A new RNN based machine learning model to forecast COVID-19 incidence, enhanced by the use of mobility data from the bike-sharing service in Madrid.

As a respiratory virus, COVID-19 propagates based on human-to-human interactions with positive COVID-19 cases. The temporal evolution of new COVID-19 infections depends on the existing number of COVID-19 infections and the people's mobility. This article proposes a new model to predict upcoming COVID-19 incidence values that combines both current and near-past incidence values together with mobility data. The model is applied to the city of Madrid (Spain). The city is divided into districts. The weekly COVID-19 incidence data per district is used jointly with a mobility estimation based on the number of rides reported by the bike-sharing service in the city of Madrid (BiciMAD). The model employs a Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) to detect temporal patterns for COVID-19 infections and mobility data, and combines the output of the LSTM layers into a dense layer that can learn the spatial patterns (the spread of the virus between districts). A baseline model that employs a similar RNN but only based on the COVID-19 confirmed cases with no mobility data is presented and used to estimate the model gain when adding mobility data. The results show that using the bike-sharing mobility estimation the proposed model increases the accuracy by 11.7% compared with the baseline model.

Space-Distributed Traffic-Enhanced LSTM-Based Machine Learning Model for COVID-19 Incidence Forecasting.

The COVID-19 virus continues to generate waves of infections around the world. With major areas in developing countries still lagging behind in vaccination campaigns, the risk of new variants that can cause re-infections worldwide makes the monitoring and forecasting of the evolution of the virus a high priority. Having accurate models able to forecast the incidence of the spread of the virus provides help to policymakers and health professionals in managing the scarce resources in an optimal way. In this paper, a new machine learning model is proposed to forecast the spread of the virus one-week ahead in a geographic area which combines mobility and COVID-19 incidence data. The area is divided into zones or districts according to the location of the COVID-19 measuring points. A traffic-driven mobility estimate among adjacent districts is proposed to capture the spatial spread of the virus. Traffic-driven mobility in adjacent districts will be used together with COVID-19 incidence data to feed a new deep learning LSTM-based model which will extract patterns from mobility-modulated COVID-19 incidence spatiotemporal data in order to optimize one-week ahead estimations. The model is trained and validated with open data available for the city of Madrid (Spain) for 3 different validation scenarios. A baseline model based on previous literature able to extract temporal patterns in COVID-19 incidence time series is also trained with the same dataset. The results show that the proposed model, based on the combination of traffic and COVID-19 incidence data, is able to outperform the baseline model in all the validation scenarios.

Deep Spatiotemporal Model for COVID-19 Forecasting.

COVID-19 has caused millions of infections and deaths over the last 2 years. Machine learning models have been proposed as an alternative to conventional epidemiologic models in an effort to optimize short- and medium-term forecasts that will help health authorities to optimize the use of policies and resources to tackle the spread of the SARS-CoV-2 virus. Although previous machine learning models based on time pattern analysis for COVID-19 sensed data have shown promising results, the spread of the virus has both spatial and temporal components. This manuscript proposes a new deep learning model that combines a time pattern extraction based on the use of a Long-Short Term Memory (LSTM) Recurrent Neural Network (RNN) over a preceding spatial analysis based on a Convolutional Neural Network (CNN) applied to a sequence of COVID-19 incidence images. The model has been validated with data from the 286 health primary care centers in the Comunidad de Madrid (Madrid region, Spain). The results show improved scores in terms of both root mean square error (RMSE) and explained variance (EV) when compared with previous models that have mainly focused on the temporal patterns and dependencies.

Tourist Experiences Recommender System Based on Emotion Recognition with Wearable Data.

The collection of physiological data from people has been facilitated due to the mass use of cheap wearable devices. Although the accuracy is low compared to specialized healthcare devices, these can be widely applied in other contexts. This study proposes the architecture for a tourist experiences recommender system (TERS) based on the user's emotional states who wear these devices. The issue lies in detecting emotion from Heart Rate (HR) measurements obtained from these wearables. Unlike most state-of-the-art studies, which have elicited emotions in controlled experiments and with high-accuracy sensors, this research's challenge consisted of emotion recognition (ER) in the daily life context of users based on the gathering of HR data. Furthermore, an objective was to generate the tourist recommendation considering the emotional state of the device wearer. The method used comprises three main phases: The first was the collection of HR measurements and labeling emotions through mobile applications. The second was emotional detection using deep learning algorithms. The final phase was the design and validation of the TERS-ER. In this way, a dataset of HR measurements labeled with emotions was obtained as results. Among the different algorithms tested for ER, the hybrid model of Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks had promising results. Moreover, concerning TERS, Collaborative Filtering (CF) using CNN showed better performance.

