Traumatic brain injury (TBI) and hemorrhagic shock (HS) are among the leading causes of trauma-related mortality worldwide. Studies have shown that when TBI is complicated by HS (TBI + HS), the mortality rate of patients increases significantly due to their complex pathophysiological interactions. Early and precise mortality risk assessment is essential for identifying high-risk TBI + HS patients, allocating healthcare resources, and implementing early treatments to improve clinical outcomes. This study proposes STran-TM, a novel mortality prediction framework designed to enhance prognostic accuracy in TBI + HS patients. STran-TM introduces two key innovations: the incorporation of missing-label variables to extract latent clinical semantics from missing data patterns, and the application of self-supervised learning (SSL) to pre-train a feature extractor for vital sign temporal representations. Using clinical data from the first 12 h after TBI + HS diagnosis, STran-TM predicts 48hour, 7-day, and in-hospital mortality. A total of 272 and 181 TBI + HS patients were selected from the MIMIC-IV and eICU databases, respectively. The model was trained and internally validated on MIMIC-IV, and externally validated on eICU. STran-TM achieved excellent predictive performance for 48-hour mortality (AUC: 0.9530 on MIMIC-IV, 0.8722 on eICU) and 7-day mortality (AUC: 0.8831 on MIMIC-IV, 0.8938 on eICU). Although performance slightly declined for in-hospital mortality (AUC: 0.8314 on MIMIC-IV, 0.8430 on eICU), the model maintained high reliability. By integrating missing-label variables and leveraging SSL, STran-TM enhances predictive performance and demonstrates robust generalizability across databases, offering a promising tool for early risk stratification in TBI + HS patients.

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.

IntroductionEmergency rescue scenes and pre-hospital emergency stages commonly encounter trauma victims. Life-saving measures must be taken at the scene if a trauma patient has pneumothorax; if the patient is not evaluated and diagnosed right away, their life may be in jeopardy. Ultrasound, which has the advantages of being non-invasive, non-radioactive, portable, rapid, and repeatable, can be used to diagnose pneumothorax. However, those who interpret ultrasound images must undergo extensive, specialized, and rigorous training. Deep learning technology allows for the intelligent diagnosis of ultrasound images, allowing general healthcare professionals to quickly and with minimal training diagnose pneumothorax in lung ultrasound patients.MethodsPrevious studies focused primarily on the lung-sliding characteristics of M-mode images, neglecting other key features in lung ultrasonography pneumothorax, and used similar technological techniques. Our study team used video understanding technology for medical ultrasound imaging diagnostics, training the TSM video understanding model on the ResNet50 network with 657 clips and testing the model with untrained 164 lung ultrasound clips.ResultsThe model’s sensitivity was 99.21%, specificity was 89.19%, and average accuracy was 96.95%. The F1 score was 0.929, and the AUC was 0.97.DiscussionThis study is the first to apply video understanding models to the multi-feature fusion diagnosis of pneumothorax, demonstrating the feasibility of using video understanding technology in medical image diagnosis.

OBJECTIVES:Accurately predicting the mortality risk of hypothermia patients is crucial for clinical decision-making, offering ample time for physicians to intervene. However, existing methods are invasive and difficult to implement in pre-hospital settings. METHODS:In this study, records of 2,700 hypothermia patients from 125 hospitals were extracted from the eICU Collaborative Research database, comprising 360 non-survivors and 2,340 survivors. Four machine learning methods were utilized to develop a mortality prediction model for hypothermia patients based on non-invasive physiological parameters. Data from 122 hospitals were used for model training, while the remainder were utilized for performance validation. RESULTS:The proposed machine learning prediction model achieved an area under the receiver operating characteristic curve (AUC) of 0.869 (95%CI: 0.840-0.895), demonstrating good mortality predictive performance for hypothermia patients only using non-invasive physiological parameters. Optimal and minimal feature subsets were selected for each machine learning method. The optimal feature subsets contained only 70.6% of the overall features for XGBoost models, while the AUC values increased by 0.039 compared to that of the entire feature subset. The interpretability analysis results highlight the vital importance of the temperature feature in predicting mortality during episodes of hypothermia in patients. CONCLUSIONS:This study developed a mortality prediction method for hypothermia patients only using non-invasive parameters, demonstrating robustness and reliability during multi-center validation. It can offer decision support for remote areas and disaster sites where it is difficult to access invasive parameters.

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.

