The presented research faces the problem of early detection of sepsis for patients in the Intensive Care Unit. The PhysioNet/Computing in Cardiology Challenge 2019 facilitated the development of automated, open-source algorithms for the early detection of sepsis from clinical data. A labeled clinical records dataset for training and verification of the algorithms was provided by the challenge organizers. However, a relatively small number of records with sepsis, supported by Sepsis-3 clinical criteria, led to highly unbalanced dataset (only 2% records with sepsis label). A high number of unbalanced data records is a great challenge for machine learning model training and is not suitable for training classical classifiers. To address these issues, a method taking into the account the amount of time the patients spent in the intensive care unit (ICU) was proposed. The proposed method uses two separate ensemble models, one trained on patient records under 56 h in the ICU, and another for patients who stayed longer than 56 h. A solution including feature selection and weighting based training on imbalanced data was proposed in this paper. In addition, several performance metrics were investigated. Results show, that for successful prediction, a particular model having few or more predictors based on the length of stay in the Intensive Care Unit should be applied.

The presented research addresses the real-time object detection problem with small and moving objects, specifically the surface-mount component on a conveyor. Detecting and counting small moving objects on the assembly line is a challenge. In order to meet the requirements of real-time applications, state-of-the-art electronic component detection and classification algorithms are implemented into powerful hardware systems. This work proposes a low-cost system with an embedded microcomputer to detect surface-mount components on a conveyor belt in real time. The system detects moving, packed, and unpacked surface-mount components. The system’s performance was experimentally investigated by implementing several object-detection algorithms. The system’s performance with different algorithm implementations was compared using mean average precision and inference time. The results of four different surface-mount components showed average precision scores of 97.3% and 97.7% for capacitor and resistor detection. The findings suggest that the system with the implemented YOLOv4-tiny algorithm on the Jetson Nano 4 GB microcomputer achieves a mean average precision score of 88.03% with an inference time of 56.4 ms and 87.98% mean average precision with 11.2 ms inference time on the Tesla P100 16 GB platform.

Migraine is one of the most common neurological disorders, characterized by moderate-to-severe headache episodes. Autonomic nervous system (ANS) alterations can occur at phases of migraine attack. This study investigates patterns of ANS changes during the pre-ictal night of migraine, utilizing wearable biosensor technology in ten individuals. Various physiological, activity-based, and signal processing metrics were examined to train predictive models and understand the relationship between specific features and migraine occurrences. Data were filtered based on specified criteria for nocturnal sleep, and analysis frames ranging from 5 to 120 min were used to improve the diversity of the training sample and investigate the impact of analysis frame duration on feature significance and migraine prediction. Several models, including XGBoost (Extreme Gradient Boosting), HistGradientBoosting (Histogram-Based Gradient Boosting), Random Forest, SVM, and KNN, were trained on unbalanced data and using cost-sensitive learning with a 5:1 ratio. To evaluate the changes in features during pre-migraine nights and nights before migraine-free days, an analysis of variance (ANOVA) was performed. The results showed that the features of electrodermal activity, skin temperature, and accelerometer exhibited the highest F-statistic values and the most significant p-values in the 5 and 10 min frames, which makes them particularly useful for the early detection of migraines. The generalized prediction model using XGBoost and a 5 min analysis frame achieved 0.806 for accuracy, 0.638 for precision, 0.595 for recall, and 0.607 for F1-score. Despite identifying distinguishing features between pre-migraine and migraine-free nights, the performance of the current model suggests the need for further improvements for clinical application.

Monitoring insect activity at hive entrances is essential for advancing precision beekeeping practices by enabling non-invasive, real-time assessment of the colony’s health and early detection of potential threats. This study evaluates deep learning models for detecting worker bees, pollen-bearing bees, drones, and wasps, comparing different YOLO-based architectures optimized for real-time inference on an RTX 4080 Super and Jetson AGX Orin. A new publicly available dataset with diverse environmental conditions was used for training and validation. Performance comparisons showed that modified YOLOv8 models achieved a better precision–speed trade-off relative to other YOLO-based architectures, enabling efficient deployment on embedded platforms. Results indicate that model modifications enhance detection accuracy while reducing inference time, particularly for small object classes such as pollen. The study explores the impact of different annotation strategies on classification performance and tracking consistency. The findings demonstrate the feasibility of deploying AI-powered hive monitoring systems on embedded platforms, with potential applications in precision beekeeping and pollination surveillance.

Taking smartphone-made videos for photogrammetry is a convenient approach because of the easy image collection process for the object being reconstructed. However, the video may contain a lot of relatively similar frames. Additionally, frames may be of different quality. The primary source of quality variation in the same video is varying motion blur. Splitting the sequence of the frames into chunks and choosing the least motion-blurred frame in every chunk would reduce data redundancy and improve image data quality. Such reduction will lead to faster and more accurate reconstruction of the 3D objects. In this research, we investigated image quality evaluation in the case of human 3D head modeling. Suppose a head modeling workflow already uses a convolutional neural network for the head detection task in order to remove non-static background. In that case, features from the neural network may be reused for the quality evaluation of the same image. We proposed a motion blur evaluation method based on the LightGBM ranker model. The method was evaluated and compared with other blind image quality evaluation methods using videos of a mannequin head and real faces. Evaluation results show that the developed method in both cases outperformed sharpness-based, BRISQUE, NIQUE, and PIQUE methods in finding the least motion-blurred image.