 ± 0.5162.69 ± 0.66

Table R3b. AP on the larger unseen-patient test sets. Best results in bold, second-best in italics.

MethodMIMIC-IV-IHMMIMIC-IV-LOSeICU-IHMeICU-LOS
GRU-D42.86 ± 0.9650.61 ± 0.9124.54 ± 1.0551.49 ± 0.83
GRU-D + Note40.82 ± 3.4050.74 ± 2.5526.15 ± 1.0550.98 ± 0.57
mTAND45.87 ± 1.3649.30 ± 0.8223.06 ± 0.8553.59 ± 1.04
mTAND + Note46.73 ± 1.5749.03 ± 1.2823.35 ± 0.8752.65 ± 1.00
KEDGN43.87 ± 1.9648.37 ± 2.4930.44 ± 1.3152.58 ± 1.32
KEDGN + Note36.09 ± 5.2942.94 ± 3.6030.37 ± 1.7753.97 ± 0.87
UTDE42.15 ± 6.8053.57 ± 2.2524.65 ± 0.4253.39 ± 1.43
FuseMoE47.04 ± 3.8754.39 ± 0.4825.67 ± 0.9254.17 ± 0.86
VITAL-stats45.25 ± 1.7752.53 ± 0.8726.98 ± 1.0952.41 ± 1.25
VITAL-stats + Note47.55 ± 0.7851.45 ± 1.2027.79 ± 1.0353.47 ± 1.74
VITAL-LLM28.44 ± 0.9944.75 ± 1.9020.18 ± 0.5052.94 ± 1.71
VITAL-LLM + Note27.09 ± 1.0045.53 ± 1.3221.19 ± 1.1553.32 ± 2.00
ISTS-PLM46.25 ± 1.2551.94 ± 0.5028.34 ± 0.9353.64 ± 0.61
ISTS-PLM + Note47.17 ± 0.6352.17 ± 1.0628.45 ± 0.5753.58 ± 0.87
Qwen3-4B (zero-shot)38.0548.2122.7848.72
MILM53.17 ± 0.6657.47 ± 1.0332.50 ± 0.9555.94 ± 0.82

MILM ranks first on all four unseen-patient test sets in both metrics, showing that its advantage holds for patients the model has never seen.

Q2 & W2: Interpreting results under charttime

Following common convention, our main experiments build inputs from charttime. Because this treats every charted value as available at prediction time, the main results are an optimistic estimate of real-time performance for all methods. We designed the value pending evaluation in Section 5.5 to address this issue. It uses storetime to identify values that are not yet available at prediction time. This evaluation also indicates the size of the gap. Only about 3% of lab observations and fewer than one note per stay are pending (Table 13), and MILM's AU-ROC is 1.0 to 1.9 points lower than in the main results (Table 4). At prediction time, the time and channel of a pending observation are already known. MILM can use this information through show presence, whereas the baselines have to drop the observation because they require a numeric value for every observed entry (Appendix E.2). In this realistic comparison, MILM outperforms every baseline (Table 15).

Q3: Comparison with simple sampling statistics

We trained a model on hand-engineered sampling features. The features are computed for every channel:

  • observation count, i.e., the measurement frequency within the fixed 24-hour window
  • first observation time
  • mean and standard deviation of the time gaps between consecutive observations
  • time since the last observation

We also add counts across all channels, including the total number of lab observations and the numbers of lab observations and notes in each 6-hour block. Together with the per-channel features, this gives 121 features on MIMIC-IV and 136 on eICU. The model is an MLP with 3 hidden layers, trained with early stopping on validation AU-ROC.

Table R4. Sampling-feature MLP vs. MILM.

MethodMIMIC-IV-IHM AU-ROCMIMIC-IV-IHM APMIMIC-IV-LOS AU-ROCMIMIC-IV-LOS APeICU-IHM AU-ROCeICU-IHM APeICU-LOS AU-ROCeICU-LOS AP
Sampling-feature MLP66.83 ± 3.4340.61 ± 3.5471.00 ± 1.3047.39 ± 1.3761.63 ± 0.8931.70 ± 1.0960.99 ± 0.8653.21 ± 1.83
MILM81.30 ± 0.7063.86 ± 1.4780.02 ± 0.5559.73 ± 1.2172.47 ± 1.4942.27 ± 2.1066.05 ± 1.7956.22 ± 1.20

An MLP trained on sampling statistics alone reaches 61 to 71 AU-ROC, which supports the premise that the sampling process carries predictive signal. These statistics do not explain the performance of MILM, which exceeds the feature model by 5 to 14 AU-ROC points on every dataset. MILM instead reads observations in the LLM's native vocabulary and draws on pretrained knowledge to interpret lab values, channel names, and clinical notes in context, with no hand-designed features.

