Hypertension is one of the most common health conditions in modern society. Accurate blood pressure monitoring in free-living conditions is important for the precise diagnosis and management of hypertension. In tandem with the advances in wearable and ubiquitous technologies, a medical-grade wearable blood pressure monitor--Omron HeartGuide wrist watch--has recently entered the consumer market.It uses the same mechanism as the upper arm blood pressure monitors and has been calibrated in laboratory settings. Nevertheless, its accuracy "in the wild" has not been investigated. This study aims to investigate the accuracy of the HeartGuide against a medical-grade upper arm blood pressure monitor HEM-1022 in free-living environments. Analysis results suggest that the HeartGuide significantly underestimated systolic pressure and diastolic pressure by an average of 16 mmHg and 6 mmHg respectively. Lower discrepancy between the two devices on diastolic pressure was observed when diastolic pressure increased. In addition, the two devices agreed well on heart rate readings. We also found that device accuracy was related to systolic pressure, heart rate, body temperature and ambient temperature, but was not related salivary cortisol level, diastolic pressure, ambient humidity and air pressure.
Authors: Zilu Liang (Kyoto University of Advanced Science), Mario Chapa-Martell (Silver Egg Technology),
Hide Authors & Abstract

We present a prototype whereby we enabled a humanoid robot to be used to assist mental health patients and their families. Our approach removes the need for Cloud-based automatic speech recognition systems to address healthcare privacy expectations. Furthermore,we describe how the robot could be used in a mental health facility by giving directions from patient selection to metrics for evaluation. Our overarching goal is to make the robot interaction as natural as possible to the point where the robot can develop artificial empathy for the human companion through the interpretation of vocals and facial expressions to infer emotions.
Authors: Miguel Vargas Martin (Ontario Tech University, Canada), Eduardo Perez Valle (Instituto Tecnologico y de Estudios Superiores de Monterrey, Mexico), Sheri Horsburgh (Ontario Shores Centre for Mental Health Sciences, Canada),
Hide Authors & Abstract

Many research articles used Machine Learning (ML) for early detection of Alzheimer's Disease (AD) especially based on Magnetic Resonance Imaging (MRI). Most ML algorithms depend on a large number of hyperparameters. Those hyperparameters have a strong influence on the model performance and thus choosing good hyperparameters is important in ML. In this article, Bayesian Optimization (BO) was used to time-efficiently find good hyperparameters for Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) models, which are based on four and seven hyperparameters and promise good classification results. Those models are applied to distinguish if mild cognitive impaired (MCI) subjects from the Alzheimer's disease neuroimaging initiative (ADNI) dataset will prospectively convert to AD. The results showed comparable cross-validation (CV) classification accuracies for models trained using BO and grid-search, whereas BO has been less time-consuming. The initial combinations for BO were set using Latin Hypercube Design (LHD) and via Random Initialization (RI). Furthermore, many models trained using BO achieved better classification results for the independent test dataset than the model based on the grid-search. The best model achieved an accuracy of 73.43 % for the independent test dataset. This model was an XGBoost model trained with BO and RI.
Authors: Louise Bloch (University of Applied Sciences and Arts Dortmund), Christoph Friedrich (University of Applied Sciences and Arts Dortmund),
Hide Authors & Abstract

