Li, Yanke, Qi, Yahang, Kim, Yumi, Paez-Granados, Diego · Nov 6, 2026
ICare-DT updates intensive-care forecasts from new patient data
Markdown · JSON · Original source
In an paper, Yanke Li, Yahang Qi, Yumi Kim and Diego Paez-Granados present ICare-DT, a digital twin that updates a patient state hourly as new measurements arrive, without retraining. Trained on 140,473 eICU-CRD stays, it forecast vital signs crossing their training-set means; later forecasts outperformed the model’s own hour-6 forecasts across 12 test settings. Validation controls that withheld or swapped new hours traced the gain to each patient’s own measurements. Without refitting, ICare-DT retained gains in 8 of 9 settings on a second database, in an evaluation using a previously analyzed MIMIC-IV test split. The paper draws a boundary between forecasting and treatment-response prediction: its uses real records altered with simulated treatment doses and injected effects. Standard learners did not beat predicting zero effect for a dosing regime no patient had received; a separate learner recovered simulated effects when given the way those effects were generated. The authors present routine records as support for forecasts that improve with new data, and the benchmark as a testbed for learning how responses are generated. These results concern the paper’s datasets, forecast targets and simulated-response benchmark; they do not establish performance in clinical use.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 6, 2026 (2026-11-06).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-09T05:49:29.033Z. First observed: 2026-11-09T05:07:20.937Z. Retrieved: 2026-11-09T05:48:26.315Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The response-prediction replaces treatments in real records with simulated doses and injects responses of known size; it is not evidence of demonstrated treatment-response prediction in clinical use. The abstract reports forecasting vital signs crossing their training-set means and results on specified datasets and settings; it does not establish performance in clinical use.
Open question
What approaches, beyond being told how the simulated responses were generated, can predict effects for dosing regimes no patient has received?
Source-stated
The authors propose ICare-DT, a Transformer-based twin that updates a latent patient state hourly from new measurements without retraining.
We propose ICare-DT, a Transformer-based twin that updates a latent patient state hourly from new measurements without retraining.
Source-stated
On test data, later forecasts beat the twin’s own hour-6 forecasts in all 12 settings, with reported AUROC gains of 0.073 to 0.153.
On test data, later forecasts beat the twin's own hour-6 forecasts of the same windows in all 12 settings (+0.073 to +0.153 AUROC).
Source-stated
Validation controls withholding or swapping new hours traced the forecasting gain to the patient’s own measurements.
On validation data, controls that withheld or swapped the new hours traced the gain to the patient's own measurements.
Source-stated
Without refitting, the model retained gains in 8 of 9 settings on a second database, evaluated on a previously analyzed MIMIC-IV test split.
Without refitting, the model kept its gain on a second database (8 of 9 settings, +0.025 to +0.114 AUROC) in a prespecified evaluation of the previously analyzed MIMIC-IV test split.
Source-stated
For simulated treatment responses at an out-of-support query, no standard learner outperformed predicting zero effect; a separate learner recovered effects when told how they were generated.
No standard learner outperformed predicting zero effect at an out-of-support query. When told how the simulated responses were generated, a separate learner recovered the effects
Source and contributor context
Li, Yanke, Qi, Yahang, Kim, Yumi, Paez-Granados, Diego · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Transformer architecture
The neural-network design behind most current language models; it weighs how each part of the input relates to the others.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: a2bfe5f0-3a6e-42dc-8dfe-59f76028385d.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T05:49:29.033Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: df6c668aea9340a22caef6bfe669f8d24ea3a0369d6136473ed9bc46b85eb001. Source text hash: 835a04c9ccd15cc45f333fed062ababa0b1263040d6cd3f55e9079e03e7b2f21.
Revision and source-check history
Version 1 · 2026-11-09T05:49:29.033Z
ICare-DT updates intensive-care forecasts from new patient data
In an paper, Yanke Li, Yahang Qi, Yumi Kim and Diego Paez-Granados present ICare-DT, a digital twin that updates a patient state hourly as new measurements arrive, without retraining. Trained on 140,473 eICU-CRD stays, it forecast vital signs crossing their training-set means; later forecasts outperformed the model’s own hour-6 forecasts across 12 test settings. Validation controls that withheld or swapped new hours traced the gain to each patient’s own measurements. Without refitting, ICare-DT retained gains in 8 of 9 settings on a second database, in an evaluation using a previously analyzed MIMIC-IV test split. The paper draws a boundary between forecasting and treatment-response prediction: its uses real records altered with simulated treatment doses and injected effects. Standard learners did not beat predicting zero effect for a dosing regime no patient had received; a separate learner recovered simulated effects when given the way those effects were generated. The authors present routine records as support for forecasts that improve with new data, and the benchmark as a testbed for learning how responses are generated. These results concern the paper’s datasets, forecast targets and simulated-response benchmark; they do not establish performance in clinical use.
