13:30
Behavior & Monitoring
Chair: Joris Jaspers
13:30
15 mins
|
Bathroom Activity Monitoring Via Wearable Accelerometer Device
Yiyuan Zhang, Ine D'Haeseleer, Vero Vanden Abeele, Bart Vanrumste
Abstract: Background: Activities of daily living (ADLs) can be monitored to determine the physical health status of older adults. In particular, bathroom activities seem promising, and may be detected via the acoustic system, camera, or internet-of-things (IoTs) sensors, with or without activity monitors (accelerometer, gyroscope, and magnetometer).
The aim of the research: In this research, the aim is to test whether it is feasible to monitor bathroom activities only using one tri-axis accelerometer.
Methodology: The activities of interest in this research are the common bathroom activities: dressing, undressing, washing hands, washing the face, brushing teeth and toileting. During the experiment, participants wore one Empatica E4 (including a tri-axis accelerometer) on their dominant wrist. Additionally, they were asked to annotate the starting and stopping time of each activity, via an APP. Each participant performed the experiment for two consecutive weeks. The experiment took place in each participant’s own home. Data from 9 participants (aged over 65 years old) were analyzed. The RBF kernel support vector machine (SVM) multi-class classification model was trained using the one-against-one method, with 5-fold cross-validation. The performance of the classification model was trained intra- and inter- subject.
Results: 908 activities were successfully recorded: Dressing, undressing, washing face, washing hands, brushing teeth and toileting with fraction of instances of 9.5%, 9.7%, 6.4%, 25.7%, 10.8% and 37.8%, respectively. In total, 54 features were extracted in the time and frequency domain. In the intra-subject, the ratio of training and the testing dataset is 7:3. The average f1-score of each activity over all the subjects ranged from 36.51% (dressing) to 85.26% (toileting), for the test data. In the inter-subject case, the data of 3 subjects were assigned to be the testing data, others as the train data. The f1-score result ranged from 80.4% (washing face) to 89.5% (toileting) for the training data and from 29.4% (dressing) to 55.6% (brushing teeth) for the test data, which illustrates the subject variability of physical activities, especially for non-periodic activities.
Conclusion & discussion: The performance of the classification model varies among activities and participants. In future work, we will analyze the misclassification error of each activity and check whether it is a bias or variance problem. Additionally, considering the complexity of some of the activities (e.g. toileting includes actions such as washing hand, dressing, flushing, undressing, etc.) another important step is to firstly classify the actions of the activities.
Acknowledgement: This research study was possible with the support of KIC EIT Health funding for GRaCE-AGE.
|
13:45
15 mins
|
Carewear: Integrating Wearable Technology in Mental Healthcare
Glen Debard, Romy Sels, Marc Mertens, Nele De Witte, Tom Van Daele, Bert Bonroy
Abstract: Wearables can collect physiological data continuously. This can give information both on vulnerability factors and the process of recovery in mental disorders. However, current technological applications in this field are limited. Burnout and depression are highly prevalent mental disorders that have a large impact on psycho-emotional wellbeing and are associated with substantial societal and economical costs. The Carewear project project (Vlaio Tetra HBC.2016.0099) aims to enrich current employee assistance programs and the treatment of depression with the implementation of wearable technology. We have developed an online software platform and accompanying clinical guidelines that allow healthcare professionals to use physiological data as a useful addition to their current practices. Two use cases are defined to investigate the added value of this implementation of wearable technology to help prevent burnout and treat depression. Clients are asked to wear a wristband that registers several physiological parameters. These are blood volume pulse, skin conductance, skin temperature and movement. Algorithms using artificial intelligence are being developed to translate these physiological parameters into data that can be used to assist in the assessment of the mental health of the subject. The data consists of the heart rate variability, stress peaks using skin conductance, skin temperature, and heart rate, and physical activity. This physiological data can be inspected and completed on the online Carewear dashboard and consequently discussed in regular consults. Altogether, the Carewear project aims to encourage the use of wearable technology in mental healthcare by providing a user-friendly platform and clinical guidelines. Both are tailored for elevated stress and depressive symptoms, which makes physiological data accessible and comprehensible for both healthcare professionals and clients.
