Can an ICU forecast model see kidney injury earlier?
A model trained on 13,277 ICU admissions forecast eight physiological targets and improved retrospective discrimination for 45 moderate-to-severe kidney-injury events in an external cohort. That small event count and untested workflow effects keep it far from clinical deployment.
The short version
- What happened
- OmniTFT trained on 13,277 MIMIC-IV admissions and was evaluated without retraining on 4,188 MIMIC-III CareVue and 23,238 eICU admissions.
- Who it affects
- ICU clinicians could gain trajectory forecasts across several measurements, but only if local validation shows that added information is timely and does not worsen alarm burden.
- What remains unproven
- The study was retrospective and no clinician acted on the forecasts.
Keep in mind
The study was retrospective and no clinician acted on the forecasts.
Only 45 of 4,685 stays had the primary stage 2-or-worse kidney-injury outcome.
Cohorts were enriched for sepsis, ARDS or AKI and do not represent all ICU patients.
External uncertainty bands were miscalibrated for some targets and performance varied by cohort.
All patient data were historical US records; clinical practice and documentation have changed.
Source analysis · 1 linked source. Primary analysed source date: 16 October 2026.
Read the full report · 7 min The evidence, methods, implications and limitations
The answer: a stronger retrospective signal, not a proven warning system
OmniTFT improved a retrospective model's ability to separate ICU stays that would and would not develop moderate-to-severe acute kidney injury over the next day. In an external eICU analysis, adding summaries of forecast creatinine values to the observed record increased AUROC from 0.693 to 0.763. At a review capacity of 30% of eligible stays, the forecast-augmented pipeline identified 73.3% of later events compared with 57.8% using the observed record alone.
The result is promising but fragile. Only 45 of 4,685 eligible stays developed the primary stage 2-or-worse outcome. That makes calibration, precision and subgroup estimates uncertain even when an AUROC difference reaches statistical significance. The study was retrospective; it did not send an alert, change treatment, reduce kidney injury or measure false-alarm burden for clinicians.
The peer-reviewed paper was published in Frontiers in Digital Health on 16 October by researchers at the University of Tokyo and Kazusa DNA Research Institute. It used public de-identified US critical-care databases. JST SPRING and a JSPS fellowship funded the work. The authors declared no commercial conflict; one author was a Frontiers editorial-board member, which the journal says did not affect review.[1]
The three cohorts and the prediction task
Only MIMIC-IV was used for development. Its 13,277 admissions were split by patient 70:20:10 for training, validation and internal testing. MIMIC-III CareVue contributed 4,188 older, non-overlapping admissions for temporal external validation. The multicentre eICU database contributed 23,238 admissions from a different hospital population. Neither external cohort was used for fine-tuning, hyperparameter selection or model choice.
All cohorts were restricted to adults meeting at least one definition for sepsis, acute respiratory distress syndrome or acute kidney injury. That disease enrichment is important: this is not evidence about an unselected ICU population. The model forecast blood pressure, heart rate, oxygen saturation, respiratory rate, temperature, creatinine, lactate and an oxygenation ratio derived from SpO2 and inspired oxygen.
Vital signs used 12 hours of history to forecast the next two hours on ten-minute grids. Laboratory values used 24 hours to forecast the next 24 hours on two-hour grids. Inputs were forward-filled only from observations available at or before each grid time. Missing values and scaling used parameters learned from the MIMIC-IV training set rather than statistics from external cohorts, reducing one common source of leakage.[1]
What the model added
OmniTFT extends the Temporal Fusion Transformer with one shared backbone and separate output heads for each target. Training deliberately sampled similar numbers of stable and volatile time windows. It also regularised rare categorical embeddings, encouraged concentration among groups of historical inputs and aligned short-range attention with changes in the learned patient representation. Each final target checkpoint contained about 1.83 to 1.85 million parameters.
On the MIMIC-IV internal test set, the model had the lowest mean absolute error among tested systems for five of eight targets, ranked second for creatinine and respiratory rate, and third for temperature. External errors generally worsened for several targets: for example heart-rate MAE rose from 4.12 beats per minute internally to 6.29 in MIMIC-III and 6.65 in eICU; oxygen-saturation MAE rose from 1.04 percentage points to 1.90 and 2.15. External validation therefore demonstrates transfer, not invariant performance.
Ablations across ten matched training seeds found that removing the balanced sampler increased mean absolute error by about 37% to 60% for blood pressure, heart rate, oxygen saturation and respiratory rate. Removing the grouped regulariser caused the largest deterioration for the other targets. These comparisons support the design choices within this dataset, but paired seed tests are not a substitute for prospective hospital validation.[1]
How the kidney-injury analysis was constructed
The researchers fixed a decision point 24 hours after ICU admission and excluded stays that already met stage 1 kidney-injury criteria. The primary outcome was incident stage 2 or worse AKI in the following 24 hours. Three pipelines used the same stays and labels: creatinine from hours 12 to 24, all observed creatinine up to hour 24, or the observed record plus summaries of OmniTFT's forecast for hours 24 to 48.
Within each setting, the downstream risk model was selected independently from logistic-regression and gradient-boosting candidates using hospital-grouped folds. This means the comparison tests complete pipelines rather than adding forecasts to one frozen risk model. The authors reported AUROC, average precision, Brier score, calibration and decision curves, with 2,000 clustered bootstrap resamples. Some secondary differences, including average precision and decision-curve estimates at several thresholds, remained uncertain.
Forecast uncertainty was not consistently calibrated after transfer. The reported 10th-to-90th percentile bands departed from their intended coverage for some external targets, including temperature and the oxygenation ratio. Before deployment, each hospital would need target-specific recalibration, monitoring for missingness and drift, and a workflow study measuring whether a forecast changes decisions without causing harmful distraction.[1]
What clinicians and health systems should take from it
The useful idea is a shared forecasting layer that can reuse one representation while respecting different clinical time scales. That could reduce the engineering burden of maintaining separate models for every measurement. It might also support a dashboard that displays expected trajectories instead of another binary alarm. The open code and public datasets make independent replication possible.
It is not ready for bedside care. None of the data is contemporaneous: MIMIC-IV covers one US academic centre from 2008 to 2019, MIMIC-III CareVue is older, and eICU covers 208 US hospitals from 2014 to 2015. Treatment, documentation and patient populations have changed. The model omitted continuous waveforms such as ECG and arterial-line traces, and no study tested clinician comprehension, alert fatigue or treatment effects.
The assessment would change with a preregistered silent deployment across several current hospitals, followed by a controlled trial if calibration and fairness remain acceptable. Investigators should publish event counts, alert workload, calibration by site and patient group, treatment changes, missed events and clinical outcomes. A higher AUROC alone cannot show that patients benefit.[1]