Coronary Artery heart Disease (CAD) is the leading cause of mortality in the world. It is a complex and multifactorial disease resulting in several acute coronary syndromes and also to death. In healthcare, an accurate clinical decision support system (CDSS) for CAD prediction has become increasingly important for making granted decision at premature stage. Intensive research have been conducted on improving classification performance using machine learning techniques and metaheuristics algorithms. But, most of these studies introduced the “classic risk factors” for CAD diagnosis such as demographic and clinical data. In this study, we present a novel CDSS based on ensemble learning for CAD prediction and we emphasize also on studying other medical markers i.e. therapy data, some genetic polymorphisms along with classical factors. The new proposed ensemble exploits the potential of three base classifiers to including Support Vector Machines (SVM), Naïve Bayes (NB) and Decision Tree (DT) C4.5 in order to improve the prediction performance. Six experimental data used to build the proposed framework: the first one is collected from a Tunisian biotechnology center and the five other datasets from the University of California at Irvine repository. The analysis of the results of the six studied datasets shows that the proposed CDSS has the highest rate on classification accuracy, precision, recall and F-measure when compared with existing ensemble models
Authors: rawia sammout (the national school of computer science tunisia), Kais Ben SALAH (Computing and Information Technology Faculty of Computing and Information Technology Jeddah, SA), Kahled Ghedira (SSOIE COSMOS National School of Computer Sciences Manouba, Tunisia), Rania Abdelhedi (Laboratory of Molecular and Cellular Screening Processes Centre of Biotechnology of Sfax Sfax, Tunisia), Najla Kharrat (Laboratory of Molecular and Cellular Screening Processes) Centre of Biotechnology of Sfax Sfax, tunisia),
Hide Authors & Abstract

The automated in vitro segmentation of axonal phase-contrast images to allow axonal tracing over time is highly desirable to understand axonal biology in the context of health and disease. While deep learning has become a powerful tool in biomedical image analysis for semantic segmentation tasks, segmentation performance has been limited so far since axons are long and thin objects that are sensitive to under- and/or over-segmentation. We here propose the use of an ensemble-based convolutional neural network (CNN) framework for the segmentation of axons on phase-contrast microscopic images. We show that the mean ResNet-50 ensemble performed better than the max u-net ensemble on the axon segmentation task. We estimated an upper limit for the expected improvement using an oracle machine. Additionally, we introduce a soft version of the Dice coefficient that describes the visually perceived quality of axon segmentation better than the standard Dice. Finally, we show that the mean ResNet-50 ensemble reaches the performance level of human experts. Taken together, we developed a CNN to robustly segment axons on phase-contrast microscopy that will foster further investigations of axonal biology in health and disease.
Authors: Philipp Grüning (University of Lübeck), Alex Palumbo (Fraunhofer EMB Lübeck, University of Lübeck), Svenja Landt (Fraunhofer EMB Lübeck, University of Lübeck), Lara Heckmann (Fraunhofer EMB, University of Lübeck), Leslie Brackhagen (University of Lübeck), Marietta Zille (Fraunhofer EMB Lübeck, University of Lübeck), Amir Madany Mamlouk (University of Lübeck),
Hide Authors & Abstract

The use of contrast agents in CT angiography examinations holds a potential health risk for the patient. Despite this, often unin-tentionally an excessive contrast agent dose is administered. Our goal is to provide a support system for the medical practitioner that advises to adjust an individually adapted dose. We propose a comparison between different means of feature encoding techniques to gain a higher accuracy when recommending the dose adjustment. We apply advanced deep learning approaches and standard methods like principle component analysis to encode high dimensional parameter vectors in a low dimensional feature space. Our experiments showed that features encoded by a regression neural network provided the best results. Especially with a focus on the 90 % precision for the "excessive dose" class meaning that if our system classifed a case as "excessive dose" the ground truth is most likely accordingly. With that in mind a recommendation for a lower dose could be administered without the risk of insuffcient contrast and therefore a repetition of the CT angiography examination. In conclusion we showed that Deep-Learning-based feature encoding on clinical parameters is advantageous for our aim to prevent excessive contrast agent doses.
Authors: Marja Fleitmann (Institute of Medical Informatics, University of Lübeck), Hristina Uzunova (Institute of Medical Informatics, University of Lübeck), Andreas Stroth (Department of Radiology and Nuclear Medicine, UKSH Lübeck), Jan Gerlach (Department of Radiology and Nuclear Medicine, UKSH Lübeck), Alexander Fürschke (Department of Radiology and Nuclear Medicine, UKSH Lübeck), Jörg Barkhausen (Department of Radiology and Nuclear Medicine, UKSH Lübeck), Arpad Bischof (Department of Radiology and Nuclear Medicine, UKSH Lübeck), Heinz Handels (Institute of Medical Informatics, University of Lübeck),
Hide Authors & Abstract