Algorithm for the Comparison of Human Periodic Movements Using Wearable Devices.

In the context of teaching-learning of motor skills in a virtual environment, videos are generally used. The person who wants to learn a certain movement watches a video and tries to perform the activity. In this sense, feedback is rarely thought of. This article proposes an algorithm in which two periodic movements are compared, the one carried out by an expert and the one carried out by the person who is learning, in order to determine how closely these two movements are performed and to provide feedback from them. The algorithm starts from the capture of data through a wearable device that yields data from an accelerometer; in this case, the data of the expert and the data of the person who is learning are captured in a dataset of salsa dance steps. Adjustments are made to the data in terms of Pearson iterations, synchronization, filtering, and normalization, and DTW, linear regression, and error analysis are used to make the corresponding comparison of the two datasets. With the above, it is possible to determine if the cycles of the two signals coincide and how closely the learner's movements resemble those of the expert.

Using Absorption Models for Insulin and Carbohydrates and Deep Leaning to Improve Glucose Level Predictions.

Diabetes is a chronic disease caused by the inability of the pancreas to produce insulin or problems in the body to use it efficiently. It is one of the fastest growing health challenges affecting more than 400 million people worldwide, according to the World Health Organization. Intensive research is being carried out on artificial intelligence methods to help people with diabetes to optimize the way in which they use insulin, carbohydrate intakes, or physical activity. By predicting upcoming levels of blood glucose concentrations, preventive actions can be taken. Previous research studies using machine learning methods for blood glucose level predictions have mainly focused on the machine learning model used. Little attention has been given to the pre-processing of insulin and carbohydrate signals in order to mimic the human absorption processes. In this manuscript, a recurrent neural network (RNN) based model for predicting upcoming blood glucose levels in people with type 1 diabetes is combined with several carbohydrate and insulin absorption curves in order to optimize the prediction results. The proposed method is applied to data from real patients suffering type 1 diabetes mellitus (T1DM). The achieved results are encouraging, obtaining accuracy levels around 0.510 mmol/L (9.2 mg/dl) in the best scenario.

Learning Carbohydrate Digestion and Insulin Absorption Curves Using Blood Glucose Level Prediction and Deep Learning Models.

Type 1 diabetes is a chronic disease caused by the inability of the pancreas to produce insulin. Patients suffering type 1 diabetes depend on the appropriate estimation of the units of insulin they have to use in order to keep blood glucose levels in range (considering the calories taken and the physical exercise carried out). In recent years, machine learning models have been developed in order to help type 1 diabetes patients with their blood glucose control. These models tend to receive the insulin units used and the carbohydrate taken as inputs and generate optimal estimations for future blood glucose levels over a prediction horizon. The body glucose kinetics is a complex user-dependent process, and learning patient-specific blood glucose patterns from insulin units and carbohydrate content is a difficult task even for deep learning-based models. This paper proposes a novel mechanism to increase the accuracy of blood glucose predictions from deep learning models based on the estimation of carbohydrate digestion and insulin absorption curves for a particular patient. This manuscript proposes a method to estimate absorption curves by using a simplified model with two parameters which are fitted to each patient by using a genetic algorithm. Using simulated data, the results show the ability of the proposed model to estimate absorption curves with mean absolute errors below 0.1 for normalized fast insulin curves having a maximum value of 1 unit.

Gait-Based Identification Using Deep Recurrent Neural Networks and Acceleration Patterns.

This manuscript presents an approach to the challenge of biometric identification based on the acceleration patterns generated by a user while walking. The proposed approach uses the data captured by a smartphone's accelerometer and gyroscope sensors while the users perform the gait activity and optimizes the design of a recurrent neural network (RNN) to optimally learn the features that better characterize each individual. The database is composed of 15 users, and the acceleration data provided has a tri-axial format in the X-Y-Z axes. Data are pre-processed to estimate the vertical acceleration (in the direction of the gravity force). A deep recurrent neural network model consisting of LSTM cells divided into several layers and dense output layers is used for user recognition. The precision results obtained by the final architecture are above 97% in most executions. The proposed deep neural network-based architecture is tested in different scenarios to check its efficiency and robustness.