We have added Tables R1 to R4 to the appendix in the revised paper.

We thank the reviewer again for suggesting the paired significance tests and the sampling-statistics baseline, which strengthened the paper. We hope these results and clarifications address the reviewer's questions.

Reviewer_z3H9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes a multimodal time series foundation model that handles irregular time series. To do this, the work proposes a two-stage fine-tuning procedure that first trains with a template that has all time series values removed, before fine-tuning as normal on complete observations. The evaluation focuses primarily on Qwen3-4B with QLoRA on clinical datasets (MIMIC-IV and eICU), which shows promising performance compared to other baselines.

Critical strengths
  1. The design is simple, straightforward, and the work is presented clearly. The design makes sense and seems plausible, without being seemingly over engineered.

  2. Experiments cover a breadth of dimensions and are detailed. There is also an application to Yelp's data, showing the approach generalizes to other domains.

Critical weaknesses
  1. The methodological novelty is incremental and limited. The scenario explored (primarily LLMs for processing medical time series) has been explored in the past (e.g., MedTsLLM), and the methods used are primarily standard fine-tuning methods that are common in almost all foundation models now. The idea of masking out time series values to improve prediction performance has been applied across a wide range of scenarios, including irregular/missing time series (e.g., GRU-D). Moreover, the gains from stage 1 barely add 0.5-1% performance improvement over directly finetuning, showing that there is not much practical significance to the added fine-tuning step. This method also requires conforming to the specific template that is introduced in the paper, which doesn't seem generalizable in the long run.

  2. Evaluation can be strengthened. The clinical datasets used are small and heavily watered down from the MIMIC-IV and eICU. Also, experimental procedure can be more detailed, as the description on the data splits could still cause the same patients/settings to appear in both training and evaluation sets.

Questions for authors
  1. The main question is whether or not the proposed work is significantly novel from previous fine-tuning works. Based on the experimental results and the proposed approach, the answer is no, and the one part that may have some novelty (stage 1) seems to add very little to the overall performance.

  2. Are train/test splits done properly so that stays from the same patient/hosiptals don't appear in both training and testing data?

  3. Can the method hold up in performance under different template formats and patient/hospital held-out configurations? The main point I'm trying to assess here is the generalizability and robustness of the approach.

5 comments

We thank the reviewer for the careful review and for finding the method sensible and the experiments comprehensive. We address each point below.

Q1 & W1: Novelty

We would like to restate the scope of our contributions. First, this is the first study of LLMs for multimodal irregular time series (MITS) classification and the first investigation of whether LLMs can learn to exploit informative sampling. Second, MILM introduces a time-ordered serialization of MITS into time-channel-value triplets and value-redacted fine-tuning on this serialization, so the LLM learns from sampling patterns before it sees full observations. Third, we design two diagnostic evaluations, value redaction and value pending, to test whether the model uses sampling patterns.

As discussed in Section 2, MedTsLLM [1] and GRU-D [2] address different settings. MedTsLLM models regularly sampled waveforms and uses text as context, whereas MILM treats notes as observations within an irregular time series. GRU-D feeds the missingness mask together with the observed values, whereas Stage 1 of MILM removes all values so that the model must learn to predict from the sampling pattern alone before it sees any value. GRU-D is also one of our baselines, and MILM outperforms it on all four datasets (Table 1).

Q1 & W1: Contribution of Stage 1

The direct fine-tuning method referred to in the review is MILM-Direct, our ablation of MILM without Stage 1. It is built on our time-ordered serialization, so comparing it with MILM measures what Stage 1 adds on top of our own model. To test whether the gain is reliable, we ran two paired tests that pool all 20 dataset-seed pairs: a paired t-test over seeds and a paired bootstrap over test stays (10,000 resamples).

Table R1. MILM minus the comparison method in points, with 95% intervals.

