How can an explainable hybrid deep learning model predict prolonged ICU stay in sepsis?

A 4-day threshold defines the prediction target

The outcome is not death or sepsis onset but prolonged ICU length of stay, defined as more than 4 days in the ICU. The authors chose this cutoff because ICU length of stay in the MIMIC-IV development cohort was strongly right-skewed, with a median of 3.15 days and an interquartile range of 1.90 to 6.80 days; 4 days sits just above the median and below the 75th percentile, keeping enough events to train on. The paper notes that prior critical care studies used other cutoffs such as 6 or 8 days, so the 4-day label is cohort-specific rather than a universal definition.

That choice matters for reading every number that follows. A model predicting more than 4 days is answering a different question from one predicting more than 8 days, and its sensitivity and precision cannot be transferred to a longer threshold without new evidence. The same predefined threshold was applied unchanged to both external cohorts, which the authors describe as eliminating threshold-tuning bias.

Seventeen routine features carry the signal

The model uses 17 variables collected within the first 24 hours of ICU admission, split into SOFA-related measures and baseline clinical characteristics. The SOFA-related group covers vasopressor dose, mechanical ventilation, 24-hour urine output, serum total bilirubin, platelet count, serum creatinine, Glasgow Coma Scale, and the PaO2/FiO2 ratio. The clinical group covers age, sex, body mass index, systolic, diastolic and mean blood pressure, lactate, and atrial fibrillation.

The authors state that no data-driven feature selection was applied and the same predefined set was used across all models and all cohorts, which keeps comparisons consistent and avoids selection bias. This is also the paper's practical claim: the inputs are routine, structured tabular data available at admission, not waveforms or longitudinal time series, so the model is designed for early use rather than for retrospective analysis of a completed stay.

Gated fusion combines two tabular modeling styles

FT-TabNet joins two existing tabular architectures. TabNet contributes feature-attentive sparse selection, focusing the model on informative SOFA-related variables; FT-Transformer contributes transformer-based learning of interactions among features. An adaptive gated fusion step computes a weighted combination of the two representations, with weights that vary by patient, so the model can lean toward global interaction modeling or sparse feature selection depending on the input.

The fused representation passes through a fully connected prediction head with a sigmoid output, giving a probability between 0 and 1 for prolonged stay. Hyperparameters were tuned only inside the inner loop of a nested cross-validation framework, with early stopping based on a validation subset drawn from training data, so the held-out outer test folds were not used for tuning. The authors frame the fusion as the key methodological contribution rather than either architecture alone.

Discrimination holds across three cohorts

Development and internal validation used 4,391 sepsis-related ICU admissions from MIMIC-IV, where 40.6% had a prolonged stay. Nested cross-validation gave a mean outer-loop AUROC of 0.848 (95% CI 0.832 to 0.856), with an AUPRC of 0.780, positive predictive value of 0.691, sensitivity of 0.783 and specificity of 0.752 at the Youden threshold. The authors also report a high negative predictive value of 0.863, which they interpret as reliable identification of low-risk patients.

External validation used 728 admissions from Chungbuk National University Hospital, where 53.6% had prolonged stays, and 162 from Chungnam National University Hospital, where 38.3% did. AUROC was 0.823 in the first and 0.781 in the second. AUPRC was 0.844 and 0.841 respectively, which the authors describe as comparable across the external cohorts. The AUROC drop in the smaller CNUH cohort is the clearest sign that transportability is good but not uniform, and the authors attribute the variability to institutional differences and the smaller sample.

The hybrid model leads the baselines it was tested against

The comparison set included FT-Transformer, FT-Transformer with skip-connected tokens, TabNet, a multilayer perceptron, tree-based models, and a meta ensemble of logistic regression, random forest and gradient boosting. The meta ensemble reached an AUROC of 0.808 but had lower sensitivity and AUPRC than the hybrid model. The skip-token FT-Transformer reached 0.803 AUROC and 0.730 AUPRC, and standard FT-Transformer reached 0.779 and 0.677.

These gaps are the paper's evidence that combining the two representations beats either alone on this tabular task. The authors also argue the model outperforms SOFA alone, which they read as evidence that nonlinear interactions among routine variables carry prognostic information that traditional severity scores do not fully capture. That comparison is the authors' interpretation of their results, not a separate clinical trial of SOFA-based decision-making.

SHAP points to organ dysfunction, not spurious patterns

SHAP values attribute each patient's predicted risk to individual features, and the summary plots across all three cohorts consistently rank mechanical ventilation, Glasgow Coma Scale, PaO2/FiO2 ratio, bilirubin and vasopressor dose among the top contributors. The direction is clinically intuitive: more organ support, lower neurological function and worse respiratory function associate with higher predicted risk.

The authors treat this consistency across geographically distinct cohorts as support for mechanistic validity, arguing the model learned biologically plausible relationships rather than spurious statistical associations. That is an interpretation built on feature ranking and direction, not a causal claim about what causes prolonged stay. The SHAP output is also what makes the model patient-level interpretable, which the authors present as the reason a clinician could inspect an individual risk estimate rather than accept a black-box score.

Net benefit appears in a 0.10 to 0.40 risk window

Decision curve analysis compares the model's net benefit against treating everyone and treating no one. The hybrid model showed higher net benefit across a threshold probability range of roughly 0.10 to 0.40 in internal validation and in both external cohorts. The authors explain that in this setting lower-to-moderate thresholds are the clinically meaningful window, because actions such as closer monitoring, early multidisciplinary involvement or anticipatory resource planning may be justified even at modest predicted risk.

The internal and CNUH net benefit curves overlapped across that range, while CBNUH showed a modestly lower but still superior benefit relative to the reference strategies. This is the paper's most decision-relevant result: the model's value is framed as early risk stratification and selective identification of patients likely to need extended ICU care, not as a stand-alone trigger for a specific intervention.

Survivor-only cohorts limit the clinical reach

The analysis included only patients who survived their ICU stay, which the authors say was intended to characterize prolonged ICU resource use. They acknowledge that mortality is a competing trajectory that can truncate length of stay, so this restriction may limit generalizability to the full ICU population. The external cohorts are two Korean hospitals, and the authors note that differences in institutional practice patterns and data distributions may require local recalibration before deployment.

The authors also list practical barriers: real-time data extraction and standardization inside electronic health records, clinician trust in model outputs, and workflow design that avoids alert fatigue. They call for prospective evaluation and system-level integration before routine clinical use. These limits sit next to the performance claims rather than replacing them, and they define how far the 0.848, 0.823 and 0.781 AUROCs should be carried.

The 4-day label and the fusion weights are the priorities

The paper's distinctive contribution is not a new clinical score but a specific combination: a predefined 4-day outcome, 17 routine first-24-hour features, and patient-specific gated fusion of TabNet and FT-Transformer representations, with SHAP exposing the drivers. For a reader deciding whether this matters, the most important comparison is the gap between the hybrid model and the single-architecture baselines on the same features, and the most important boundary is the survivor-only, two-hospital external validation.

The practical reading is that this model is positioned as an early screening aid for extended ICU resource use in sepsis, strongest where the predicted risk falls in the 0.10 to 0.40 decision window and where the negative prediction is used to identify low-risk patients. Whether that positioning holds in a prospective setting, in non-survivor populations, or under a different length-of-stay threshold remains outside what this evidence can answer.

Sources used in this answer