Accurate and timely prediction of high-flow nasal cannula (HFNC) treatment failure in patients with acute hypoxemic respiratory failure (AHRF) can lower patient mortality. Previous studies have highlighted inconsistencies in the predictive performance of existing indices, such as ROX and mROX, which are limited by their reliance on oxygenation parameters alone. To address this, we developed a machine learning-based predictive model using temporal data from AHRF patients, aimed at facilitating quicker development of individualized treatment plans and intervention strategies for healthcare professionals. We extracted 15 non-invasive and 15 laboratory features, including patient demographic characteristics, Glasgow Coma Scale, blood gas analysis, chemical assay, and complete blood cell count features. In addition to five machine learning models and an ensemble classifier, an long short-term memory (LSTM) network was included to assess deep learning performance on time-series data. Our study enrolled 427 patients with 498 treatment records. The soft-voting ensemble algorithm achieved an optimal predictive performance with an AUC of 0.839 (95% CI 0.786–0.889) for the all-features model, while logistic regression using common features achieved an AUC of 0.767 (95% CI 0.704–0.825), outperforming ROX and mROX indices. Incorporating blood gas analysis features improved the non-invasive model’s performance by 0.104. This study introduces a machine learning model integrated with a dynamic real-time alert system for predicting HFNC treatment failure in AHRF patients, demonstrating improved performance over traditional indices in internal validation and showing potential for decision support in select healthcare settings.

Sedation is an essential component of the anesthesia process. Inadequate sedation during anesthesia increases the risk of patient discomfort, intraoperative awareness, and psychological trauma. Conventional electroencephalography (EEG) based depth of anesthesia monitoring is often impractical in out-of-hospital settings due to equipment limitations and signal artifacts. Alternative non-EEG-based approaches are therefore required. In this study, we developed a machine learning model to detect inadequate sedation using 27 feature parameters, including demographics, vital signs, and heart rate variability metrics, from the open-access VitalDB database. Patient states were defined as inadequate sedation when the bispectral index (BIS) > 60. We systematically evaluated four temporal windows and four algorithms, and assessed model interpretability using Shapley Additive Explanations (SHAP). The Light Gradient Boosting Machine (LGBM) achieved the best performance, with an area under the receiver operating characteristic curve (AUC) of 0.825 and an accuracy (ACC) of 0.741 using a 2 s time window. Extending the time window to 20 s improved both metrics by approximately 0.012. Feature selection identified 12 key parameters that maintained comparable accuracy, confirming robustness with reduced complexity. These findings demonstrate the feasibility of using non-EEG-based physiological data for real-time detection of inadequate sedation. The developed model is interpretable, resource-efficient, scalable, and shows strong potential for integration into portable monitoring systems in prehospital, emergency, and low-resource surgical settings.

Objectives: Multiple organ dysfunction syndrome (MODS) is a serious, prognostically poor complication in trauma sepsis. We developed an interpretable, multicenter-validated prediction model to enable early, individualized risk assessment and guide timely care. Methods: Using MIMIC-IV and eICU data, we built a pre-trained transfer-learning model with a separation processing strategy and assessed interpretability with SHAP. Results: Internal validation included 700 MIMIC-IV patients; external validation included 110 eICU patients. Across 6-, 12-, and 24-h prediction windows, the best pre-trained model achieved an average AUC of 0.906. Notably, fine-tuning on only 100 trauma sepsis cases (3.6% of the training set) still yielded an AUC of 0.846, surpassing the non-pre-trained model by 0.165. SHAP analysis further revealed that platelet count was one of the most important variables contributing to MODS prediction. Conclusions: Overall, the pre-trained MODS model demonstrated robust discrimination, generalizability, and clear interpretability in both internal and external validations, highlighting its portability and clinical potential for early identification of high-risk trauma sepsis patients.