ComparisonMetricPaired t-test over seedsPaired bootstrap over test stays
MILM vs. MILM-DirectAU-ROC+1.10 [0.39, 1.82], p = 0.004+1.11 [0.40, 1.83], p = 0.003
AP+0.51 [−0.27, 1.29], p = 0.19+0.52 [−0.64, 1.70], p = 0.36
MILM vs. MILM-Direct-EqAU-ROC+1.21 [0.24, 2.18], p = 0.017+1.22 [0.44, 2.00], p = 0.002
AP+0.41 [−0.56, 1.39], p = 0.39+0.42 [−0.74, 1.64], p = 0.47

The AU-ROC gain is significant under both tests, including against MILM-Direct-Eq, which receives the same total number of training epochs. Although its intervals are wider, MILM's mean AP is also higher in both comparisons. Stage 1 also matters in the value pending setting (Table 4), where some observations are recorded but their values are not yet available. MILM keeps its lead over MILM-Direct when pending observations are dropped, and the lead widens in all four dataset-metric combinations when their time and channel are shown.

Q3 & W1: Dependence on the prompt template

The prompt template mainly enables zero-shot prediction, and we document it in full for transparency. To test whether fine-tuning depends on it, we trained MILM with only the serialized time series as input, without the system prompt or task instructions.

Table R2. Full prompt vs. no instructions. Updated with all 5 runs on every dataset.

MethodMIMIC-IV-IHM AU-ROCMIMIC-IV-IHM APMIMIC-IV-LOS AU-ROCMIMIC-IV-LOS APeICU-IHM AU-ROCeICU-IHM APeICU-LOS AU-ROCeICU-LOS AP
Qwen3-4B zero-shot (full prompt)69.8248.3667.9844.8264.5329.4355.7145.58
MILM, full prompt81.30 ± 0.7063.86 ± 1.4780.02 ± 0.5559.73 ± 1.2172.47 ± 1.4942.27 ± 2.1066.05 ± 1.7956.22 ± 1.20
MILM, w/o instructions80.71 ± 0.4762.90 ± 0.8577.64 ± 3.0157.48 ± 3.1173.42 ± 1.2539.94 ± 2.6564.83 ± 3.2252.25 ± 2.56

Averaged over the four datasets, MILM improves AU-ROC over the zero-shot model by 10.45 points with the full prompt and by 9.64 points without instructions. MILM therefore retains most of its gain without the prompt template, so its performance does not substantially depend on this specific format.

Q2, Q3 & W2: Patient and hospital overlap across splits

Our splits are random at the ICU-stay level, so some test stays belong to patients who also have stays in training or validation. These account for 9.0%, 6.7%, 9.8%, and 9.1% of the test stays on MIMIC-IV-IHM, MIMIC-IV-LOS, eICU-IHM, and eICU-LOS. To measure their effect, we removed them and re-evaluated all methods on the remaining patient-disjoint test stays.

Table R3a. AU-ROC on the patient-disjoint test sets. Best results in bold, second-best in italics.

MethodMIMIC-IV-IHMMIMIC-IV-LOSeICU-IHMeICU-LOS
GRU-D76.97 ± 0.9369.68 ± 1.0066.54 ± 1.1860.05 ± 1.68
GRU-D + Note75.51 ± 1.4370.25 ± 2.5166.93 ± 0.9557.66 ± 3.34
mTAND78.42 ± 0.8970.48 ± 1.1568.24 ± 1.4158.01 ± 1.56
mTAND + Note78.87 ± 1.4873.99 ± 0.9068.30 ± 1.4465.21 ± 0.91
KEDGN80.38 ± 1.4672.12 ± 1.7472.88 ± 1.8964.87 ± 1.79
KEDGN + Note73.84 ± 3.1062.81 ± 5.3371.60 ± 1.2262.73 ± 2.33
UTDE75.84 ± 3.3375.53 ± 2.9173.41 ± 1.0164.06 ± 3.15
FuseMoE76.88 ± 1.8776.26 ± 0.4871.52 ± 1.0565.64 ± 2.08
VITAL-stats78.41 ± 0.7673.35 ± 0.9971.07 ± 1.3058.99 ± 2.21
VITAL-stats + Note78.98 ± 0.6773.67 ± 1.0670.12 ± 1.2559.30 ± 0.81
VITAL-LLM67.42 ± 2.7768.29 ± 1.8262.19 ± 3.0358.94 ± 2.04
VITAL-LLM + Note67.89 ± 2.6768.62 ± 1.2861.29 ± 2.7359.11 ± 2.23
ISTS-PLM81.24 ± 1.0773.96 ± 1.2269.58 ± 1.1062.63 ± 1.62
ISTS-PLM + Note81.16 ± 0.6973.13 ± 1.7569.59 ± 1.0164.15 ± 1.04
Qwen3-4B (zero-shot)70.8967.9866.4456.69
MILM82.49 ± 0.6979.75 ± 0.5773.00 ± 1.2766.01 ± 1.80

Table R3b. AP on the patient-disjoint test sets. Best results in bold, second-best in italics.

