Business processes in complex and dynamic environments, such as hospitals, are subject to constant changes. This could arise from adaptation to unforeseen events, creating uncertainty and inefficiency in clinical pathways. Traditionally grounded in sub-symbolic AI, Process Mining provides insights into process behavior through actions such as concept drift analysis, including detection, characterization, and explainability. However, most existing studies focus on drift detection, making it challenging to identify the root causes of process deviations. The literature indicates that this limitation largely stems from the difficulty of integrating domain-specific knowledge into process mining. This article targets this issue and presents the Minuscule Movement of business Processes (MMP) approach to diagnose drifts and deviations in patients’ pathways through two main steps. First, it defines a meta-model to embed domain knowledge such as potential causes of drifts into process discovery analyses and to generate artificial traces that simulate process deviations. In the second step, these traces are assessed using the proposed ProDIST algorithm to find the process most similar to the discovered workflow. The identified process is then used to diagnose and determine the root causes of the drift. MMP’s practical applicability is assessed through a real-life experiment in healthcare. Accordingly, MMP successfully detected process drifts in the healthcare case study, achieving consistent accuracy and providing interpretable insights that enabled domain experts to identify potential causes of process drifts and deviations.

Video-based fall detection plays a crucial role in telemonitoring as a key component in ensuring timely intervention and safety for older adults who live alone. Despite promising advances, most vision-based fall detection methods remain insufficiently robust and fail to generalize effectively to real-world scenarios. In this paper, we present a comprehensive experimental framework designed to rigorously and generically assess the robustness of fall detection approaches. This framework encompasses three evaluation settings: intra-subject, intersubject, and cross-dataset evaluation. It is validated using a custom architecture that combines a convolutional neural network (CNN) with a bidirectional LSTM (BiLSTM), evaluated on three public datasets: URF, Le2i-FD, and MCFD. Experimental results demonstrate strong generalization, with recall exceeding 80% in cross-dataset settings. These findings underscore the importance of diverse evaluation strategies in developing reliable fall detection systems.

Objective: This study aims to create a complete Medication Error (ME) dataset. This will help to address the challenge of limited access to real-world data for developing machine learning models in healthcare applications. Methods: We use transformer-based models (GPT-4, LLAMA3, and Mistral) to create our synthetic dataset in French. These models generate a diverse range of descriptions that capture the variability of ME types. We assess the effectiveness of our synthetic dataset through expert evaluations by healthcare professionals and an AI-driven analysis, to test its realism and its utility in training machine learning models for ME classification. Results: The synthetic dataset demonstrates high accuracy and realism in representing diverse ME scenarios. Expert evaluation confirms that the dataset is similar to real-world ME data. The AI-driven evaluation also shows that models trained on synthetic data achieved robust classification performance, validating the dataset's utility for the development of effective ME classification tools. Conclusion: The proposed approach demonstrates the potential of large language models to generate realistic synthetic ME reports in French. Out of 200 evaluated reports, 70% of zero-shot outputs were deemed below expectations, while 80% of one-shot and few-shot outputs were considered valid or valid with minor revisions by clinical experts. Furthermore, classifiers trained on 800 synthetic reports attained an F1-score of up to 0.78 when tested on real data. These results confirm that synthetic data can effectively support AI-driven ME analysis in contexts where real-world data is limited or unavailable.

Human Activity Recognition (HAR) plays a critical role in healthcare monitoring and smart home systems, enabling tracking of patient movements, fall detection, and daily activity monitoring. However, HAR faces challenges due to the scarcity of diverse, large-scale datasets and the absence of sufficient abnormal activity samples necessary for detecting rare but critical health events. This paper addresses these challenges through advanced synthetic data generation and state-of-the-art classification techniques. We introduce a Generative Adversarial Network (GAN) for time-series data to generate synthetic samples, significantly expanding the WISDM dataset and incorporating an ‘abnormal’ activity class to enhance dataset diversity and real-world applicability. The fidelity of the synthetic data is rigorously evaluated using Dynamic Time Warping (DTW), achieving an average distance of 56.1, demonstrating strong alignment with real data distributions. For classification, we leverage transformer-based models, which have shown superior performance over traditional HAR methods such as CNNs and LSTMs. Our approach achieves 91.4

In this paper, we present a new application of Discrete Event Simulation (DES) as a Digital Twin (DT) for real-time monitoring of patient pathways (PPs) in a hospital. This application, which we call HospiT'Win, enables hospital managers and decision-makers to monitor and manage PPs in real time, supporting informed decision-making. Our approach relies on real-time synchronization between the DT and the PPs at each detected event in the hospital. This approach requires that the initial state of the simulation reflects the current state of the PPs in the hospital. For consistency, in the remainder of this paper, we refer to our approach as the DT of PPs. In this paper, we propose a real-time synchronization mechanism, based on DES, to synchronize the DT of PPs in a hospital with real PPs. This synchronization mechanism was tested using an emulator that mimics the stochastic processes of an outpatient clinic. The use of the proposed synchronization mechanism resulted in a dynamic DT that can track real PPs in hospitals in real time for each detected event. Moreover, this synchronization mechanism allowed for deeper analysis of the behavior of the emulated clinic through the DT and facilitated the detection of unexpected events. Quantitative evaluation showed an average synchronization latency of 0.73 s and an accuracy of 97.7

