The research team, led by investigators at Xiaogan Hospital Affiliated to Wuhan University of Science and Technology, built their models using clinical features recorded during the first 24 hours of ICU admission. Thirty-one candidate predictors were drawn from routinely obtainable data: demographics, vital signs, complete blood counts, liver and kidney function tests, electrolytes, metabolic parameters such as glucose and lactate, and therapeutic variables including mechanical ventilation, vasopressor use, and continuous renal replacement therapy. The definition of SALI was deliberately strict, requiring total bilirubin above 2.0 mg/dL and an international normalised ratio above 1.5, both documented within the first 24 hours after ICU admission. Patients with pre-existing chronic liver disease were excluded, ensuring the models target acute liver injury rather than cirrhosis or chronic hepatitis.

To develop and test the models, the researchers turned to MIMIC-IV, a freely accessible electronic health record database containing ICU data from 2008 through 2022. Among 94,458 adult ICU admissions, 34,216 met the Sepsis-3 criteria, and 4,241 met the prespecified SALI biochemical criteria. After excluding 2,691 patients for short ICU stays, excessive missingness, or pre-existing chronic liver disease, 1,550 adults remained. These were allocated to a training cohort of 1,119 patients and a held-out internal validation cohort of 431. The external validation cohort came from Xiaogan Central Hospital in Hubei, China, where 229 patients met the same eligibility criteria between January 2018 and October 2025. The 28-day in-hospital mortality rate in this external cohort was 45.9 percent, underscoring how lethal SALI remains.

The methodological approach was unusually thorough. Rather than testing a single algorithm, the team evaluated 100 candidate pipelines, combining ten feature selection methods with ten classifiers, including logistic regression, random forest, XGBoost, LightGBM, support vector machines, k-nearest neighbours, naive Bayes, decision trees, extra trees, and multilayer perceptrons. Each pipeline was assessed using stratified 10-fold cross-validation within the training cohort, and each prediction horizon was modelled independently. The final specifications retained all 31 predictors and selected LightGBM for 7-day mortality and logistic regression for 14- and 28-day mortality. Class imbalance was handled at the classifier level using balanced class weights, and no over-sampling or under-sampling was performed.

Internal validation performance was strong across all three horizons. The area under the receiver operating characteristic curve, or AUROC, reached 0.840 for 7-day mortality, 0.841 for 14-day mortality, and 0.850 for 28-day mortality on the held-out internal validation set. Cross-validation AUROCs within the training cohort were 0.829, 0.813, and 0.818, respectively. Event counts in the training set were 257, 348, and 418 deaths at 7, 14, and 28 days, providing substantial statistical power for model development. These figures suggest the models captured meaningful prognostic signal from first-24-hour data alone.

External validation told a more nuanced story. Applied without updating to the independent Chinese cohort, the models achieved external AUROCs of 0.720, 0.717, and 0.763 for 7-, 14-, and 28-day mortality, respectively. The drop from internal to external performance ranged from 0.066 to 0.106 across the three models, illustrating the effect of cross-system case mix differences. The two cohorts differed substantially in oxygenation, transaminases, lactate, blood counts, and use of mechanical ventilation, vasopressors, and CRRT. Despite these differences, the 28-day model demonstrated moderate transportability to one independent retrospective centre.

Perhaps the most clinically striking finding was the comparison against conventional bedside scores. In the external cohort, the 28-day machine learning model outperformed a modified SOFA score, which omitted the respiratory component and used binary organ components, achieving an AUROC of 0.763 versus 0.673, a statistically significant difference with P equal to 0.011. It also outperformed MELD-Na, a liver-specific scoring system, with an AUROC of 0.763 versus 0.636, P equal to 0.002. At 7 and 14 days, the differences against SOFA were not statistically significant, though the 14-day comparison against MELD-Na remained significant. The researchers caution that these comparisons address incremental prognostic information rather than complexity-matched equivalence, since the machine learning models integrated 31 variables while SOFA and MELD-Na rely on fewer.

A parsimonious sensitivity analysis added an important twist. A properly specified logistic regression model using only the SOFA component variables and corrected Glasgow Coma Scale achieved an external 28-day AUROC of 0.762, essentially matching the full machine learning model at 0.763, with no statistically significant difference. This suggests that greater model complexity did not translate into better external discrimination, and that a well-designed regression benchmark can rival more elaborate algorithms when the underlying feature set is the same. The strongest supported comparative claim is therefore that the model provides better discrimination than conventional bedside scores in this cohort, not that machine learning is inherently superior to every lower-dimensional regression approach.

Calibration, the alignment between predicted probabilities and observed event rates, revealed another layer of complexity. In the external cohort, calibration slopes were 0.851, 0.563, and 0.779 for the 7-, 14-, and 28-day models, respectively. The 14-day model was clearly miscalibrated. Exploratory logistic recalibration, fitted and evaluated on the same external cohort, improved the calibration slope from 0.563 to 0.970 and moved the intercept from −0.178 to approximately −0.01. However, the Brier score, a measure of overall probabilistic accuracy, remained essentially unchanged at 0.208 before and 0.209 after recalibration, indicating that correcting the slope did not meaningfully improve overall accuracy. Because recalibration was performed on the same cohort used for evaluation, the researchers treat this as exploratory model updating rather than independent validation.

