: Machine Learning (ML) models can predict diseases with noteworthy results. However, when implemented, their generalization are compromised, resulting in lower performances and render healthcare professionals more susceptible into delivering erroneous diagnostics. This study focuses on the use of uncertainty measures to abstain from classifying samples and use the rejected samples as a selection criterion for active learning. For the multi-label classification of cardiac arrhythmias different methods for uncertainty quantification were compared using three Deep Learning (DL) models: a single model and two pseudoensemble models using Monte-Carlo (MC) Dropout and Deep Ensemble (DE) techniques. When tested with an external dataset, the models’ performances dropped from a F1-Score of 96% to 70%, indicating the possibility of dataset shift. The uncertainty measures for classification with rejection resulted in an increase of the rejection rate from 10% in the training set to a range between 30% to 50% on the external dataset. For the active learning approach, 10% of the highest uncertainty samples were used to retrain the models and their performance increased by almost 5%. Although there are still challenges to the implementation of ML models, the results show that uncertainty quantification is a valuable method to employ in safety mechanisms under dataset shift conditions.
: A vast increase in the percentage of elderly people over the past few decades has induced a serious concern among the research fraternity worldwide. Consequently, the large increase in the number of elderly needing assistance because of chronic diseases is expected to take place. Dementia, depression and mental stress are among the most disabling diseases with dangerous consequences such as wandering into hazardous or insecure areas. This wandering, particularly in urban areas can be life threatening. Recently, with the rapid emergence of disruptive technologies like Internet of Things (IoT), Radio Frequency Identification (RFID) and wireless bio sensors, it has become feasible to build systems that combine IoT and the cloud for monitoring the elderly suffering from dementia or depression. Furthermore, mental chronic diseases, such as stress and depression, are becoming a major concern for governments around the globe. The American Psychological Association (APA) categorizes stress, anxiety and depression as main factors for diverse mental health problems. The cost for treating work-related stress, anxiety and depression, is estimated to be around 617 billion euros per year in Europe alone. Wearable devices for monitoring chronic diseases such as mental stress and depression have been considered as game-changers to the way diseases are managed, by measuring vital signs like skin conductance and changes in the levels of biological stress, and sending warnings remotely to an online server. This paper proposes a work in progress Arduino based real-time stress recognition and localization system using wearable RFID and vital sign sensors for elderly suffering from Dementia and mental stress. The current work utilizes the heart rate variability and Electro Dermal Activity wearable sensors based on the Bitalino development system for measuring mental stress and anxiety in a smart home setting for elderly living alone by exposing a number of subjects to stress and anxiety stimulating horror videos. The system was tested successfully in the university lab.
Insufficient fluid intake in older adults is a prevalent and concerning health issue with far-reaching implications. Monitoring fluid intake is particularly important in various healthcare settings, including hospitals, long-term care facilities, and home care, as well as for specific populations such as older adults or individuals with certain medical conditions. This paper presents an investigation to estimate the fluid intake volume using surface Electromyographic (sEMG) sensors. Eleven subjects participated in the experiment, and sEMG recordings of swallows from cups, bottles, and straws were collected. Four features were extracted from the EMG signals. Seven regression algorithms were implemented for quantifying the volume of swallowed fluid: Random Forest (RF), Support Vector Regressor, K-nearest neighbour (KNN), Linear Regressor (LR), Decision Tree (DT), Lasso, and Ridge. The mean sip volume across subjects was 14.85 +/- 5.05 ml. Results showed that using Random Forest as a subject-specific regressor, the root mean square (RMSE) for estimating fluid intake volume using the Mean Absolute Value feature gave 1.37 +/- 1.1 ml, and using Support Vector as a global regressor, the RMSE was 2.5 +/- 1.2 ml using the Waveform Length feature. When applied as global regressors, SVR gave 6.04 +/- 1.7 ml with the Mel Frequency Cepstrum Coefficients feature and 6.36 +/- 1.6 ml with the Willison Amplitude feature. Random Forest gave 6.04 +/- 1.7 ml with the Willison Amplitude feature. These results indicate a step forward in estimating fluid intake volume based on sEMG for hydration monitoring.