Original source
cs.AI · Article
Source details
Original source date: Nov 6, 2026 (2026-11-06).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-09T05:49:29.033Z. First observed: 2026-11-09T05:07:20.937Z. Retrieved: 2026-11-09T05:48:26.315Z.
Limits
The response-prediction replaces treatments in real records with simulated doses and injects responses of known size; it is not evidence of demonstrated treatment-response prediction in clinical use. The abstract reports forecasting vital signs crossing their training-set means and results on specified datasets and settings; it does not establish performance in clinical use.
Open question
What approaches, beyond being told how the simulated responses were generated, can predict effects for dosing regimes no patient has received?
Source-stated
The authors propose ICare-DT, a Transformer-based twin that updates a latent patient state hourly from new measurements without retraining.
We propose ICare-DT, a Transformer-based twin that updates a latent patient state hourly from new measurements without retraining.
Source-stated
On test data, later forecasts beat the twin’s own hour-6 forecasts in all 12 settings, with reported AUROC gains of 0.073 to 0.153.
On test data, later forecasts beat the twin's own hour-6 forecasts of the same windows in all 12 settings (+0.073 to +0.153 AUROC).
Source-stated
Validation controls withholding or swapping new hours traced the forecasting gain to the patient’s own measurements.
On validation data, controls that withheld or swapped the new hours traced the gain to the patient's own measurements.
Source-stated
Without refitting, the model retained gains in 8 of 9 settings on a second database, evaluated on a previously analyzed MIMIC-IV test split.
Without refitting, the model kept its gain on a second database (8 of 9 settings, +0.025 to +0.114 AUROC) in a prespecified evaluation of the previously analyzed MIMIC-IV test split.
Source-stated
For simulated treatment responses at an out-of-support query, no standard learner outperformed predicting zero effect; a separate learner recovered effects when told how they were generated.
No standard learner outperformed predicting zero effect at an out-of-support query. When told how the simulated responses were generated, a separate learner recovered the effects
Source and contributor context
Li, Yanke, Qi, Yahang, Kim, Yumi, Paez-Granados, Diego · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Transformer architecture
The neural-network design behind most current language models; it weighs how each part of the input relates to the others.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: a2bfe5f0-3a6e-42dc-8dfe-59f76028385d.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T05:49:29.033Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: df6c668aea9340a22caef6bfe669f8d24ea3a0369d6136473ed9bc46b85eb001. Source text hash: 835a04c9ccd15cc45f333fed062ababa0b1263040d6cd3f55e9079e03e7b2f21.
uncertain · 2026-11-09T05:49:29.033Z · anthropic/claude-opus-5.5
Publication: a2bfe5f0-3a6e-42dc-8dfe-59f76028385d
attribution — pass: The paper and its claims are correctly attributed to the four listed authors on arXiv. Results are framed as what the paper reports, not as established fact.
Authors:Yanke Li, Yahang Qi, Yumi Kim, Diego Paez-Granados
summary — uncertain: The reading is mostly faithful but hard for a non-specialist to follow. Central terms such as digital twin, Transformer, eICU-CRD, MIMIC-IV, out-of-support and the AUROC metric are never explained; tooltips are needed. It also drops the condition that the separate learner recovered effects without seeing their true values, and only partly (R2 0.30 to 0.40 at one effect size). Without that, the result can read as trivial. Smallest repair: restore that condition and the partial R2. It also omits the refitted MIMIC-IV gradient-boosting results, including the acute kidney injury gain, and the prespecified nature of the evaluation.
a separate learner recovered the effects from the same records without seeing their true values (R2 0.30 to 0.40 at one effect size)
date — pass: The originalPublished date of 2026-11-06 matches the arXiv v1 submission date. It is not the later feed date.
[Submitted on 6 Nov 2026]
support — pass: Every claim excerpt is an exact substring of the abstract and supports its claim. The limitations correctly confine the results to the paper's datasets and its simulated benchmark, without implying clinical benefit. The question follows from the reported out-of-support failure and is not a disguised claim.
No standard learner outperformed predicting zero effect at an out-of-support query.
media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct way for a reader to reach the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivThe neural-network design behind most current language models; it weighs how each part of the input relates to the others.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivThe neural-network design behind most current language models; it weighs how each part of the input relates to the others.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference