Melanoma is the deadliest form of skin cancer. Early detection of melanoma is vital, as it helps in decreasing the death rate as well as treatment costs. Dermatologists are using image-based diagnostic tools to assist them in decision-making and detecting melanoma at an early stage. We aim to develop a novel handheld medical scanning device dedicated to early detection of melanoma at the primary healthcare with low cost and high performance. However, developing this particular device is very challenging due to the complicated computations required by the embedded diagnosis system. In this paper, we propose a hardware-friendly design for implementing an embedded system by exploiting the recent hardware advances in reconfigurable computing. The developed embedded system achieved optimized implementation results for the hardware resource utilization, power consumption, detection speed and processing time with high classification accuracy rate using real data for melanoma detection. Consequently, the proposed embedded diagnosis system meets the critical embedded systems constraints, which is capable for integration towards a cost- and energy-efficient medical device for early detection of melanoma.

This paper presents a novel hybrid model for classifying time series of heart sound signal using time-growing neural network. The proposed hybrid model takes segmental behaviour of heart sound signal into account by combining two different deep learning methods, the Static and the Moving Time-Growing Neural Network, which we call STGNN and MTGNN, respectively. Flexibility of the model in learning both deterministic and stochastic segments of signal allows it to learn those complicated characteristics of heart sound signal caused by any obstruction on semilunar heart valve. The model is trained to distinguish between a patient group and a reference group. The patient group is comprised of the subjects with the semilunar heart valve abnormalities including aortic stenosis, pulmonary stenosis and bicuspid aortic valve, whereas the reference group which is composed of the individuals with the heart abnormalities other than those of the reference group or the healthy ones. The model is validated using two different databases: one comprised of 140 children with various heart conditions, and the other one constituted of 50 elderly patients with aortic stenosis. Both the datasets were collected from the referrals to the University hospitals. The overall accuracy and sensitivity of the model are estimated to be 84.2% and 82.8%, respectively. The results show that the model exhibits sufficient capability to distinguish between the patient and the reference group in such a demanding clinical application. 

Eight ECG beat detection algorithms, from the PhysioNet's WFDR and Cardiovascular Signal toolboxes, were tested on twenty measurements, obtained by the Savvy patch ECG device, for their accuracy in beat detection. On each subject, one measurement is obtained while sitting and one while running. Each measurement lasted from thirty seconds to one minute. The measurements obtained while running were more challenging for all the algorithms, as most of them almost perfectly detected all the beats on the measurements obtained in sitting position. However, when applied on the measurements obtained while running, all the algorithms have performed with decreased accuracy. Considering overall percentage of the faulty detected peaks, the four best algorithms were jqrs, from the Cardiovascular Signal Toolbox, and ecgpuwave, gqrs, and wqrs, from the WEDB Toolbox, with percentages of faulty detected beats 1.7, 2.3, 2.9, and 3, respectively.

Eight ECG beat detection algorithms, from the PhysioNet's WFDB and Cardiovascular Signal toolboxes, were tested on twenty measurements, obtained by the Savvy patch ECG device, for their accuracy in beat detection. On each subject, one measurement is obtained while sitting and one while running. Each measurement lasted from thirty seconds to one minute. The measurements obtained while running were more challenging for all the algorithms, as most of them almost perfectly detected all the beats on the measurements obtained in sitting position. However, when applied on the measurements obtained while running, all the algorithms have performed with decreased accuracy. Considering overall percentage of the faulty detected peaks, the four best algorithms were jqrs, from the Cardiovascular Signal Toolbox, and ecgpuwave, gqrs, and wqrs, from the WFDB Toolbox, with percentages of faulty detected beats 1.7, 2.3, 2.9, and 3, respectively. 

The atrial irregular rhythm, often reflected in atrial fibrillation, undulation or flutter, is recognized as one of the major causes of brain stroke and entails an increased risk of thromboembolic events because it increases the likelihood of blood clots formation. Its early detection is becoming an increasingly important preventive measure. The paper presents a simple methodology for the detection of atrial irregular rhythm by ECG body gadget that can perform long-term measurements, e.g. several weeks or more. Multichannel ECG, on the body surface, gives a more detailed insight into the atrial activity in comparison to standard 12-lead ECG. The information from MECG is compared with single-channel patch ECG. The obtained results suggest that the proposed methodology could be useful in treatments of atrial irregular rhythm. One can obtain a reliable information about the time and duration of fibrillation events, or determine arrhythmic focuses and conductive pathways in heart atria, or study the effects of antiarrhythmic drugs on existing arrhythmias and on an eventual development of new types of arrhythmias. 

