Time-Updated Prognostic Modeling in ICU Patients with Documented Coma or Unresponsiveness Using Routine Arterial Blood Gas Trajectories: An Exploratory Explainable Machine-Learning Study

Background/Objectives: Prognostication in ICU patients with documented coma or unresponsiveness is a high-stakes task that informs escalation of care, goals-of-care discussions, and family counselling. Conventional scores are often based on static snapshots and may not reflect early physiological evolution in heterogeneous real-world ICU populations. Routine arterial blood gases (ABG) and SpO2 are repeatedly measured during early ICU care and may capture clinically meaningful trajectories that can be leveraged by explainable machine learning. To develop and internally validate exploratory, time-updated explainable machine-learning models for ICU outcome in ICU patients with clinically documented coma or unresponsiveness using routine ABG/SpO2 measurements and physiological trajectories available at admission, 24 h, and 72 h, and to evaluate whether trajectory information adds prognostic information within a staged internal-validation framework. Methods: We conducted a retrospective single-centre study of 108 adult ICU patients with clinically documented coma or unresponsiveness. Predictors included demographics, comorbidity burden, COVID-19 status, baseline ABG/SpO2 at ICU admission, inflammatory and coagulation biomarkers, and derived ABG/SpO2 trajectory variables at 24 h and 72 h. Trajectory variables were defined as changes from admission to 24 h and to 72 h and were retained as missing when follow-up measurements were unavailable. The primary ICU-course outcome was ICU death versus transfer to ward. Three staged models were evaluated: Model A using baseline variables, Model B adding 24 h trajectory features, and Model C adding 72 h trajectory features. For each stage, models were analyzed with and without the derived respiratory_support index; models excluding respiratory_support were treated as the main interpretive analyses. Logistic regression, random forest, and gradient boosting (XGBoost) classifiers were assessed using repeated stratified 5-fold cross-validation with 20 repeats and aligned out-of-fold predictions. Performance was reported using AUC-ROC, precision–recall AUC, Brier score, and operating-point metrics; clinical utility was examined with decision-curve analysis. Model interpretation used SHAP and partial dependence plots. Robustness analyses included feature-exclusion sensitivity analysis for respiratory_support and a label-permutation sanity check. Results: ICU mortality was 65.7% (71/108). Follow-up ABG completeness was 75.9% at 24 h and 61.1% at 72 h. Because respiratory_support summarized the highest support level during the first 72 h and strongly separated outcome groups, models excluding respiratory_support were treated as the primary interpretive analyses. In the primary NoRS logistic-regression models, discrimination was moderate-to-strong, with AUC-ROC 0.822 for Model A_noRS, 0.848 for Model B_noRS, and 0.895 for Model C_noRS; bootstrap 95% confidence intervals were 0.739–0.897, 0.766–0.919, and 0.830–0.951, respectively. Measurement-availability sensitivity analyses and simple benchmark models were added to contextualize trajectory-related performance. Respiratory_support-enriched models were retained only as secondary severity-aware analyses, not as admission-only prediction models. Label permutation reduced discrimination toward chance (AUC ≈ 0.55). SHAP and partial-dependence analyses identified oxygenation variables, inflammatory burden, acid–base status, and ΔPaO2 at 72 h as clinically coherent contributors to predicted risk; when included, respiratory_support dominated feature attribution, consistent with its role as an organ-support intensity marker. Conclusions: In ICU patients with clinically documented coma or unresponsiveness, explainable machine-learning models using routine ABG/SpO2 trajectories within the first 72 h are feasible and may provide time-updated prognostic information, but the incremental value of trajectory-enriched models over simpler admission-only benchmarks remains unproven. Trajectory-enriched NoRS models retained meaningful discrimination after removing organ-support severity, suggesting a possible physiologically meaningful signal beyond support intensity alone, although definitive incremental value over parsimonious admission-only benchmarks was not established. These findings should be interpreted as exploratory and internally validated only; they do not establish a deployable ICU mortality score, do not demonstrate superiority over established ICU severity scores, and require external validation in larger multicentre cohorts before clinical deployment.