Background/Objectives: Sepsis remains one of the leading causes of mortality worldwide, characterized by a complex and heterogeneous clinical presentation. Despite advances in patient monitoring and biomarkers, early detection of sepsis in the intensive care unit (ICU) is often hampered by incomplete data and diagnostic uncertainty. In recent years, machine learning models have been proposed as predictive tools, but many function as opaque “black boxes”, meaning that humans are unable to understand algorithmic reasoning, poorly suited to the uncertainty-laden clinical environment of critical care. Even when post-hoc interpretability methods are available for these algorithms, their explanations often remain difficult for non-expert clinicians to understand. Methods: In this clinical perspective, we explore the specific advantages of probabilistic graphical models, particularly Bayesian Networks (BNs) and their dynamic counterparts (DBNs), for sepsis prediction. Results: Recent applications of AI models in sepsis prediction have demonstrated encouraging results, such as DBNs achieving an AUROC of 0.94 in early detection, or causal probabilistic models in hospital admissions (AUROC 0.95). These models explicitly represent clinical reasoning under uncertainty, handle missing data natively, and offer interpretable, transparent decision paths. Drawing on recent studies, including real-time sepsis alert systems and treatment-effect modeling, we highlight concrete clinical applications and their current limitations. Conclusions: We argue that BNs present a great opportunity to bridge the gap between artificial intelligence and bedside care through human-in-the-loop collaboration, transparent inference, and integration into clinical information systems. As critical care continues to move toward data-driven decision-making, Bayesian models may offer not only technical performance but also the epistemic humility needed to support clinicians facing uncertain, high-stakes decisions.

Background: High incidence of ventilator associated pneumonia (VAP) has been reported in critically ill patients with COVID-19. Among these patients, we aimed to assess the incidence, outcomes and risk factors of VAP recurrences. Methods: We conducted an observational retrospective study in three French intensive care units (ICUs). Patients admitted for a documented COVID-19 from March 2020 to May 2021 and requiring mechanical ventilation (MV) for ≥48 h were included. The study main outcome was the incidence of VAP recurrences. Secondary outcomes were the duration of MV, ICU and hospital length of stay and mortality according to VAP and recurrences. We also assessed the factors associated with VAP recurrences. Results: During the study period, 398 patients met the inclusion criteria. A total of 236 (59%) of them had at least one VAP episode during their ICU stay and 109 (46%) of these patients developed at least one recurrence. The incidence of VAP recurrence considering death and extubation as competing events was 29.6% (IC = [0.250–0.343]). Seventy-eight percent of recurrences were due to the same bacteria (relapses). Patients with a VAP recurrence had a longer duration of MV as compared with one VAP and no VAP patients (41 (25–56) vs. 16 (8–30) and 10 (5–18) days; p < 0.001) and a longer ICU length of stay (46 (29–66) vs. 22 (12–36) and 14 (9–25) days; p < 0.001). The 90-day mortality was higher in the recurrence group as compared with the no VAP group only (31.2 vs. 21.0% (p = 0.021)). In a multivariate analysis including bacterial co-infection at admission, the use of immunosuppressive therapies and the bacteria responsible for the first VAP episode, the duration of MV was the only factor independently associated with VAP recurrence. Conclusion: In COVID-19 associated respiratory failure, recurrences affected 46% of patients with a first episode of VAP. VAP recurrences were mainly relapses and were associated with a prolonged duration of MV and ICU length of stay but not with a higher mortality. MV duration was the only factor associated with recurrences.

Objectives: To describe clinical characteristics and management of intensive care units (ICU) patients with laboratory-confirmed COVID-19 and to determine 90-day mortality after ICU admission and associated risk factors. Methods: This observational retrospective study was conducted in six intensive care units (ICUs) in three university hospitals in Marseille, France. Between 10 March and 10 May 2020, all adult patients admitted in ICU with laboratory-confirmed SARS-CoV-2 and respiratory failure were eligible for inclusion. The statistical analysis was focused on the mechanically ventilated patients. The primary outcome was the 90-day mortality after ICU admission. Results: Included in the study were 172 patients with COVID-19 related respiratory failure, 117 of whom (67%) received invasive mechanical ventilation. 90-day mortality of the invasively ventilated patients was 27.4%. Median duration of ventilation and median length of stay in ICU for these patients were 20 (9–33) days and 29 (17–46) days. Mortality increased with the severity of ARDS at ICU admission. After multivariable analysis was carried out, risk factors associated with 90-day mortality were age, elevated Charlson comorbidity index, chronic statins intake and occurrence of an arterial thrombosis. Conclusion: In this cohort, age and number of comorbidities were the main predictors of mortality in invasively ventilated patients. The only modifiable factor associated with mortality in multivariate analysis was arterial thrombosis.

Background: Ventilator-associated pneumonia (VAP) is a common and serious ICU complication, affecting up to 40% of mechanically ventilated patients. The diagnosis of VAP currently relies on retrospective clinical, radiological, and microbiological criteria, which often delays targeted treatment and promotes the overuse of broad-spectrum antibiotics. The early prediction of VAP is crucial to improve outcomes and guide antimicrobial use related to this disease. This study aimed to develop and validate PREDICT (Pneumonia Risk Evaluation and Diagnostic Intelligence via Computational Technology), a deep learning algorithm for early VAP prediction that is based solely on vital signs. Methods: We conducted a retrospective cohort study using the MIMIC-IV database, which includes ICU patients who were ventilated for at least 48 h. Five vital signs (respiratory rate, SpO₂, heart rate, temperature, and mean arterial pressure) were structured into 24 h temporal windows. The PREDICT model, based on a long short-term memory neural network, was trained to predict the onset of VAP 6, 12, and 24 h in the future. Its performance was compared to that of conventional machine learning models (random forest, XGBoost, logistic regression) using their AUPRC, sensitivity, specificity, and predictive values. Results: PREDICT achieved high predictive accuracy with AUPRC values of 96.0%, 94.1%, and 94.7% at 6, 12, and 24 h before the onset of VAP, respectively. Its sensitivity and positive predictive values exceeded 85% across all horizons. Traditional ML models showed a drop in performance over longer timeframes. Analysis of the model’s explainability highlighted the respiratory rate, SpO₂, and temperature as key predictive features. Conclusions: PREDICT is the first deep learning model specifically designed for early VAP prediction in ICUs. It represents a promising tool for timely clinical decision-making and improved antibiotic stewardship.