We present modified antenna-like devices - applicators - for direct detection of internal inhomogeneities such as breast tumours and brain haemorrhages, at a frequency about 1 GHz. This direct detection provides the possibility of using a simple microwave generator and simple rectification and position registration of the received signals. Direct readouts are thus possible, without any massive computing resources as with tomographic imaging. The transmitting applicator is non-contacting and in free air close to the object. It generates an essentially quasistatic axial magnetic field which induces a circular electric field in the tissue. The receiving 3D contacting applicator contains a high-permittivity ceramic and is resonant. Its mode field provides the desired polarisation sensitivity and filters out the main electric field. The overall system sensitivity for detection of internal inhomogeneities is accomplished by optimised use of the orthogonality of the primary magnetic, induced electric, and diffracted electric fields. When developments are completed, the system will replace or complement existing commercial technologies at a low cost.

ESS-H+ will continue the work in the KKS research profile Embedded Sensor Systems for Health (ESS-H) at Mälardalen University. ESS-H has during six years provided important collaboration between researchers, industrial partners and healthcare organizations within three focus areas (Health monitoring at home, Health monitoring at work, and Infrastructure and communication).

The focus of the new research profile ESS-H+ is that monitoring of humans should be able to be performed anytime, anywhere. Our research challenges are focused to the areas of Reliable acquisition and management of physiological data, and Reliable distributed data analysis. Reliable acquisition and management of physiological data is a fundamental prerequisite for advancing anytime, anywhere health monitoring, towards enabling remote monitoring of more serious health conditions than what is safely possible today. Reliable distributed data analysis is a fundamental prerequisite for enabling large-scale deployment of anytime, anywhere health monitoring. Both research challenges are complex and require multi-disciplinary research.

The following important research goals will be addressed within ESS-H+: 

• Reliable data acquisition, and design of suitable sensor systems for achieving this. 
• Development of analysis and classification algorithms for physiological parameters. 
• Achieve efficient distributed data fusion and decision support. 
• Better utilization of the compute power of sensor nodes, and increased communication reliability: safety as well as security. 
• Efficient integration of scientific results, from different scientific areas, to an efficient and user-friendly embedded sensor system. 

The research in ESS-H+ will include research within the scientific areas (Biomedical sensor technology, Biomedical signal processing, Intelligent decision support, and Reliable and secure data communication) but also a strong integration between these areas, our collaborating companies, and the end users. Our main challenge within ESS-H+ will be this interdisciplinary integration that aims towards fully operating systems, thud providing efficient integration of scientific results, from different scientific areas, to an efficient and user-friendly embedded sensor system.

Continuous blood pressure (BP) monitoring can produce a significant amount of digital data, which increases the chance of early diagnosis and improve the rate of survival for people diagnosed with hypertension and Cardiovascular diseases (CVDs). However, mining and processing this vast amount of data are challenging. This research is aimed to address this challenge by proposing a deep learning technique, convolutional neural network (CNN), to estimate the systolic blood pressure (SBP) using electrocardiogram (ECG) and photoplethysmography (PPG) signals. Two different methods are investigated and compared in this research. In the first method, continuous wavelet transform (CWT) and CNN have been employed to estimate the SBP. For the second method, we used random sampling within the stochastic gradient descent (SGD) optimization of CNN and the raw ECG and PPG signals for training the network. The Medical Information Mart for Intensive Care (MIMIC III) database is used for both methods, which split to two parts, 70% for training our network and the remaining used for testing the performance of the network. Both methods are capable of learning how to extract relevant features from the signals. Therefore, there is no need for engineered feature extraction compared to previous works. Our experimental results show high accuracy for both CNN-based methods which make them promising and reliable architectures for SBP estimation.

Traditionally, research education is performed within the universities, and the PhD students are working within a research group. However, technical development and also research is performed within companies, and the need to keep up with the latest findings in research and to strengthen the competence within the private business sector is increasing. At Mälardalen University, we have experience from working in several Industrial Graduate Schools. The collaboration with the companies gets intensified and deepened trough such programmes, and the university tends to keep the good contact with previous PhD students and their companies also many years after their graduation. The Graduate Schools also give the companies good insight in the university world. Presently, we are involved in two Graduate Schools, and several of the PhD projects are focusing within Biomedical Engineering. Further, one of the graduate schools is linked to the research profile Embedded Sensor Systems for Health, which is supported from the same financier. Companies are involved also in the research profile, and through these activities, the Industrial PhD students form a critical mass and can exchange both experience and knowledge with other companies and with university researchers.