: Inertial measurement units are commonly used in motion analysis applications, such as sports training aid, gait analysis, medical diagnosis, or rehabilitation assistance. Linear acceleration and orientation obtained from sensor fusion are employed for the detection and classification of actions, as well as for measuring relevant parameters of the motion. Typically, in multi-sensor setups, a single model of the device is used. However, considering potential end-users, it could be beneficial to allow heterogeneous setups, particularly by including everyday-use devices with built-in inertial sensors, such as smartwatches. In this work, we perform experiments with several different sensors in order to analyze agreement in their measurements. Results indicate that devices of different models are not directly interchangeable, however, in some applications, heterogeneous setups may be viable.
: Electrodermal activity (EDA) reflects changes in electrical conductivity of the skin via activation of the sympathetic nervous system. Ambulatory EDA measurements bring multiple challenges regarding quality assessment and response detection. A signal quality indicator (SQI) is one method to overcome these. This study aimed to investigate the transferability and generalizability of several open-source state-of-the-art SQIs and response detectors regarding their performance against manually annotated EDA of participants in rest. Three annotators identified artifacts and physiological responses in wrist EDA of 45 participants (10.75 hours). The F1-score, precision, and recall of several state-of-the-art SQIs and response detectors were computed on a subset of the annotated data (n=28). The SQIs and response detectors resulted in F1 scores between 3-16% and 18-32%, respectively. These results indicated that current SQIs and response indicators are not performant enough for EDA of subjects in rest, implying similar or worse outcomes for ambulatory EDA. It is suggested that SQIs must be adjusted based on the used device and set-up.
: In recent years, we have been improving the accuracy of semantic segmentation by deepening segmentation models, but large amount of computational resources are required due to the increase in computational complexity. Therefore knowledge distillation has been studied as one of model compression methods. We propose a knowledge distillation method in which the output distribution of a teacher model learned for each class is used as a target of the student model for the purpose of memory compression and accuracy improvement. Experimental results demonstrate that the segmentation accuracy was improved without increasing the computational cost on two different datasets.
: Rehabilitation and motor skill learning approaches based on Action Observation (AO) and Motor Imagery (MI) rely on the assumption that the sensorimotor system is stimulated by AO and MI tasks similarly to the actual execution of a movement. An advantage of AO over MI is that it is less dependent on subject’s imagination ability, and a direct comparison of their effect on cortical activations during complex upper limb movements has been rarely examined. Therefore, in this study we compare sensorimotor event related desynchronization (ERD) patterns, as a measure of cortical activation, collected from 46 healthy volunteers performing AO and MI protocols. In both mu and beta sensorimotor rhythms a stronger ERD was elicited by AO, characterized by an evident lateralization in the contralateral side of the brain with respect to the limb involved in the observed movement.
: The prevalence and severity of depression make it imperative to develop a means to automatically detect it, so as to alleviate the associated mental effort and cost of seeing a dedicated professional. Depression can also co-exist with other conditions, such as Obstructive Sleep Apnea Syndrome (OSAS). In this paper, we build upon our previous work involving sleep staging, detection of OSAS, and detection of depression in OSAS patients, but focus solely on the latter of the three. We use features extracted from EEG, ECG, and breathing signals of 80 subjects suffering from OSAS and half of which also with depression, using 75 % of this 80-subject dataset for training and 10-fold cross-validation and the remainder for testing. We train three models to classify depression: a random forest (RF), a three-layer artificial neural network (3-ANN), and a gated-recurrent unit long short-term memory (GRU-LSTM) recurrent neural network. Our analysis shows that, like our previous work, the 3-ANN is still the best performing model, with the GRU-LSTM following closely behind at an accuracy of 79.0 % and 78.6 %, respectively, but with a smaller F1-score at 80.0 % and 81.6 %. However, we believe that the large increase in computation time and number of learnable parameters does not justify the use of GRU-LSTM over a simple ANN.