MethodMIMIC-IV-IHMMIMIC-IV-LOSeICU-IHMeICU-LOS
GRU-D56.15 ± 0.5846.98 ± 1.5736.18 ± 3.0252.04 ± 2.81
GRU-D + Note54.36 ± 3.9647.44 ± 2.9137.23 ± 0.9247.02 ± 3.87
mTAND58.66 ± 1.5948.68 ± 1.8638.38 ± 3.4551.81 ± 1.32
mTAND + Note60.41 ± 1.2953.93 ± 2.0739.87 ± 2.4553.44 ± 0.67
KEDGN61.78 ± 3.4248.97 ± 2.1443.73 ± 1.3056.15 ± 3.28
KEDGN + Note49.39 ± 3.8440.36 ± 5.0342.01 ± 1.0652.40 ± 3.06
UTDE57.23 ± 3.6256.61 ± 3.6042.26 ± 1.2051.64 ± 4.21
FuseMoE57.11 ± 1.5057.45 ± 1.6242.01 ± 1.4053.98 ± 3.01
VITAL-stats56.02 ± 2.3452.01 ± 3.2440.46 ± 2.6948.65 ± 1.48
VITAL-stats + Note55.95 ± 1.7253.70 ± 3.5439.24 ± 2.0049.66 ± 1.54
VITAL-LLM40.95 ± 3.1246.88 ± 3.4329.29 ± 2.2850.46 ± 2.45
VITAL-LLM + Note40.37 ± 3.5348.27 ± 2.1328.76 ± 2.3050.27 ± 1.80
ISTS-PLM61.20 ± 1.4953.82 ± 1.2538.64 ± 0.4954.15 ± 2.13
ISTS-PLM + Note60.90 ± 0.8853.01 ± 1.9937.86 ± 0.6554.35 ± 1.62
Qwen3-4B (zero-shot)49.9345.2131.5647.48
MILM66.14 ± 1.3160.30 ± 1.5943.28 ± 1.8756.97 ± 1.05

MILM still ranks first on every dataset except eICU-IHM, where it ranks second in both metrics.

To evaluate generalization to new patients and new hospitals at a larger scale, we built new test sets from patients who appear in neither training nor validation using relaxed cohort filters. Each set is three times the size of the original test set. For eICU, we also built unseen-hospital test sets that exclude every hospital present in training or validation.

Table R4a. AU-ROC on the new test sets. Best results in bold, second-best in italics.

MethodMIMIC-IV-IHM unseen patientMIMIC-IV-LOS unseen patienteICU-IHM unseen patienteICU-LOS unseen patienteICU-IHM unseen hospitaleICU-LOS unseen hospital
GRU-D76.52 ± 0.3364.35 ± 0.3265.69 ± 0.8656.09 ± 1.7566.65 ± 1.1256.81 ± 1.79
GRU-D + Note76.51 ± 0.8164.87 ± 1.5567.48 ± 0.8856.41 ± 0.7467.57 ± 1.0454.58 ± 2.76
mTAND76.23 ± 1.3063.53 ± 0.6964.41 ± 0.2658.65 ± 1.5168.62 ± 1.1758.61 ± 0.78
mTAND + Note78.29 ± 0.8164.03 ± 0.7665.26 ± 0.5359.19 ± 0.5968.36 ± 0.9655.96 ± 1.46
KEDGN78.34 ± 0.5864.04 ± 1.9270.12 ± 0.6659.54 ± 1.4772.31 ± 1.6559.80 ± 1.59
KEDGN + Note74.32 ± 2.6556.97 ± 4.4870.01 ± 1.0060.20 ± 0.6373.49 ± 0.5960.13 ± 1.44
UTDE77.60 ± 2.8567.55 ± 1.8867.03 ± 0.6660.31 ± 0.8367.69 ± 0.8255.93 ± 1.24
FuseMoE80.11 ± 1.5268.24 ± 0.7067.85 ± 0.5060.69 ± 0.3067.99 ± 0.8056.49 ± 1.48
VITAL-stats78.24 ± 0.7565.47 ± 0.8268.61 ± 0.8259.11 ± 1.2070.87 ± 1.5058.67 ± 0.37
VITAL-stats + Note78.77 ± 0.5165.44 ± 1.2968.87 ± 1.1759.64 ± 1.4070.55 ± 0.7659.31 ± 1.00
VITAL-LLM71.42 ± 1.0059.31 ± 1.1857.77 ± 0.4958.95 ± 1.2157.82 ± 0.8356.02 ± 1.06
VITAL-LLM + Note70.83 ± 0.8960.53 ± 1.0658.16 ± 0.7259.64 ± 1.0356.76 ± 0.5654.06 ± 0.73
ISTS-PLM78.78 ± 0.6065.83 ± 0.9168.84 ± 0.1360.40 ± 0.4873.12 ± 1.5361.91 ± 0.85
ISTS-PLM + Note78.87 ± 0.5265.73 ± 0.5769.40 ± 0.5260.44 ± 0.7872.81 ± 0.8061.81 ± 0.41
Qwen3-4B (zero-shot)75.6762.0564.8753.1270.5659.17
MILM84.20 ± 0.2870.36 ± 0.4371.63 ± 0.5162.69 ± 0.6675.57 ± 0.7860.66 ± 0.79