Wearable technology has a large potential in the field of mental healthcare but there are still some challenges in the practical implementation. The current wrist-worn wearables still have difficulties to produce valid signals which leads to more postprocessing which decreases the accuracy of the physiological data. Also most commercial systems use low sample rates and are often closed systems, which makes it difficult to use the raw data. Current research will determine their usability in our context.
|
14:00
15 mins
|
Transfer Learning Improves Event-Based Modeling in Small and Heterogeneous Datasets
Leontine Ham, Vikram Venkatraghavan, Alle Meije Wink, Wiesje van der Flier, Rebecca Steketee, Marion Smits, Wiro Niessen, Stefan Klein, Esther Bron
Abstract: Understanding the progression of dementia is important for early diagnosis. Event-based models (EBMs) [Fonteijn, 2012] estimate disease progression, which can be represented as a sequence of events, using cross-sectional data. In addition, they output a patient stage which can be used for diagnosis and prognosis. EBMs use mixture modelling, e.g. Gaussian Mixture Model (GMM), to fit the patient- and control distributions of each biomarker. In small and heterogeneous datasets, GMM fitting may fail and thus impede the reliability of EBM.
Therefore, we propose a transfer learning solution that enables EBM of small and heterogeneous datasets by adding biomarker data for a similar disease. Data from a well-defined cohort of a similar disease is expected to aid more accurate estimation of Gaussian parameters. We jointly fit GMM such that Gaussian parameters are shared, while the mixing parameters are estimated independently for the different datasets. We use discriminative EBM (DEBM) [Venkatraghavan, 2018] to estimate the event sequence.
We applied the proposed solution to vascular cognitive impairment (VCI). VCI presents itself with cognitive decline of cerebrovascular origin, has high heterogeneity and often co-occurs with Alzheimer’s disease (AD). VCI datasets are generally small, without a homogeneously diagnosed group. Since AD has similar grey matter (GM) atrophy effects, we used data from AD-subjects to aid the GMM fitting in VCI-subjects. We included subjects from the Heart-Brain Connection study [Hooghiemstra, 2017]; 148 VCI patients, of which 35 with ≥2 cognitive domains affected (VCI2), and 124 cognitively normals (CN). External biomarker data [Bron, 2017; Binnewijzend, 2015] was used in our transfer learning approach: 88 AD, 22 mild cognitive impairment and 125 CN. For both datasets, global- and 74 region GM volumes were computed [Bron, 2014]. We selected 30 significantly different GM volumes (29 ROIs + entire brain) (VCI2-CN; t-test; p<0.05). We evaluated patient staging classification performance (VCI2-CN) using 50-repetition 5-fold cross-validation. AUC was slightly higher for the novel approach (0.866 ± 0.084) than for HBC alone (0.847 ± 0.098) and standard deviation was smaller.
The proposed transfer learning approach improved and stabilized AUC. Hence, DEBM combined with transfer learning is promising for small and heterogeneous datasets.
|
14:15
15 mins
|
Detection of Infant Cries Using Acoustic Features: a Comparison of Classification Techniques
Ahnjili ZhuParris, Max van Gent, Robert-Jan Doll, Geert Jan Groeneveld, Adam Cohen
Abstract: The total duration and frequency of an infant’s cry each day can be indicative of a serious health condition. To track and quantify an infant’s crying behavior in challenging acoustic environments, a cry detection algorithm is needed. Here, we present and compare the performance of four types of classifiers to detect infant crying.
A synthetic dataset was created containing about 400 five second recordings of either infants crying or common domestic sounds (e.g., infants laughing, people talking, and doors closing). We used the openSMILE software for feature extraction; features included pitch and energy related parameters (e.g., Mel-frequency Cepstral coefficients). Classification was done using four different classifiers: Logistic Regression, Random Forest, Support Vector Machines (SVM), and Multi-layer Perceptron classifier to correctly classify crying recordings from non-crying recordings. We then evaluated the hyper-parameter tuning of each classifier using 5 fold cross validation and assessed the predictive power of the network based on a 66% train and test set split.
Here we present the design, implementation and comparison of these algorithms. The preliminary results reflect which audio features and algorithms produce the optimal sensitivity and selectivity for clinical application. The best results were obtained with the random forest and the least favorable results were obtained with the SVM.