To make the models interpretable rather than opaque, the team applied SHAP, or Shapley additive explanations, using model-matched methods for each algorithm. The feature rankings differed strikingly across prediction horizons. For the 7-day model, peripheral oxygen saturation, INR, and lactate ranked highest, suggesting that acute hypoxaemia, coagulation activation, and tissue hypoperfusion dominate early mortality risk. For the 14- and 28-day models, haematocrit, vasopressor use, and mean arterial pressure rose to prominence, patterns that may reflect capillary leak, fluid accumulation, and marrow suppression in prolonged critical illness. The prominence of INR among 7-day attributions is consistent with immunothrombosis, in which coagulation activation acts as an intravascular effector of innate immunity rather than solely as a downstream marker of illness severity. The researchers emphasise that SHAP values represent model attribution rather than causal effects, and that differences across horizons are descriptive and hypothesis-generating.

Sensitivity analyses supported the stability of the primary findings. Reclassification using survival-file-derived mortality dates, which can also capture post-discharge deaths, increased external AUROCs to 0.747, 0.765, and 0.799 at 7, 14, and 28 days, respectively. Excluding 16 discordant records yielded similar results. Treating clearly implausible monitor values as missing produced small, generally favourable changes, with the 28-day external AUROC increasing from 0.763 to approximately 0.771. Decision curve analysis showed potential net benefit across clinically relevant threshold probabilities, though this finding remains retrospective and threshold-dependent.

The study, reported in accordance with the TRIPOD+AI statement, distinguishes itself from prior SALI literature in two ways. First, most existing models target 28-day mortality as the sole endpoint and implicitly treat predictors as time-invariant, even when multiple timepoints are analysed. Second, most models are trained and validated within the same dataset, and external validation, particularly in non-Western populations, remains rare. A recent systematic review of artificial intelligence models for ICU sepsis outcomes emphasised that external validation, calibration, prospective evaluation, and workflow-level clinical utility remain recurrent gaps. Against this background, the multi-horizon design and geographically independent Chinese external cohort represent meaningful contributions rather than claims that machine learning itself is unprecedented in SALI prediction.

Two properties matter for potential clinical use beyond discrimination: feasibility and interpretability. Most of the 31 predictors are routinely obtainable within the first 24 hours of ICU admission, which may facilitate future evaluation of integration into electronic health record workflows. SHAP-derived patterns provide attribution information that may assist interpretation. However, because external validation consisted of one independent centre and no prospective clinical impact assessment was performed, the models should currently be regarded as research-stage risk stratification tools. At any future implementation stage, poor-quality or unavailable inputs should be flagged and handled according to prespecified pre-processing rules, and use should be limited to trained healthcare professionals with appropriate clinical oversight.

Several limitations temper the findings. The external validation cohort was a single-centre Chinese sample of 229 patients, limiting precision and generalisability. Predictors were restricted to the first 24 hours after ICU admission, so the analysis does not represent a dynamic prediction model using time-varying covariates. Patients discharged alive before the corresponding horizon were classified as survivors because post-discharge mortality was not ascertained, meaning the endpoints represent time-limited in-hospital mortality rather than complete 7-, 14-, or 28-day mortality. The modified SOFA comparator omits the respiratory component and should not be interpreted as conventional full SOFA. No formal a priori sample size calculation was performed for the external validation, and no prespecified subgroup fairness or health-inequality analysis was undertaken. The exclusion of patients with pre-existing chronic liver disease restricts the models to acute SALI.

The researchers conclude that the models showed moderate transportability to one independent retrospective cohort, and that multicentre prospective validation is needed before clinical implementation. The analysis code, cohort extraction script, pre-processing parameters, and frozen model objects have been deposited in a public repository to facilitate independent reproduction, while the external validation cohort data could not be deposited publicly because of institutional data governance restrictions. For now, the study offers a rigorous, transparent, and externally tested framework for mortality risk stratification in SALI across three time horizons, providing a foundation for the prospective multicentre evaluation that must precede any clinical deployment.

Subject of Research: Machine learning mortality prediction in sepsis-associated liver injury

Article Title: Development and external validation of interpretable machine learning models for 7-, 14-, and 28-day mortality in sepsis-associated liver injury

Article References: Yao, H., Li, Z., Lü, J., & Shen, W. (2026). Development and external validation of interpretable machine learning models for 7-, 14-, and 28-day mortality in sepsis-associated liver injury. Frontiers in Medicine, 13, Article 1942203. https://doi.org/10.3389/fmed.2026.1942203

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1942203

Keywords: machine learning, sepsis-associated liver injury, mortality prediction, external validation, SHAP, SOFA, MELD-Na, intensive care, Development, external, validation, interpretable