Table R4b. AP on the new test sets. Best results in bold, second-best in italics.

MethodMIMIC-IV-IHM unseen patientMIMIC-IV-LOS unseen patienteICU-IHM unseen patienteICU-LOS unseen patienteICU-IHM unseen hospitaleICU-LOS unseen hospital
GRU-D42.86 ± 0.9650.61 ± 0.9124.54 ± 1.0551.49 ± 0.8322.13 ± 1.5959.14 ± 1.24
GRU-D + Note40.82 ± 3.4050.74 ± 2.5526.15 ± 1.0550.98 ± 0.5722.07 ± 1.1257.28 ± 2.70
mTAND45.87 ± 1.3649.30 ± 0.8223.06 ± 0.8553.59 ± 1.0420.30 ± 1.5060.85 ± 0.42
mTAND + Note46.73 ± 1.5749.03 ± 1.2823.35 ± 0.8752.65 ± 1.0020.01 ± 1.0858.41 ± 1.26
KEDGN43.87 ± 1.9648.37 ± 2.4930.44 ± 1.3152.58 ± 1.3224.21 ± 3.0261.50 ± 1.29
KEDGN + Note36.09 ± 5.2942.94 ± 3.6030.37 ± 1.7753.97 ± 0.8725.44 ± 1.4061.72 ± 1.51
UTDE42.15 ± 6.8053.57 ± 2.2524.65 ± 0.4253.39 ± 1.4319.70 ± 0.5958.35 ± 1.25
FuseMoE47.04 ± 3.8754.39 ± 0.4825.67 ± 0.9254.17 ± 0.8618.31 ± 1.5458.58 ± 1.86
VITAL-stats45.25 ± 1.7752.53 ± 0.8726.98 ± 1.0952.41 ± 1.2521.48 ± 2.7860.19 ± 0.70
VITAL-stats + Note47.55 ± 0.7851.45 ± 1.2027.79 ± 1.0353.47 ± 1.7422.63 ± 1.4260.71 ± 0.97
VITAL-LLM28.44 ± 0.9944.75 ± 1.9020.18 ± 0.5052.94 ± 1.7116.17 ± 2.1158.42 ± 1.47
VITAL-LLM + Note27.09 ± 1.0045.53 ± 1.3221.19 ± 1.1553.32 ± 2.0016.59 ± 0.7656.45 ± 1.34
ISTS-PLM46.25 ± 1.2551.94 ± 0.5028.34 ± 0.9353.64 ± 0.6125.62 ± 0.8763.01 ± 0.75
ISTS-PLM + Note47.17 ± 0.6352.17 ± 1.0628.45 ± 0.5753.58 ± 0.8724.96 ± 0.8462.61 ± 0.61
Qwen3-4B (zero-shot)38.0548.2122.7848.7217.9460.61
MILM53.17 ± 0.6657.47 ± 1.0332.50 ± 0.9555.94 ± 0.8230.00 ± 0.7661.56 ± 0.70

MILM ranks first in both metrics on 5 of the 6 new test sets.

We have added Tables R1 to R4 to the appendix in the revised paper.

We thank the reviewer again for the constructive feedback, which led to the additional evaluations above and strengthened the paper. We hope these results and clarifications address the reviewer's concerns.

References

[1] Chan et al. MedTsLLM: Leveraging LLMs for Multimodal Medical Time Series Analysis. MLHC 2024.

[2] Che et al. Recurrent Neural Networks for Multivariate Time Series with Missing Values. Scientific Reports 2018.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.