Automated process discovery as one of the paradigms of process mining has attracted both industries and academic researchers. These methods offer visibility and comprehension out of complex and unstructured event logs. Over the past decade, the classic heuristic miner and applied heuristic-based process discovery algorithms showed promising results in revealing the hidden process patterns in information systems. One of the challenges related to such algorithms is the arbitrary selection of recorded behaviors in an event log. The offered filtering thresholds are manually adjustable, which could lead to the extraction of a non-optimal process model. This is also visible in commercial process mining solutions. Recently, the first version of the stable heuristic miner algorithm targeted this issue by evaluating the statistical stability of an event log. However, the previous version was limited to evaluating only activities’ behaviors. In this article, we’ll be evaluating the statistical stability of both activities and edges of a graph, which could be discovered from an event log. As a contribution, the stable heuristic miner 2 is introduced. Consequently, the definition of the descriptive reference process model has improved. The novel algorithm is evaluated by using two real-world event logs. These event logs are the familiar Sepsis data set and the urology department patients’ pathways event log, which is recorded by monitoring the interpreted location data of patients on hospital premises and is shared with the scientific community in this article.

Human Activity Recognition (HAR) is essential for applications like health monitoring and fall detection using mobile sensors. However, obtaining large datasets necessary for training HAR systems is prohibitively expensive. Generative Adversarial Networks (GANs) have been proposed to generate synthetic HAR data, simplifying and enhancing the development of HAR systems. While these methods have improved accuracy, existing GAN-based approaches struggle with generating clear abnormal patterns, weakening anomaly detection capabilities. Available HAR data predominantly focuses on normal activities and does not target anomalies, making it challenging to detect anomalous situations. To address this, we propose a novel controllable GAN that generates realistic HAR data as well as distinct abnormal classes. This advancement enhances anomaly detection accuracy through more standardized activity recognition and quantification by HAR systems. We evaluate our approach on the WISDM dataset, demonstrating significant improvements in the balance and quality of synthetic data, leading to better performance in anomaly detection.

The growing elderly population in developed countries highlights the critical need for preventing frailty, which poses significant challenges to health systems due to increased risks of severe health issues. This paper introduces Senselife, a framework that provides explainable service recommendations specifically tailored for frailty prevention. It begins by outlining the medical context and challenges associated with aging, followed by an overview of existing recommender systems with similar objectives. We detail the integration of three key resources-ROR, RNA, and Data Laregion-within the Senselife framework to represent service supply. The paper explains how service supply is structured and the transformation of available data for use within our recommender engine. We introduce the concept of operational activities derived from the ROR and leverage the capabilities of LLMs to incorporate RNA data into Senselife. Additionally, we illustrate how these services are ultimately compiled into recommended service packages. Finally, the paper concludes by summarizing key findings and suggesting potential directions for future research.

As the global elderly population continues to grow, social isolation emerges as a critical factor contributing to the incidence of frailty. This paper explores the integration of social interaction functionalities within the Senselife framework, a service recommendation platform designed for frailty prevention in older adults. We propose enhancements to Senselife that facilitate community-building and social engagement through technology-driven interactions. By leveraging user-centered design, the paper discusses how enhanced social features can significantly improve the efficacy of our frailty prevention strategies, offering a holistic approach to elderly care. Our methodology includes the development of social interaction modules that encourage active participation and connectivity among elderly users, ultimately aiming to enhance their quality of life and reduce the risks associated with social isolation.

The aging global population presents unique challenges, particularly in managing frailty-a condition defined by declines in physical, cognitive, and social capacities. This paper introduces Senselife, a recommender system tailored for frailty management in elderly individuals. Senselife leverages hypergraph-based knowledge models to intelligently recommend personalized services aimed at mitigating frailty and enhancing life quality. Our methodology integrates diverse data types through Heterogeneous Information Networks (HINs), allowing for nuanced userservice interactions that significantly improve recommendation accuracy and relevance. This paper details the development of these models, emphasizing the transition from conventional data handling to advanced, knowledge-driven approaches that consider both user and service complexities. By incorporating these sophisticated models, Senselife aims to provide a scalable solution for frailty prevention, offering a significant contribution to personalized elderly care.

