• Home
• Preliminary Program

14:00   eHealth/Monitoring Chair: Joris Jaspers 14:00 15 mins Study Design for Detection of Circulatory Occlusion in Smart Watch Data Using Probabilistic Predictive Coding Roelof Hup, Xi Long, Reinder Haakma, Rik Vullings Abstract: Unwitnessed out-of-hospital cardiac arrest (OHCA) is a major health problem, with an incidence rate of 40.8 to 100.2 individuals per 100,000 population and survival rates between 4.6% and 18%. Early recognition and response by bystanders are crucial for increasing the chance of survival. However, 29.7% to 63.4% of the cases go unwitnessed, inevitably resulting in death. The BEating Cardiac Arrest (BECA) project aims to develop smartwatch technology to detect OHCA, even in unwitnessed cases [1]. Given the low incidence rate of OHCA, we collected a proxy dataset simulating OHCA conditions through the temporary circulatory occlusion of the arm, which mimics the lack of pulsatile behavior in the smartwatch’s photoplethysmogram (PPG) data. In this study, we asked 100 volunteers to wear a smartwatch for an hour, ending with a circulatory occlusion of the arm by inflating a blood pressure cuff to 200 mmHg for 20 seconds. To simulate OHCA as accurately as possible, volunteers were asked to sit still during this inflation, mimicking the typical unconsciousness that follows OHCA. The smartwatch captured both PPG and three-axes accelerometer data. To detect circulatory occlusion, we aim to use the Probabilistic Predictive Coding (PPC) anomaly detection method [2], an unsupervised machine learning pipeline that seeks to predict the distribution of future data given data from the past. At application time, the method can express the likelihood of actual data instances occurring according to this earlier predicted data distribution. By training the pipeline on data without circulatory occlusions, we expect that the model will be able to point out circulatory occlusions as unlikely to occur and therefore anomalous. One of the main data requirements for PPC is predictability, which artifact-free PPG signals generally comply with. However, during the daily use of smartwatches, motion artifacts are introduced in the PPG signal, which are inherently unpredictable. Therefore, we aim to adapt the PPC model so that the forecasting module considers not only PPG data, but also other types of information such as accelerometric data. The method will be tested using cross-validation and diagnostic performance will be expressed in the area under the Receiver Operating Characteristic curve (ROC-AUC). 14:15 15 mins Wearable Sensors and Machine Learning for Pelvic Rotation Monitoring Meixing Liao, Eva Libeert, Joeri van Cauwelaert, Simon Brumagne, Jean-Marie Aerts, Bart Vanrumste Abstract: A sedentary lifestyle is increasingly recognized as a major contributor to global health problems, including low back pain (LBP). Monitoring pelvic rotation, a key factor in posture and movement control, is essential to LBP prevention and management. However, current research in daily health monitoring has limited attention to pelvic rotation. To fill this gap, this study explores the feasibility of using IMUs to distinguish pelvic rotation from other common activities. To address the lack of data on pelvic rotation, we collected a dataset from 45 sedentary participants using the Xsens motion capture system (sampled at 60 Hz frequency), annotated with camera recordings. The dataset includes 8251 segments of sitting, standing, walking, sit/stand transition, and pelvic rotations in both directions. The time-series data were first converted into 0.5-second windows with 0.1-second strides. For each window, we calculated 16 hand-crafted features over 10 degrees of freedom (acceleration xyz, magnetic field xyz, and quaternions) to represent its patterns. The dataset is inherently imbalanced, with higher portions of sitting, standing, and walking, which imitate the daily scenarios—the duration of these activities is longer than pelvic rotations. To explore the feasibility of pelvic rotation recognition, six machine learning classifiers including Decision Tree, Random Forest, Support Vector Machine, XGBoost, Logistic Regression, and K-Nearest Neighbor were evaluated on the full sensor data. Our results showed that XGBoost consistently outperformed others, achieving a weighted overall F1 score of 0.916 in the 3-class configuration (anterior pelvic rotation, posterior pelvic rotation, and others). Another focus was to optimize sensor location. F1 scores indicated the sensor placed in the legging pocket on the sacrum as the optimal location for detecting pelvic rotation, compared with the other 17 sensor locations: the pelvic sensor achieved an F1 score of 0.692 for anterior pelvic rotation and 0.654 for posterior pelvic rotation, compared to scores of 0.775 and 0.738, respectively, when using the full sensor set. This research highlights the potential of IMUs for pelvic rotation detection, though it is limited by the relatively low variance in classes, underscoring the need for expanded data on diverse postures and activities in real-world scenarios. 14:30 15 mins Missing Data of the Philips HealthDot: insights into monitoring at home and level of physical activity Ilse Waanders, Harry Vaassen, Daan Lips, Annemieke Witteveen, Arlene John Abstract: Rationale: Continuous vital sign monitoring of patients after colorectal and pancreatic resections is crucial for assessing health status and enabling the early detection of potential deterioration. The Philips Healthdot, a wearable designed for post-discharge patient monitoring, uses accelerometer measurements to track heart rate (HR), respiration rate (RR), and patient activity, and transmits the data through the Long Range (LoRa) network. Data loss or periods of missing data can affect the effectiveness of this monitoring strategy compromizing patient safety. This study evaluates the extent of HR and RR data loss from the Philips HealthDot in both home and hospital settings. Methods: To quantify HR and RR data loss, thirty patients who underwent oncological colorectal or pancreatic resections were monitored with the HealthDot for 14 days. The percentage and duration of missing HR and RR data were retrospectively determined both inhospital and at-home. Data loss was classified as either non-transmission via the LoRa network or as missing values within transmitted data. The relationship between activity level and missing HR/RR values in transmitted data was analyzed using Students’ t-test. Results: From all thirty patients, 36.5% of RR and 43.8% of HR data was missing. Nine patients experienced more than four consecutive hours of missing data, with five of these patients missing data for more than nine consecutive hours. Over the entire monitoring period resulting in 116239 time datapoints, 30339 (26.1%) time datapoints were not transmitted via the LoRa network. In-hospital, the percentage of missing values in the transmitted data was 2.0%, whereas at home, it was 38.1%. Higher activity levels, such as walking, were significantly associated with the loss of HR and RR data in transmitted data (p<0.001). Conclusions: Missing HR and RR data is closely linked to LoRa network transmission limitations and patient activity levels. Implementing the HealthDot for at-home monitoring may reduce the capacity to detect and respond to patient deterioration, potentially compromising patient safety. Therefore, combining the healthdot with patient-reported measurements is advisable until improved remote patient monitoring solutions are available. 14:45 15 mins Optimizing telemonitoring protocols for early discharge after colorectal surgery; an evaluation of one year of telemonitoring. Ilse Waanders, Jian Bickes, Annemiek Kwast, Daan Lips, Annemieke Witteveen Abstract: Rationale: Early discharge after colorectal resections is facilitated through telemonitoring to detect symptoms of complications in early stage, since 2024 in the Medical Spectrum Twente (MST). Patients fill out a symptom related questionnaire with 10 questions in the morning and 3 in the afternoon and 3 vital sign measurements both times, which strains the patient. Frequent false positive alarms increase the alarm burden of the healthcare professional. This research aims to evaluate the current telemonitoring questions and measurements. The questions and their corresponding alarm thresholds are optimized to detect complications early while limiting the alarm burden. Methods: All patients who underwent an elective colorectal resection and were discharged early with home monitoring as a part of their regular care are included in this single-center retrospective cohort analysis. The total number of alarms per day of patients who were readmitted to the hospital was compared to the number of alarms per day of patients without readmission. The percentage of alarms based on an individual measurement was determined for the groups. Results: Preliminary data, from January until October of 2024, consist of 78 patients who were discharged early with telemonitoring in the MST after colorectal resection, of which 18 were readmitted. The number of alarms per patient per day was 3.63 ±1.85 alarms within readmitted patients and 2,56 ±1.33 in patients without readmission(p=0.033). Individual questions show that the level of appetite has a significant contribution of 66.7% in patients with readmission and 16.7% in patients without readmission(p=0.001). Temperature did not contribute towards determining complication status as there were zero alarms in the population. More comprehensive results are expected in January. Conclusion: Patients without readmission can be differentiated from those with readmission by questions and vital sign measurements from the telemonitoring protocol. Patients who were readmitted had significantly more alarms than those without readmission. Further analysis on the alarm thresholds is expected to lower the false positive rate, decreasing alarm fatigue in healthcare professionals. Some questions pose more contribution towards differentiation between groups whereas other questions show no contribution. Implicating that the telemonitoring protocol can be optimized to limit the burden on patients. 15:00 15 mins Predicting the length of geriatric rehabilitation stay following hip fracture surgery using machine learning: an exploratory study Sanne M. Krakers Abstract: Approximately 50% of patients aged ≥70 years are admitted to a geriatric rehabilitation department at a skilled nursing home following hip fracture surgery1. Given the wide variation in the length of geriatric rehabilitation stay (LOGRS), predicting the LOGRS would help manage patients’ recovery expectations and create appropriate therapy schedules2. This study aims to predict the LOGRS for older patients following hip fracture surgery using data collected upon the first week of geriatric rehabilitation. Proposed methods: Continuous physical activity data from at least 101 patients aged ≥70 years will be monitored using the MOX accelerometer during their first week of geriatric rehabilitation following hip fracture surgery. Statistical, amplitude, and morphological features of the physical activity data will be calculated to predict the LOGRS. Additionally, patient characteristics such as age, surgery type, mobility scores, and cognition scores will be collected. The LOGRS will be divided into three categories based on previous research (e.g., ≤ 28 days, 29-42 days and ≥ 43 days)2. Physical activity features and patient characteristics from 80% of patients in each category will be randomly assigned to two separate (one physical activity and one patient characteristics) training sets, with 20% allocated to two separate test sets. Principal component analysis will be applied separately to both training sets to identify the features with the highest explained variance, before merging them into a single training set. Multiple machine learning (ML) models will be trained using the MATLAB2023b Classification Learner App, with the model showing the highest accuracy selected for further analysis. Preliminary Results: Among the ML models tested in this preliminary phase, the Bilayered Neural Network emerged as one of the ML models that yielded promising results, achieving a classification accuracy of 75.0%, a recall of 73.9% and a precision of 77.2% on the test set. Conclusions: This study will explore the feasibility of predicting the LOGRS. Accurately predicting LOGRS can assist healthcare professionals in managing patient expectations and create appropriate therapy schedules that promote faster recovery. This may help address potential future capacity issues by improving patient flow, allowing for the treatment of an increasing number of hip fracture patients. 15:15 15 mins Neural interfaces for bioelectronic medicine Geert Langereis, Vojkan Mihajlovic, Nicoló Rossetti, Philipp Schnepel, Konstantinos Petkos Abstract: The regulation of most physiological processes in the human body relies on a homeostatic equilibrium. This is achieved through a biological control loop involving physiological receptors (sensors) and molecular actuators, connected from the central nervous system to local tissue and organs. The connection from the brain to the organs, and vice versa, is done through the peripheral nervous- and vascular systems. However, tissue damage, disease and aging can disrupt these control loops, leading to impaired homeostasis. Bioelectronic devices offer the opportunity to interface artificial technology with biological systems, enabling the measurement, stimulation and modulation of specific processes using sensors and actuators . The assumption is that bioelectronic medicine in the form of neuromodulation devices can restore homeostasis, meaning, bringing back a worry-free life for chronically ill people and prevent the evolvement of comorbidities and development of other chronic conditions. To increase the efficacy of bioelectronic medicine delivered via neuromodulation devices, substantial improvements are required in terms of selective activation of fibers innervating target vs non-target end organs. Functionally selective activation of vagus nerve fibers could lead to precise control of the response of e.g., the human inflammatory system that could result in benefits for chronic inflammatory diseases while reducing side-effects such as neck muscle contraction and voice alterations. However, the technical needs for chronic neuromodulation devices require improvements in short-term adaptivity and functional selectivity. This means that peripheral nerves must be stimulated at sufficient spatial and temporal resolution, and control loops must be closed using real time nerve read-out. Chronic use requires miniaturized, low-power implants that have a functional life of many years. We have developed a stimulation method that offers improved functional selectivity over the state-of-the-art. A dedicated ASIC was developed to enable advanced high-frequency stimulation methods achieving low power consumption , and a benchtop system was built achieving real-time readout of compound action potentials and facilitation of close-loop applications. Additionally, the system features advanced, fast settling stimulation artefact suppression.

end %-->