Hypothermia, a component of the “lethal triad,” commonly complicates the condition of critically injured trauma patients, thereby substantially elevating the risk of mortality. This study develop and evaluate a dynamic warning system based on non-invasive features, aimed at predicting the likelihood of hypothermia occurring in trauma patients within the next hour. 462 patients from the eICU database were selected on the basis of meeting the inclusion criteria, and 19 non-invasive and 17 invasive features were extracted. Five classic machine learning methods were employed to develop dynamic early warning model for hypothermia based on various observation windows, with multi-center data used for model validation. The shapley additive explanations (SHAP) algorithm was utilized to analyze the interpretability of the model, and ablation experiments were conducted to further evaluate the contribution of significant feature to the prediction performance. The AUC values of the optimal models based on non-invasive features in the same test set are 0.838. When using cross-hospital data as the validation set, the highest AUC values for the same models based on non-invasive features decrease by only 0.015. In addition, ablation experiments reveal that the model’s AUC exhibited a 0.010 improvement when the three most influential invasive features were incorporated into the non-invasive feature set. The results show that machine learning models have shown significant potential in predicting hypothermia through the utilization of solely non-invasive features.

PURPOSE:Early mortality in major trauma has decreased, but MODS remains a leading cause of poor outcomes, driven by trauma-induced cytokine storms that exacerbate injuries and organ damage. METHODS:This prospective cohort study included 79 major trauma patients (ISS >15) treated in the National Center for Trauma Medicine, Peking University People's Hospital, from September 1, 2021, to July 31, 2023. Patients (1) with ISS >15 (according to AIS 2015), (2) aged 15-80 years, (3) admitted within 6 h of injury, (4) having no prior treatment before admission, were included. Exclusion criteria were (1) GCS score <9 or AIS score ≥3 for TBI, (2) confirmed infection, infectious disease, or high infection risk, (3) pregnancy, (4) severe primary diseases affecting survival, (5) recent use of immunosuppressive or cytotoxic drugs within the past 6 months, (6) psychiatric patients, (7) participation in other clinical trials within the past 30 days, (8) patients with incomplete data or missing blood samples. Admission serum inflammatory cytokines and pathophysiological data were analyzed to develop machine learning models predicting MODS within 7 days. LR, DR, RF, SVM, NB, and XGBoost were evaluated based on the area under the AUROC. The SHAP method was used to interpret results. RESULTS:This study enrolled 79 patients with major trauma, and the median (Q1, Q3) age was 51 (35, 59) years (52 males, 65.8%). The inflammatory cytokine data were collected for all participants. Among these patients, 35 (44.3%) developed MODS, and 44 (55.7%) did not. Additionally, 2 patients (2.5%) from the MODS group succumbed. The logistic regression model showed strong performance in predicting MODS. Ten key cytokines, IL-18, Eotaxin, MCP-4, IP-10, CXCL12, MIP-3α, MCP-1, IL-1RA, Cystatin C, and MRP8/14 were identified as critical to the trauma-induced cytokine storm and MODS development. Early elevation of these cytokines achieved high predictive accuracy, with an AUROC of 0.887 (95% CI 0.813-0.976). CONCLUSION:Trauma-induced cytokine storms are strongly associated with MODS. Early identification of inflammatory cytokine changes enables better prediction and timely interventions to improve outcomes.

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.

The use of invasive mechanical ventilation (IMV) is crucial in rescuing patients with respiratory dysfunction. Accurately predicting the demand for IMV is vital for clinical decision-making. However, current techniques are invasive and challenging to implement in pre-hospital and emergency rescue settings. To address this issue, a real-time prediction method utilizing only non-invasive parameters was developed to forecast IMV demand in this study. The model introduced the concept of real-time warning and leveraged the advantages of machine learning and integrated methods, achieving an AUC value of 0.935 (95% CI 0.933-0.937). The AUC value for the multi-center validation using the AmsterdamUMCdb database was 0.727, surpassing the performance of traditional risk adjustment algorithms (OSI(oxygenation saturation index): 0.608, P/F(oxygenation index): 0.558). Feature weight analysis demonstrated that BMI, Gcsverbal, and age significantly contributed to the model's decision-making. These findings highlight the substantial potential of a machine learning real-time dynamic warning model that solely relies on non-invasive parameters to predict IMV demand. Such a model can provide technical support for predicting the need for IMV in pre-hospital and disaster scenarios.