The severity of COVID-19 in patients varies drastically, ranging from asymptomatic infection to severe respiratory failure and death. Currently, few prognostic markers for disease outcomes exist, impairing triaging in hospitals. Here, we train feed-forward neural networks on electronic health records of 819 confirmed SARS-CoV-2 positive patients admitted to a two-site NHS Trust hospital in London, England. To allow early risk assessment, the models ingest data collected in the emergency department (ED) to predict subsequent admission to intensive care, need for mechanical ventilation and in-hospital mortality. We apply univariate selection and recursive feature elimination to find the minimal subset of clinical variables needed for accurate prediction. Our models achieve AUC-ROC scores of 0.78 to 0.87, outperforming standard clinical risk scores. This accuracy is reached with as few as 13% of clinical variables routinely collected within the ED, which increases the practical applicability of such algorithms. Hence, state-of-the-art neural networks can predict severe COVID-19 accurately and early from a small subset of clinical variables.
Authors: Sophie Peacock (Sensyne Health plc, Schrodinger Building, Heatley Road, Oxford Science Park, Oxford, OX4 4GE), Mattia Cinelli (Sensyne Health plc, Schrodinger Building, Heatley Road, Oxford Science Park, Oxford, OX4 4GE), Frank Heldt (Sensyne Health plc, Schrodinger Building, Heatley Road, Oxford Science Park, Oxford, OX4 4GE), Lachlan McLachlan (Sensyne Health plc, Schrodinger Building, Heatley Road, Oxford Science Park, Oxford, OX4 4GE), Marcela Vizcaychipi (Chelsea and Westminster Hospital NHS Foundation Trust, 369 Fulham Road, London, SW10 9NH, UK), Alex McCarthy (Chelsea and Westminster Hospital NHS Foundation Trust, 369 Fulham Road, London, SW10 9NH, UK), Nadezda Lipunova (Sensyne Health plc, Schrodinger Building, Heatley Road, Oxford Science Park, Oxford, OX4 4GE), Robert Fletcher (Sensyne Health plc, Schrodinger Building, Heatley Road, Oxford Science Park, Oxford, OX4 4GE), Anne Hancock (Sensyne Health plc, Schrodinger Building, Heatley Road, Oxford Science Park, Oxford, OX4 4GE), Robert Dürichen (Sensyne Health plc, Schrodinger Building, Heatley Road, Oxford Science Park, Oxford, OX4 4GE), Fernando Andreotti (Sensyne Health plc, Schrodinger Building, Heatley Road, Oxford Science Park, Oxford, OX4 4GE), Rabia Khan (Sensyne Health plc, Schrodinger Building, Heatley Road, Oxford Science Park, Oxford, OX4 4GE),
Hide Authors & Abstract

Neural Networks are powerful classifiers. However, they are black boxes and do not provide explicit explanations for their decisions. For many applications, particularly in health care, explanations are essential for building trust in the model. In the field of computer vision, a multitude of interpretability methods have been developed to analyze Neural Networks by explaining what they have learned during training and what factors influence their decisions. This work provides an overview of these explanation methods in form of a taxonomy. We adapt and benchmark the different methods to time series data, including electrocardiogram (ECG). Further, we introduce quantitative explanation metrics that enable us to build an objective benchmarking framework with which we extensively rate and compare interpretability methods. As a result, we show that the Grad-CAM++ algorithm outperforms all other methods. Finally, we identify the limits of existing explanation methods for specific datasets, with feature values close to zero.
Authors: Thomas Frick (IBM Research Zurich), Stefan Glüge (Zurich University of Applied Sciences), Abbas Rahimi (ETH Zurich), Luca Benini (ETH Zurich), Thomas Brunschwiler (IBM Research Zurich),
Hide Authors & Abstract