This approach of automating the classification of infant cries allows for the detection and quantification of crying frequency and duration which can then inform parents and physicians if an infant is in fact “crying excessively”. Our next step is to validate the predictive power of random forest to classify crying and non-crying audio recordings of infants in a hospital setting. The aim is to verify the accuracy of our infant cry detection classifier using recordings sourced from a mobile phone app in challenge acoustic environments.
|
14:30
15 mins
|
Digital Biomarkers in Early Phase Clinical Trials Using a Remote Monitoring System: a Method Description
Ghobad Maleki, Robert-Jan Doll, Ernst-Jan Bos, Geert Jan Groeneveld, Adam Cohen
Abstract: The pharmacological treatment of mood disorders with available drugs is (partially) ineffective. Therefore, more effective antidepressant drugs are needed. However, a lack of objective biomarkers and reliance on subjective psychometric questionnaires to quantify drug effects hampers the development of novel drugs. Smartphone-based remote monitoring technologies may act as an addition to questionnaires in demonstrating pharmacodynamic (PD) effects in clinical trials. The Centre for Human Drug Research developed a platform allowing the measurement of social and physical activity based on smartphone and wearable sensors (REMOSTM). Here we describe this system in detail and provide an overview of future validation projects.
REMOS is a platform which allows the remote monitoring of healthy subjects and patients, secure data storage, and data management. It includes an Android application for unobtrusive collection of data coming from the smartphone and connected wearable devices (e.g., Withings). The collected data contains data recorded by the smartphone sensors (location, IMU, microphone, and ambient light), as well as phone usage logs (e.g., app usage, call duration, and number of texts). All parameters and sample frequencies are configurable per study-protocol, allowing multiple simultaneous trials. A dashboard enables researchers to monitor the connections and potential data-loss.
After validating the system according to GAMP5 procedures, REMOS will be included in upcoming trials to demonstrate the feasibility and validity. This will be done in two parts: (1) characterization and differentiation between healthy subjects and patients, and (2) studying the sensitivity to pharmacological interventions on physical and social behavior.
|
14:45
15 mins
|
Development of a Realistic On-the-Road Driving Test for Characterization of (impaired) Driving Behavior
Ingrid Koopmans, Hein van der Wall, Ernst Jan Bos, Hemme Hijma, Robert-Jan Doll, Rob Zuiker, Adam Cohen
Abstract: Driving a car is a complex task which requires the integration of various aspects related to the central nervous system (CNS). Impairments in sensory processing, cognitive functions, or motor-function have a negative effect on driving performance, which in turn can result in increased accident rates. Known origins of impaired driving behaviour include sleep deprivation, the consumption of alcohol, or the use of certain types of medication. The current standardized on-the-road-driving test to capture driving behaviour requires subjects to drive on a straight lane at a speed of 95km/h. A single camera then monitors the lateral position of the car on the road. The standard deviation of the lateral position (SDLP) is then used as a measure of driving performance. This method, however, is limited to specific highway trajectories allowing only minimal interactions with external variables (e.g., other drivers). Here, we introduce a method which aims to characterize driving behaviour in more detail, while allowing a more realistic driving test.
The Centre for Human Drug Research (CHDR) equipped a car with multiple sensors and cameras. Besides the lateral position on the road, this car is able to extract both internal (e.g., steering-rotations, throttle, brakes, and gear) and external (e.g., number of and position to surrounding vehicles) parameters. All sensor-data is frequently sampled and synchronously stored on-board. For safety purposes, a driving instructor has access to dual controls (i.e., clutch, brake, and throttle).
To demonstrate the feasibility of this car to allow a more realistic driving test, we first demonstrate the capability of measuring the SDLP. Ten healthy volunteers drove a pre-defined route twice. For both drives, the SDLP was calculated. It was found that the SDLP was reproducible between sessions with relatively low intra-subject variability (6.6-9.4%) compared to the between-subject variability (15.4-19.6%). For the next steps, we are going to (1) identify additional parameters to describe driving behaviour, and (2) study the sensitivity of these parameters to detect changes in driving behaviour. In the end, this car will provide a more elaborate characterization of driving behaviour by including interactions with other road users and allow for flexible trajectories.
|
|