Personal features can significantly enhance machine learning models by improving both personalization and prediction performance in health-related applications. In this study, we created a new subset of DiaData, an integrated continuous glucose monitoring (CGM) dataset of 2510 subjects with Type 1 Diabetes (T1D), by extracting age, sex, duration of diabetes, HbA1c, race, height, and weight. Our objective was to assess the influence of these additional features on the performance of hypoglycemia classification models. T1D is an autoimmune disorder in which the pancreas cannot produce sufficient insulin, and thereby affected individuals depend on external insulin injections. However, a major side effect of insulin therapy is hypoglycemia, defined as blood glucose levels below$\mathbf{7 0 ~ m g} \boldsymbol{/} \mathbf{d L}$. Hypoglycemia can become life-threatening if not detected on time and if appropriate preventive measures are not implemented. This risk is particularly concerning in asymptomatic episodes. To address this challenge, we propose Fully Convolutional Network (FCN) and XGBoost-based hypoglycemia classification models that incorporate personal features. Specifically, our contributions include: 1) Enriching DiaData with additional personal features by extracting a new subset comprising 8 databases and 1651 subjects. 2) Performing a correlation analysis between numerical personal features and hypoglycemic CGM glucose ranges. 3) Training both deep learning and machine learning models to explore the impact of personal features on prediction performance. 4) Evaluating the contribution of personal features to prediction outcomes in XGBoost models. Our findings show that while the HbA1c score and the age group have a moderate impact, other features such as height, weight, and the duration of diabetes have minimal influence. Although personal features were included, their addition did not lead to a significant improvement in the models' predictive performance.

There is growing interest in applying statistical and deep learning-based imputation techniques to address missing values in physiological time series data. However, traditional evaluation metrics like RMSE often fail to capture the accuracy of imputed values, particularly for heart rate (HR) signals. Additional metrics such as MAE, MAPE, Cohen's Distance Test (CDT), and Jensen-Shannon Distance (JSD) can yield inconsistent evaluations, complicating the choice of optimal imputation methods for downstream prediction tasks. To address this, we propose FRAM-SHAP, a novel framework that combines multiple predictive and statistical distance metrics into a weighted metric. Weights are derived using XGBoost optimized with Optuna and interpreted via L1-normalized SHAP values, based on each metric's ability to determine whether imputed values fall within the same distributional interval as the original data. Further, robustness is evaluated using 10 bpm and 20 bpm HR based distributional intervals across two datasets, 'D1NAMO' and 'BIG IDEAs Lab Glycemic Variability and Wearable Device Data', capturing different degrees of HR variability. Thus, FRAMSHAP provides a flexible, adaptable framework for evaluating imputation quality in application specific contexts.

The rapid evolution and continuously growing complexity of cyber-physical systems (CPS), ranging from Industry 4.0 production plants to ship drivetrains and building monitoring, have led to the abundant generation of heterogeneous, poorly-structured, and not standardized data. This situation is further aggravated by retrofitting legacy systems with new sensors for the purpose of data-driven monitoring. In this paper, we introduce Cyber-Physical System Watch (CPSWatch), a lightweight framework that aims to support the monitoring of CPS including the possibility for diagnosis, encompassing a high-level ontology, two sample datasets of different complexity as well as a use case scenario on how it can be applied. Our proposed ontology provides a unified framework for describing data across different CPS applications and aligns with OPC UA, ensuring its general applicability in modern industrial settings. CPSWatch is evaluated in terms of standard ontology evaluation measures, within the scope of condition monitoring of an automation system in the maritime domain and a benchmark in the field of process engineering.