Hospital bed requirement prediction is essential for efficient healthcare resource management and delivering high-quality patient care. Unscheduled admissions significantly impact waiting times and complicate bed management strategies. In this paper, we present different a Multi-Task Learning (MTL) models to predict bed requirements for various Inpatient Departments (IPD) simultaneously. Specifically, we target prediction of daily bed requirements for patients admitted through the Emergency Department (ED). The proposed MTL approach, demonstrates significant advantages in predictive accuracy. Among implemented models, the MTL with XGBoost outperformed than other models in 11 tasks out of 12 with sMAPE, MAE and MSE values ranging from 0.185 to 0.974 and MTL with PyTorch model achieved low sMAPE value of 0.620 with 1 task. A key contribution of this paper is our simultaneous predictive approach, that considers the interaction between departments and leverages the bed occupancy data from the previous day (lag 1) to make a prediction. This methodology aims to reduce waiting time, optimize resource utilization and enhance patient flow within the hospital by following unscheduled admissions. This study demonstrates the potential of AI-driven predictive analytics in hospital resource management and emphasize the significance of MTL methodologies in predicting bed requirements.

Because of their potentially serious consequences, Medication errors (ME) represent a major challenge for health-care facilities. To manage these errors and minimize their seriousness, healthcare professionals follow a collaborative management process that relies on reporting and then analyzing the reports filled in by them via various reporting tools, both digital and paper-based. These tools can be customized requiring specific information to be entered in multiple forms with commonly the presence of textual descriptions to fill in a free-text field. Therefore, text analysis is crucial for thoroughly understanding and effectively analyzing medication errors. Given the large volume of reports to be quickly processed, it is essential to help healthcare professionals prioritize which ME to analyze. In this context, we propose, in this work, processing ME reports with natural language processing tasks using Transformer models such as GPT. In this study, we present the extraction of key information from the reports to help structure textual descriptions of ME with the GPT-4 transformer model. The results obtained show the potential of this model to extract relevant information from ME descriptions in French language without any deep fine-tuning,

Efficient patient transport is crucial for the smooth operation of hospitals, as it directly impacts the timely delivery of medical care and the overall patient experience. The complexity of managing transport logistics, especially with limited resources, necessitates advanced optimization techniques. In this paper, we present an optimization model designed to enhance the scheduling of patient transportation missions and improve the organization of transport activities. Our primary objective is to minimize delays and reduce the number of unaccomplished missions, even when transporter availability is constrained. Using historical data from a hospital in France and the CPLEX solver, we compare our model's performance with existing models in the literature. The results demonstrate that our model significantly improves efficiency and reliability, ensuring timely patient transport and better resource utilization.

Optical Character Recognition (OCR) is extremely useful in various sectors for exploring massive archived data. This technology enables the digitization of printed and handwritten texts that are frequently present in the medical field. For instance, medication error (ME) reports were previously and still in some healthcare facilities written manually, this has led to the accumulation of numerous handwritten data that are unfortunately challenging to exploit. Their digitization through OCR allows extracting important data from these documents and using them to populate the database to implement future analysis techniques to optimize the medication error management process. This paper presents a transformer-based handwritten recognition architecture that employs the Transformer-Based Optical Character Recognition (TrOCR) model combined with image segmentation techniques. Although the TrOCR model provided by Microsoft performs reasonably well in handwritten recognition, it is limited to English text because its pretrained version was trained exclusively on English samples. This limitation is problematic for us, as our task involves digitizing French medication dictation errors. Additionally, its limitation to processing single-line text images impairs its ability to recognize paragraphs. To address these limitations, we will fine-tune the model on French handwritten data and integrate a single-line level segmentation technique, thereby overcoming these constraints. Therefore, the preliminary results from implementing our proposed architecture are promising for the digitization of medication error reports.

Predicting patient waiting times in public emergency department rooms (EDs) has relied on inaccurate rolling average or median estimators. This inefficiency negatively affects EDs resources and staff management and causes patient dissatisfaction and adverse outcomes. This paper proposes a data science-oriented method to analyze real retrospective data. Using different error metrics, we applied various Machine Learning (ML) and Deep learning (DL) techniques to predict patient waiting times, including RF, Lasso, Huber regressor, SVR, and DNN. We examined data on 88,166 patients' arrivals at the ED of the Intercommunal Hospital Center of Castres-Mazamet (CHIC). The results show that the DNN algorithm has the best predictive capability among other models. By precise and real-time prediction of patient waiting times, EDs can optimize their activities and improve the quality of services offered to patients.

The Medication Use Process in hospitals is one of the crucial processes of the care pathways. It is defined as a highly intricate and collaborative process aiming to deliver effective and safe patient care. Indeed, it involves different stakeholders working together to provide the best treatment to the patient and manage several health conditions. This complexity can give rise to medication errors, leading to serious patient health-related consequences. Being fully aware of the significance and complexity of this process, in this work, we propose a digital collaborative framework to manage medication errors. The main goals of this framework are to assist healthcare organizations in their risk management sessions, maximize the effectiveness of the care provided, and improve patient safety and quality of care.