ICare-DT improves ICU forecasts with new data—but treatment prediction remains unproven

A patient’s digital twin should track their changing condition and help anticipate treatment responses. In an arXiv preprint submitted on November 6, 2026, Yanke Li and colleagues investigate these capabilities separately. Their findings support updating forecasts with new measurements, not a system ready to recommend treatments.

What the model predicts

ICare-DT uses a Transformer to turn an ICU record into a compact mathematical representation of the patient’s condition. Inputs include 41 physiological channels, 120 treatment channels, measurement indicators and four static characteristics. Each hour, it rereads the growing history with unchanged model weights, updating the patient representation without retraining.

The researchers trained five versions, differing in random seed, on 140,473 eICU-CRD stays, with patients separated across training, validation and test groups.

The main forecasts concern heart rate, respiratory rate and two oxygen-saturation channels crossing their training-set averages. The model also accounts for discharge and death occurring before a crossing. Crossing a population average is not necessarily a clinically dangerous event.

New measurements improve forecasts

The central comparison asked whether predictions made after 12, 18 or 30 hours beat predictions the same model made at hour six for the same future windows. These covered the following six, 12 or 24 hours, producing 12 target-and-window combinations.

Updated forecasts improved AUROC by 0.073 to 0.153 in all 12 settings, surviving correction for multiple comparisons. AUROC measures how well a forecast ranks cases with and without an outcome—not the percentage of patients correctly diagnosed.

Later forecasts are also closer to the events they predict. Validation experiments therefore withheld the new measurements or substituted another patient’s later hours. Reading the patient’s own measurements performed better throughout.

Those controls have limits. Artificially holding records constant can disadvantage a model not trained on such inputs. Another patient’s measurements are also a weak reference: simply using the patient’s own latest vital-sign value had a similar advantage over the other patient’s value. The experiments did not establish personalization beyond that latest reading.

The updated model nevertheless beat the latest-value baseline in 11 of 12 eICU test settings, in an unregistered comparison. But an untuned gradient-boosted model reading the same hours was stronger on validation data. The paper makes no state-of-the-art accuracy claim.

Transfer brings a narrower result

Applied to MIMIC-IV without refitting its weights, ICare-DT retained a significant updating advantage in eight of nine evaluable settings. However, against the latest-value baseline, it was significantly better in only one of nine settings and significantly worse in two after Holm correction for multiple comparisons. The remaining six comparisons were inconclusive, despite positive or negative point estimates. Measurement practices and oxygen-saturation channel definitions differed between databases.

A separate experiment refitted clinical-risk models within MIMIC-IV. Adding an endpoint-fine-tuned twin representation to gradient boosting improved acute-kidney-injury AUROC by 0.0035, with a 95% confidence interval of 0.0007 to 0.0065. The 0.0010 gain for hypotension was not significant. Clinical importance was not assessed.

Why treatment responses are harder

Routine records show outcomes under treatment actually given—not the same patient’s outcome under an alternative. Sicker patients also receive more treatment, potentially making beneficial interventions appear harmful.

The researchers built semi-synthetic experiments combining real physiological records with simulated dosing and injected responses of known size. These establish no actual drug benefit.

Measurement timing mattered. Heart-rate values changed in roughly 57% of hours in the examined records, versus about 4% for creatinine. Ridge models supervised with true simulated effects recovered heart-rate responses when reading inferred measurement times. None cleared the recovery threshold for creatinine at real cadence. These readers do not use the twin and cannot be deployed because real individual effects are unavailable as labels.

The main treatment-effect learner benchmark concerned just one selected setting: simulated phenylephrine response in heart rate, with an amplitude cap of 0.75 reference standard deviations. Researchers selected it after an exploratory sweep, at the smallest amplitude where the cadence-aware reader separated from its comparison reader.

In this setting, none of 16 scored standard treatment-effect learners significantly outperformed predicting zero effect at the main query. That query asked about sustained treatment after a washout—a regime no patient had received. No proposed comparison within the dosing period met the required overlap criteria. The benchmark therefore cannot separate learner failure from inadequate treatment variation at the query.

In that same setting, a learner supplied with the simulator’s response structure achieved R² values of 0.30–0.40 without true effect labels. That privileged knowledge limits practical relevance. Replication independence was also limited: although 395 of its 400 stays were new, the replication panel reused generator draws, including patient traits, and carried nearly identical injected effects to discovery. It tested recovery from different records, not independently generated effects.

A separate candidate, ICare-RH, worked in simulation but failed placebo checks on real-record panels and was not scored against their true effects.

Development evidence, not bedside validation

Reproducing measurement timing helped generated data pass a preset realism audit. But the response-panel audit missed substantial artificial changes, and flexible classifiers distinguished even the audited panel records from real ones. Passing did not certify realism.

The authors disclose prior analysis of evaluation splits and choices informed by earlier results. Registrations were internal version-control protocols, not external preregistrations. Forecast evaluation also selected frequently measured patients. These remain development findings: clinical targets, recalibration, prospective evaluation and trustworthy treatment-response prediction are needed before clinical use.