Dataset description VitaStress is a multimodal wearable dataset for automated stress recognition and affective computing research. The dataset was collected in a controlled laboratory setting to support the development, evaluation, and comparison of machine learning models for detecting stress from physiological and motion signals. It contains wrist-worn vital-sign and kinematic data recorded during four experimentally defined affective and activity states: neutral baseline, physical activity, cognitive stress, and socio-evaluative stress. The dataset includes processed physiological parameters and raw sensor signals such as heart rate-related measures, RR intervals, respiration-related information, temperature, electrodermal activity, photoplethysmography, and three-axis accelerometer data. Each recording is accompanied by detailed annotations describing the timing of experimental phases, stimuli, body positions, locations, rest periods, and relevant contextual information. In addition, psychological self-reports are provided after each condition to document participants’ subjective affective states in terms of arousal, valence, and dominance. VitaStress was designed to address the limited availability of publicly accessible, well-annotated, and reusable stress-recognition datasets. The dataset follows a structured experimental protocol inspired by reproducible and reusable affective computing principles, enabling detailed reconstruction of each experiment and supporting comparability across studies. It can be used for binary stress classification, multiclass affective-state recognition, subject-independent evaluation, feature engineering, multimodal sensor fusion, missing-data analysis, and benchmarking of wearable stress-detection models. By combining physiological signals, motion data, precise annotations, and self-reported affective assessments, VitaStress provides a reusable resource for research in human-computer interaction, wearable health technology, affective computing, physiological time-series analysis, and human-centered machine learning. The VitaStress dataset is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

Abstract Forecasting the behaviour of industrial robots, power grids or pandemics under changing external inputs requires accurate dynamical models that can adapt to varying signals and capture long-term effects such as delays or memory. While recent neural approaches address some of these challenges individually, their reliance on computationally intensive solvers and their black-box nature limit their practical utility. In this work, we propose Laplace-Net, a decoupled, solver-free neural framework for learning forced and delay-aware dynamical systems. It uses the Laplace transform to (i) bypass computationally intensive solvers, (ii) enable the learning of delays and memory effects and (iii) decompose each system into interpretable control-theoretic components. Laplace-Net also enhances transferability, as its modular structure allows for targeted re-training of individual components to new system setups or environments. Experimental results on eight benchmark datasets–including linear, nonlinear and delayed systems–demonstrate the method’s improved accuracy and robustness compared to state-of-the-art approaches, particularly in handling complex and previously unseen inputs.

The increasing operational demands in maritime contexts, particularly during time-sensitive missions like search and rescue, necessitate reliable, intelligent support systems. These systems depend on semantically structured and interoperable models to integrate and interpret complex sensor data as well as facilitate informed decision-making. We introduce MontoFlow, a semantic integration framework that combines dynamic data access with domain-specific knowledge representation. It links static properties with dynamic sensory measurements, forming the foundation for advanced maritime diagnostics. At its core, MontoFlow incorporates the SHIP Ontology, a maritime-focused SSN/SOSA extension that provides a comprehensive semantic model describing onboard sensors, vessel components, and their observations. We illustrate the practical relevance and rationale behind the development of MontoFlow through real-world examples, with emphasis on ship maintenance and onboard anomaly detection. The SHIP Ontology is thoroughly evaluated based on domain coverage and a use case in the maritime context, demonstrating both high quality and practical applicability. This work presents a reusable and extensible resource for semantically enriching maritime sensory data, supporting advanced analytics and dynamic data monitoring. Ontology: https://burbachs.github.io/ShipSensoryOntology/SHIP.owl GitHub: https://github.com/BurbachS/ShipSensoryOntology License: CC BY-NC-SA 4.0 DOI: 10.5281/zenodo.15390282

