Background: Sepsis remains a significant challenge in patients with major trauma in the ICU. Early detection and treatment are crucial for improving outcomes and reducing mortality rates. Nonetheless, clinical tools for predicting sepsis among patients with major trauma are limited. This study aimed to develop and validate an artificial intelligence (AI) platform for predicting the risk of sepsis among patients with major trauma. Patients and methods: This study involved 961 patients, with a prospective analysis of data from 244 patients with major trauma at our hospital and a retrospective analysis of data from 717 patients extracted from a database in the United States. The patients from our hospital constituted the model development cohort, and the patients from the database constituted the external validation cohort. The patients in the model development cohort were randomly divided into a training cohort and an internal validation cohort at a ratio of 8:2. The machine-learning algorithms used to train models included logistic regression, decision tree, extreme gradient boosting machine (eXGBM), neural network (NN), random forest, and light gradient boosting machine (LightGBM). Results: The incidence of sepsis for the model development cohort was 43.44%. Twelve predictors, including gender, abdominal trauma, open trauma, red blood cell count, heart rate, respiratory rate, injury severity score, sequential organ failure assessment score, Glasgow coma scale, smoking, total protein concentrations, and hematocrit, were used as features in the final model. Internal validation showed that the NN model had the highest area under the curve (AUC) of 0.932 (95% CI: 0.917–0.948), followed by the LightGBM and eXGBM models with AUCs of 0.913 (95% CI: 0.883–0.930) and 0.912 (95% CI: 0.880–0.935), respectively. In the external validation cohort, the eXGBM model (AUC: 0.891, 95% CI: 0.866–0.914) had the highest AUC value, followed by the LightGBM model (AUC: 0.886, 95% CI: 0.860–0.906), and the AUC value of the NN model was only 0.787 (95% CI: 0.751–0.829). Considering the predictive performance for both the internal and external validation cohorts, the LightGBM model had the highest score of 82, followed by the eXGBM (81) and NN (76) models. Thus, the LightGBM has emerged as the optimal model, and it was deployed online as an AI application. Conclusions: This study develops and validates an AI application to effectively assess the susceptibility of patients with major trauma to sepsis. The AI application equips healthcare professionals with a valuable tool to promptly identify individuals at high risk of developing sepsis. This will facilitate clinical decision-making and enable early intervention.

Semi-supervised anomaly detection identifies abnormal (testing) observations which are different from normal (training) observations. In many practical situations, anomalies are poorly insufficient and not well defined, while the normal data are easily sampled, have a wide variety, and may not be classified. For this paradigm, we propose a novel end-to-end deep network as an anomaly detector only trained on normal samples. Our architecture consists of a conditional variational auto-encoder (CVAE), a feature discriminator (FD), and an adversarially trained WGAN-GP discriminator. The CVAE is designed as a generator to reconstruct images. It leverages underlying category information and multivariate Gaussian distributions to regularize the latent space, enabling a smooth and informative manifold. For anomalies which have a certain similarity to normal data, we perform active negative training by generating potential outliers from the latent space to limit network generative capability. In order to capture data characteristics, we maximize the mutual information between the inputs and the latent codes by the FD. It enhances the relationship between the high-dimensional image space and corresponding encoded vectors. To promote reconstruction, a structural similarity loss is applied to robustly recover local texture details and the WGAN-GP discriminator is employed to aid in generating photo-realistic images. We distinguish anomalies by computing a reconstruction-based anomaly score. Different from recent encoder-decoder or GAN-based architectures, our approach considers input categories, constructs, and exploits a useful manifold in an unsupervised manner and has a stronger reconstruction capability. The experimental results demonstrate that the proposed approach outperforms state-of-the-art methods over several benchmark datasets.

In this article, we extend the invariant extended Kalman filter (EKF) to light detection and ranging (LiDAR)-inertial odometry and mapping systems using invariant observer design and the theory of Lie groups for directly fusing LiDAR and inertial measurement unit (IMU) measurements. We consider this from two different aspects and implement two independent systems. Specifically, we propose a robo-centric invariant EKF LiDAR-inertial odometry termed Inv-LIO1. Its mapping module is an ordinary used one and two modules run in separate threads. A world-centric invariant EKF LiDAR-inertial odometry termed Inv-LIO2 is designed and implemented, which has an integrated odometry and mapping architecture. In Inv-LIO1, the output of the filter is the pose estimated by the scan-to-scan match method, which serves as the initial estimate of the mapping module that refines the odometry and constructs a 3-D map. The robo-centric formulation represents that the state in a local frame shifted at every LiDAR time to prevent filter divergence. Inv-LIO2 directly fuses LiDAR feature points and IMU data to obtain the map-refined odometry by scan-to-map match method, followed by global map update. To validate the effectiveness and robustness of the proposed method, we conduct extensive experiments in various indoor and outdoor environments. Overall, Inv-LIO1 offers pure odometry with higher accuracy than other state-of-the-art systems, improving the overall performance. Inv-LIO2 achieves superior accuracy over other state-of-the-art systems in the map-refined odometry comparison.

Pose graph optimization helps reduce drift accumulated in pure odometry of visual simultaneous localization and mapping (SLAM) systems by solving a nonlinear least square problem, including both sequential constraints and loop-closing constraints. However, the covariances of all constraints are set to constant matrices or by manual setting. In this paper, we propose a novel approach to approximate covariances of constraints in pose graph optimization to better represent the true uncertainty of the underlying visual-inertial navigation system (VINS) that fuses inertial measurements and visual observations. Specifically, for sequential constraints, we propose to utilize nonlinear factor recovery to optimally extract covariance matrices from the accumulated visual-inertial odometry (VIO). For loop-closing constraints, we propose a dynamic scale estimation method to approximate the scales of the information matrices. To evaluate the effectiveness and robustness of the proposed method, we conduct extensive experiments on public and self-collected datasets in various environments. Results show that our proposed method achieves higher accuracy compared with naively-formulated pose graph optimization adopted by several state-of-the-art visual-inertial navigation systems.

In an environment where robots coexist with humans, mobile robots should be human-aware and comply with humans' behavioural norms so as to not disturb humans' personal space and activities. In this work, we propose an inverse reinforcement learning-based time-dependent A* planner for human-aware robot navigation with local vision. In this method, the planning process of time-dependent A* is regarded as a Markov decision process and the cost function of the time-dependent A* is learned using the inverse reinforcement learning via capturing humans' demonstration trajectories. With this method, a robot can plan a path that complies with humans' behaviour patterns and the robot's kinematics. When constructing feature vectors of the cost function, considering the local vision characteristics, we propose a visual coverage feature for enabling robots to learn from how humans move in a limited visual field. The effectiveness of the proposed method has been validated by experiments in real-world scenarios: using this approach robots can effectively mimic human motion patterns when avoiding pedestrians; furthermore, in a limited visual field, robots can learn to choose a path that enables them to have the larger visual coverage which shows a better navigation performance.