Objectives: Timely and accurate prediction of the requirement for invasive mechanical ventilation (IMV) can reduce patient mortality. Existing methods (traditional risk adjustment algorithms, clinical observation, et.) use laboratory parameters requiring specialized biochemical analysis, which is difficult to obtain in the pre-hospital emergency setting and does not accurately predict the requirement for IMV. Methods: In this study, 20 non-invasive parameters including patient demographic parameters, physiological parameters, Glasgow score and ventilator parameters, were extracted from the Medical Information Mart for Intensive Care III (MIMIC III) database. A real-time early warning model of IMV requirement was developed using classical seven machine learning methods in different categories and compared with two traditional risk adjustment algorithms. Results: The prediction results using Lightgbm were 0.917 (95 %CI:0.914-0.922) for area under receiver operating characteristic curve (AUC) and 0.853 for accuracy (ACC) (95 %CI:0.850-0.856), outperforming the traditional risk adjustment algorithm, which were 0.615 and 0.533 respectively. The addition of invasive parameters increased the AUC value of the model by 0.009. Conclusions: A real-time early warning algorithm was developed in this paper for IMV requirement based on non-invasive parameters using seven learning methods, which proved to be superior to the traditional risk adjustment algorithm. Using real-time clinical data, the proposed algorithm can calculate current and future requirement for IMV requirement at any point in time during the stay of a patient in the ICU. Finally, it provides technical support for a wide range of applications in remote areas and disaster sites, where invasive parameters are unavailable.

Hydrogel, with excellent hydrophilicity biocompatibility and stimuli responsive, has been widely used in biomedical and artificial organs fields. However, synthetic hydrogel usually displays poor gel strength, poor toughness and slow speed of absorbing water. Therefore, extensive exploratory research is still required to develop new composite materials for the application of bionic viscera. In this work, we constructed a new kind of composite hydrogel by a facile wet chemistry method to cross-link three kinds of polymer resin, polyvinyl alcohol, carboxymethyl cellulose, and carboxymethyl starch. The conductivity, moisture content and elastic modulus of hydrogel can be controlled by the dosage of potassium salt, sodium salt, water and glycerol. After molding and testing, the electrical conductivity, elastic modulus, and moisture content of the artificial liver and kidney organs were similar to those of the corresponding pig viscera, which is superior to most of the current research on hydrogels. This work provides a facile approach to synthesize highly efficient hydrogel, which accelerates the application of hydrogel in bionic organs.

OBJECTIVES Timely and accurate prediction of multiple organ dysfunction syndrome (MODS) is essential for the rescue and treatment of trauma patients However, existing methods are invasive, easily affected by artifacts and can be difficult to perform in a pre-hospital setting. We aim to develop prediction models for patients with MODS using only non-invasive parameters. METHOD In this study, records from 2319 patients were extracted from the Multiparameter Intelligent Monitoring in Intensive Care Ⅲ database (MIMIC Ⅲ), based on the sequential organ failure assessment (SOFA) score. Seven commonly used machine learning (ML) methods were selected and applied to develop a real-time prediction method for MODS based on full parameters (laboratory parameter. drug and non-invasive parameters, 57 parameters in total) and non-invasive parameters only (17 parameters) and compared with four traditional scoring systems. RESULTS The prediction results using LightGBM (LGBM) and Adaboost based on the full parameter modeling were 0.959 for area under receiver operating characteristic curve (AUC), outperforming four traditional scoring systems. The removal of 40 parameters and retaining of 17 non-invasive parameters decreased the AUC value of LGBM by 0.015, which still outperformed all traditional scoring systems. CONCLUSIONS A real-time and accurate MODS prediction method was developed in this paper based on non-invasive parameters by comparing the performance of four ML methods, which proved to be superior to the traditional scoring systems. This method can help medical staff to diagnose MODS as soon as possible and can improve the survival rate of patients in a pre-hospital setting.

在基于视觉的定位分析过程中,瞳孔作为重要的生理参数,易受鬓角、眉毛和光照强度等因素干扰,导致定位结果准确率不高,因此本文提出一种基于几何形状特征的实时瞳孔定位追踪技术.该技术方法结合两种不同定义的连通域外接矩形,通过面积阈值、角度差进行初步筛选,利用外接矩形长宽比进行二次筛选,最终实现瞳孔的精确定位.将该技术方法在5种不同数据集上进行准确性检测,定位结果的最高准确率提高到99.5％,平均误差减少到7.54％.由85名被试者在不同眼球注视情况下进行瞳孔实时定位追踪检测,结果表明该方法具有良好的鲁棒性,能够实现对瞳孔的实时追踪定位.

Product service system (PSS), an undeveloped business model, provides consumers with full-life cycle services from the perspective of products, functions and services. This paper proposes a posture recognition method based on smart product service system from the aspects of technology and user’ need. A Posture-CNN model based on the convolutional neural network (CNN) is established in this paper. According to the theory of PSS, this paper analyses the difficulties that the posture recognition technology is facing in the application of products, and resolves the existing problems such as low accuracy of recognition and unsupported complex environment. A self-made posture data set is tested with the results that the method can greatly reduce network parameters and improve network speed compared with other existing classification methods. To some extent, as a core technology of the PSS control terminal, this technology can perfect users’ experiences, better serve consumers and promote the development of the PSS.