Chronic diseases, especially diabetes, are burdens for the patient since lifelong management is required, and comorbidities can occur as a consequence of insufficient prevention. Hypoglycemia, a medical condition encountered by diabetic individuals, can result in severe symptoms if untreated, necessitating prompt preventive actions. Continuous health monitoring based on data collected with wearables can enable the early prediction of extreme blood glucose states. However, integrating and using data acquired from various sensors is challenging, especially when it comes to maintaining the quality and quantity of data due to inherent noise and missing values. To this end, the review discusses dataset constraints and highlights the temporal behavior of prominent features in predicting hypoglycemia. It outlines a framework of preprocessing techniques that could be adopted to mitigate dataset limitations. A thorough analysis of the imputation procedures employed in the reviewed studies is conducted. In addition, machine learning imputation techniques applied in other healthcare domains are studied to investigate if they could be adopted to close the longer gaps of missing values in the datasets involved in the prediction of hypoglycemia. Based on a comprehensive evaluation of imputation techniques, a paradigm, Impute-Paradigm, is proposed and validated through a case study, enabling imputation tailored to variable duration time gaps. After analysing the reviewed studies, we recommend studying the rate of change of individual features and conclude that different time gaps of separate features should be treated with multiple imputation techniques. • Analysis of preprocessing and imputation techniques for wearable sensor data. • Analysis of temporal patterns in blood glucose, heart rate, and accelerometer data. • Comprehensive quantitative evaluation of various imputation techniques. • Recommendation of imputation techniques based on time gap size using Impute-Paradigm.

Medical data analysis provides valuable insights into patient contexts, supporting personalized treatments and preventive strategies. Reliable analysis requires large volumes of high-quality data, as outliers can distort results and missing values lead to information loss. Notably, for Type 1 Diabetes (T1D), data analysis explores relationships between demographics, sensor data, and behavior. To address limited data volume, DiaData-an integration of 15 separate datasets containing glucose values from 2510 subjects with T1D-was previously introduced. This study improves the quality of DiaData by identifying outliers and imputing missing values. In particular, we make the following contributions: 1) Sensor errors are determined with the interquartile range (IQR) approach and replaced with missing values. Outlier removal leads to less bias toward misleading values. 2) Gaps are classified by length to impute small gaps ($\le$ 25min) with linear interpolation and larger gaps ($\ge$30 and $\le$ 120min) with Stineman interpolation. A visual comparison shows that Stineman interpolation provides more realistic glucose estimates than linear interpolation for larger gaps. 3) After data cleaning, the correlation between glucose and heart rate is analyzed, reporting a moderate relation between 15 and 60 minutes before hypoglycemia ($\le$ 70mg/dL). 4) Finally, a benchmark for hypoglycemia classification is provided with a state-of-the-art Fully Convolutional Network (FCN). The model is trained with the main database and subdatabase II of DiaData to classify hypoglycemia onset up to 2 hours in advance. Training with more data improves performance by 3% while using quality-refined data yields a 4% gain compared to raw data.

This paper addresses the task of conversational question answering (ConvQA) over knowledge graphs (KGs). The majority of existing ConvQA methods rely on full supervision signals with a strict assumption of the availability of gold logical forms of queries to extract answers from the KG. However, creating such a gold logical form is not viable for each potential question in a real-world scenario. Hence, in the case of missing gold logical forms, the existing information retrieval-based approaches use weak supervision via heuristics or reinforcement learning, formulating ConvQA as a KG path ranking problem. Despite missing gold logical forms, an abundance of conversational contexts, such as entire dialog history with fluent responses and domain information, can be incorporated to effectively reach the correct KG path. This work proposes a contrastive representation learning-based approach to rank KG paths effectively. Our approach solves two key challenges. Firstly, it allows weak supervision-based learning that omits the necessity of gold annotations. Second, it incorporates the conversational context (entire dialog history and domain information) to jointly learn its homogeneous representation with KG paths to improve contrastive representations for effective path ranking. We evaluate our approach on standard datasets for ConvQA, on which it significantly outperforms existing baselines on all domains and overall. Specifically, in some cases, the Mean Reciprocal Rank (MRR) and [email protected] ranking metrics improve by absolute 10 and 18 points, respectively, compared to the state-of-the-art performance.

Endri Kacupaj, Joan Plepi, Kuldeep Singh, Harsh Thakkar, Jens Lehmann, Maria Maleshkova. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 2021.

