Abstract
Falls, especially in the elderly, are one of the main factors of hospitalization. Time-consuming intervention can be fatal or cause irreversible damages to the victims. On the other hand, there is currently a significant amount of smart clothing equipped with various sensors, particularly gyroscopes and accelerometers, which can be used to detect accidents. The creation of a tool that automatically detects eventual falls allows helping the victims as soon as possible. This works focuses in the automatic fall detection from sensors signals using long short-term memory networks. To train and test this approach, the Sisfall dataset is used, which considers the simulation of 23 adults and 15 older people. These simulations are based on everyday activities and the falls that may result from their execution. The results indicate that the procedure provides an accuracy score of 97.1% on the test set.

Abstract
The knowledge of hemodynamic behaviour in the abdominal aorta artery bifurcation is of great importance for the early diagnosis of several cardiovascular diseases common in this bifurcation. The work developed focuses on a case study of hemodynamic in the abdominal aorta artery bifurcation, based on a realistic 3D geometric model reconstructed from 2D medical images of a real patient. Hemodynamic quantities based on the wall shear stress (WSS) of the abdominal aorta bifurcation are analysed and is presented an alternative analysis of the well-established stress hemodynamic descriptors to identify specific zones of the artery with a higher probability of developing cardiovascular diseases. The individual analysis of different zones of the artery allowed to obtain information that can remain masked when whole artery is considered as a single zone. The reported results provide a correlation between the analysed stress hemodynamic descriptors and the area of the wall artery. Then, the aim of this work is the identification of regions at the luminal surface subject to atherosusceptible WSS phenotypes. For the patient studied, the analysis presented allowed the identification of the patient's propensity to develop atherosclerosis, according to the hemodynamic descriptors time-averaged WSS (TAWSS), oscillatory shear index (OSI), and relative residence time (RRT). Thus, this work offers a new way of looking to the stress hemodynamic descriptors.

Abstract
Background: The gap between offer and need for a kidney transplant (KT) has been increasing. The Kidney Donor Profile Index (KDPI) is a measure of organ quality and allows estimation of graft survival, but could not apply to all populations. Knowledge of our kidney donor and recipient population is vital to adjust transplant strategies. Methods: We performed a retrospective evaluation of donors and recipients of KT regarding two kidney transplant units: Centro Hospitalar Universitario de Coimbra, CHUC (Coimbra, Portugal) and Centro Hospitalar Universitario de Sao Joao, CHUSJ (Porto, Portugal), between 2013 and 2018. We then did statistical analysis and modeling, correlating these KT outcomes with donor and recipient characteristics, including KDPI. Artificial intelligence methods were performed to determine the best predictors of graft survival. Results: We analyzed a total of 808 kidney donors and 829 recipients of KT. The association between KDPI and graft dysfunction was only moderate. The decision tree machine learning algorithm proved to be better at predicting graft failure than artificial neural networks. Multinomial logistic regression revealed recipient age as an important prognostic factor for graft loss. Conclusions: In this Portuguese cohort, KDPI was not a good measure of KT survival, although it correlated with GFR 1 year post-transplant. The decision tree proved to be the best algorithm to predict graft failure. Age of the recipient was the most important predictor of graft dysfunction.

Abstract
Congestive heart failure (CHF) is a severe condition that affects the pumping power of your cardiac muscle. In this work, long-term memory (LSTM) and Bidirectional LSTM (BiLSTM) networks were used to identify congestive heart failure human beings using datasets from the PhysioNET. Two approaches were adopted, the first considers beating signals directly to feed the LSTM networks, and the second one used features signals extracted from the beating signals. The BiLSTM considering features signals obtain the best results reaching an accuracy of 90%.

Abstract
A classifier using a Long Short-Term Memory (LSTM) network to identify human beings infected with Covid-19 is proposed in this work. This classifier has significant advantages over current testing methods: it is fast, contactless, and requires few monetary resources. The data considered for this study was extracted from the Coswara dataset using 140 individuals (70 healthy and 70 infected with Covid-19). This dataset contains respiratory signals, such as people counting numbers, coughing, or breathing. The classifier uses non-linear time sequence features extracted from the signals after a preprocessing stage. The classifier was able to discriminate whether a human is infected with Covid-19 with an accuracy of 92.1%, specificity of 85.7%, and sensitivity of 98.6% using 5-fold Cross-Validation. Based on the results obtained, the classifier can be used as an alternative for the Reverse Transcription Polymerase Chain Reaction (RT-PCR) tests.

Abstract
Nowadays, independent older people stay alone for long periods, which increases the risk of being seriously damaged after a fall without the quick attendance of medical services. Several smart clothing equipment was created to detect these falls using sensors such as accelerometers and gyroscopes, allowing a short intervention to the victims. This work considers the Sisfall database, where 23 adults and 15 older people performed several daily living simulations. The signals registered by three sensors were used to train a Long Short-Term Memory network and a Bi-Long Short-Term Memory network to detect everyday activities and falls. Several experiments were performed, where the BiLSTM model outperforms the LSTM model with a mean accuracy of 99.21% on the testing set.