: ECGs are a common diagnostic method for diagnosing cardiac pathologies. In this study, the Pearson correlation coefficient is used to examine the latent linear correlations between the leads of a standard 12-lead ECG. We utilize both the original ECG signals from the PTB-XL database and the reconstructed signal generated by a deep learning model, ECGio. We find that leads III, aVL, V1, and V2 are, on average, the leads with the most unique information due to their low correlation with other leads.
In most research studies, much of the gathered information is qualitative in nature.This article focuses on items for which there are multiple rankings that should be optimally combined.More specifically, it describes a supervised stochastic approach, driven by a Boltzmann machine capable of ranking elements related to each other by order of importance.Unlike classic statistical ranking techniques, the algorithm does not need a voting rule for decision-making.The experimental results indicate that the proposed model outperforms two reference rank aggregation algorithms, ELECTRE IV and VIKOR, and it behaves more stable when encountering noisy data.
: Causal discovery is the problem of estimating a joint distribution from observational data. In recent years, hybrid algorithms have been proposed to overcome computational problems that lead to better results. This work presents a hybrid approach that combines PC algorithm independence tests with a bio-inspired Improved Physarum Learner algorithm. The combination indicates improvement in computational time spent and yet consistent structural results.
: Event detection based on biosignals continuously acquired by wearable devices has become an emergent topic. Particularly, real-time event detection with the electrocardiogram (ECG) has been explored to monitor heart conditions and epileptic seizures in the ambulatory. However, ECG acquired in the ambulatory is much more prone to noise and artifacts, due to the dynamic nature of these environments. Therefore, real-time and robust ECG denoising methods are crucial if event detection is meant to succeed. Denoising autoencoders (DAEs) are studied as robust and fast methods to attenuate ECG noise and artifacts. ECG data augmentation techniques are shown to effectively improve the performance of such a deep learning method. Activity and subject specific models are shown to output better ECG denoised estimates, than non-specific ones. And using accelerometry (ACC) as noise reference exemplifies how biosignal multimodality improves ECG attenuation of muscle and motion artifacts. Therefore, this work establishes effective design techniques to be considered when engineering ECG deep learning models.
Turning deficits have been linked to aging and movement disorders and are a common cause of falls and fractures.Despite previous works on the automatic identification of turns and on its relation to fall risk, different algorithms for turn identification have been used, but their agreement and differences have not been investigated.In this study, we compared the two most-used turn-validated algorithms (El-Gohary and Pham) using a dataset comprising real-world data from 171 community-dwelling older adults monitored for one week with a single wearable sensor.The quantity and quality of turn parameters were calculated and used as predictors of future falls.After the analysis, the El-Gohary and Pham algorithms identified 1,063,810 and 942,845 turns, respectively.The agreement of the algorithms showed a very high to moderate correlation for all turn parameters.We found that prospective fallers take longer to perform a turn, and their movements are less smooth when compared to non-fallers.A fall risk assessment model built only on turn parameters showed reasonable performance for both algorithms (AUC = 0.6).Our results show that differences between turn parameters in the algorithms, when averaged at the single-subject level, are less of a concern when looking for associations with prospective falls.
The stylohyoid and the posterior digastric muscles have essentially been measured through indirect imaging method because of the difficulty to measure them.They are small neck muscles, close to each other, that cannot easily be accessed independently.Yet, they showed promising results for a robust and safe indwelling detection of swallowing, both in terms of timing and recruitment.The possibility to thoroughly establish their relevance through their direct functional analysis would enable the development of an implantable active artificial larynx, that would protect the airway during swallowing detection.Therefore, we set up the first standardized procedure that allows their direct measurement through intramuscular electromyography (EMG) and that we report in this paper.We also used submental surface EMG and swallowing sound modalities to access the major time points of the swallowing process.Finally, various exercises, along with swallowing, were performed by the volunteers.16 peoples were measured with our new procedure, and both the stylohyoid and the posterior digastric could be measured independently with no difficulty.Timings and tasks comparison are therefore ongoing.