In this paper, we introduce the MLM (Multiple Languages and Modalities) dataset - a new resource to train and evaluate multitask systems on samples in multiple modalities and three languages. The generation process and inclusion of semantic data provide a resource that further tests the ability for multitask systems to learn relationships between entities. The dataset is designed for researchers and developers who build applications that perform multiple tasks on data encountered on the web and in digital archives. A second version of MLM provides a geo-representative subset of the data with weighted samples for countries of the European Union. We demonstrate the value of the resource in developing novel applications in the digital humanities with a motivating use case and specify a benchmark set of tasks to retrieve modalities and locate entities in the dataset. Evaluation of baseline multitask and single task systems on the full and geo-representative versions of MLM demonstrate the challenges of generalising on diverse data. In addition to the digital humanities, we expect the resource to contribute to research in multimodal representation learning, location estimation, and scene understanding.

Question Answering (QA) systems over Knowledge Graphs (KGs) aim to provide a concise answer to a given natural language question. Despite the significant evolution of QA methods over the past years, there are still some core lines of work, which are lagging behind. This is especially true for methods and datasets that support the verbalization of answers in natural language. Specifically, to the best of our knowledge, none of the existing Question Answering datasets provide any verbalization data for the question-query pairs. Hence, we aim to fill this gap by providing the first QA dataset VQuAnDa that includes the verbalization of each answer. We base VQuAnDa on a commonly used large-scale QA dataset – LC-QuAD, in order to support compatibility and continuity of previous work. We complement the dataset with baseline scores for measuring future training and evaluation work, by using a set of standard sequence to sequence models and sharing the results of the experiments. This resource empowers researchers to train and evaluate a variety of models to generate answer verbalizations.

Information on the Web is heterogeneous and available in constantly increasing quantities. Consequently, there are numerous, partly redundant data analytics services, each optimized for data with certain characteristics. Often, analytics tasks require multiple services to be pipelined to find a solution, where combinations of exchangeable services for single steps might outperform one-service-predictions. This work proposes a Multi-Agent System (MAS) perception of prior setting, where decentralized agents are considered to manage services, having to coordinate their decisions to find a consensus. We, first, propose a supervised method for service accuracy estimation and, therefore, exploit locality-sensitive features of training data. Given a committee of services managed by agents, we develop coordination strategies to handle conflicting confidences and reduce erroneous predictions due to service correlation. We evaluate our approach with Named Entity Recognition (NER)- and Named Entity Disambiguation (NED) services on text corpora with heterogeneous characteristics (i.e. news articles and tweets). Our empirical results improve the out-of-the-box performance of the original services.

The vision of the Internet of Things (IoT) promises the capability of connecting billions of devices, resources and things together. In the realisation of this vision, we are currently neglecting the interoperability between devices that is caused by a heterogeneous landscape of things and which leads to the proliferation of isolated islands of custom IoT solutions. A first step towards enabling some interoperability is to connect things to the Web and to use the Web stack, thereby conceiving the socalled Web of Things (WoT). However, even when a homogeneous access is reached through Web protocols, a common understanding is still missing. In addition, decentralised applications, advocated by the IoT vision, and a-priori unknown requirements of specific integration scenarios demand new concepts for the adaptation of things at runtime. Our work focuses on two main aspects: overcoming not only data but also device and interface heterogeneity, and enabling adaptable and scalable decentralised WoT applications. To this end we present an approach for realising decentralised WoT applications based on three main building blocks: 1) semantics of the devices' capabilities and interfaces, 2) rules to enable embedding controller logic within device's interfaces for supporting a decentralised applications, and 3) support for reconfiguring the controller logic at runtime for customising and adapting the application. We show how our approach can be applied by introducing a reference architecture, provide a thorough evaluation in terms of a proof-of-concept implementation of an example use case, and performance tests.