Deep Physiological Model for Blood Glucose Prediction in T1DM Patients.

Accurate estimations for the near future levels of blood glucose are crucial for Type 1 Diabetes Mellitus (T1DM) patients in order to be able to react on time and avoid hypo and hyper-glycemic episodes. Accurate predictions for blood glucose are the base for control algorithms in glucose regulating systems such as the artificial pancreas. Numerous research studies have already been conducted in order to provide predictions for blood glucose levels with particularities in the input signals and underlying models used. These models can be categorized into two major families: those based on tuning glucose physiological-metabolic models and those based on learning glucose evolution patterns based on machine learning techniques. This paper reviews the state of the art in blood glucose predictions for T1DM patients and proposes, implements, validates and compares a new hybrid model that decomposes a deep machine learning model in order to mimic the metabolic behavior of physiological blood glucose methods. The differential equations for carbohydrate and insulin absorption in physiological models are modeled using a Recurrent Neural Network (RNN) implemented using Long Short-Term Memory (LSTM) cells. The results show Root Mean Square Error (RMSE) values under 5 mg/dL for simulated patients and under 10 mg/dL for real patients.

Using Recurrent Neural Networks to Compare Movement Patterns in ADHD and Normally Developing Children Based on Acceleration Signals from the Wrist and Ankle.

Attention deficit and hyperactivity disorder (ADHD) is a neurodevelopmental condition that affects, among other things, the movement patterns of children suffering it. Inattention, hyperactivity and impulsive behaviors, major symptoms characterizing ADHD, result not only in differences in the activity levels but also in the activity patterns themselves. This paper proposes and trains a Recurrent Neural Network (RNN) to characterize the moment patterns for normally developing children and uses the trained RNN in order to assess differences in the movement patterns from children with ADHD. Each child is monitored for 24 consecutive hours, in a normal school day, wearing 4 tri-axial accelerometers (one at each wrist and ankle). The results for both medicated and non-medicated children with ADHD, and for different activity levels are presented. While the movement patterns for non-medicated ADHD diagnosed participants showed higher differences as compared to those of normally developing participants, those differences were only statistically significant for medium intensity movements. On the other hand, the medicated ADHD participants showed statistically different behavior for low intensity movements.

Automatic Extraction and Detection of Characteristic Movement Patterns in Children with ADHD Based on a Convolutional Neural Network (CNN) and Acceleration Images.

Attention deficit and hyperactivity disorder (ADHD) is a neurodevelopmental disorder, which is characterized by inattention, hyperactivity and impulsive behaviors. In particular, children have difficulty keeping still exhibiting increased fine and gross motor activity. This paper focuses on analyzing the data obtained from two tri-axial accelerometers (one on the wrist of the dominant arm and the other on the ankle of the dominant leg) worn during school hours by a group of 22 children (11 children with ADHD and 11 paired controls). Five of the 11 ADHD diagnosed children were not on medication during the study. The children were not explicitly instructed to perform any particular activity but followed a normal session at school alternating classes of little or moderate physical activity with intermediate breaks of more prominent physical activity. The tri-axial acceleration signals were converted into 2D acceleration images and a Convolutional Neural Network (CNN) was trained to recognize the differences between non-medicated ADHD children and their paired controls. The results show that there were statistically significant differences in the way the two groups moved for the wrist accelerometer (t-test p-value <0.05). For the ankle accelerometer statistical significance was only achieved between data from the non-medicated children in the experimental group and the control group. Using a Convolutional Neural Network (CNN) to automatically extract embedded acceleration patterns and provide an objective measure to help in the diagnosis of ADHD, an accuracy of 0.875 for the wrist sensor and an accuracy of 0.9375 for the ankle sensor was achieved.

Detecting Steps Walking at very Low Speeds Combining Outlier Detection, Transition Matrices and Autoencoders from Acceleration Patterns.

In this paper, we develop and validate a new algorithm to detect steps while walking at speeds between 30 and 40 steps per minute based on the data sensed from a single tri-axial accelerometer. The algorithm concatenates three consecutive phases. First, an outlier detection is performed on the sensed data based on the Mahalanobis distance to pre-detect candidate points in the acceleration time series that may contain a ground contact segment of data while walking. Second, the acceleration segment around the pre-detected point is used to calculate the transition matrix in order to capture the time dependencies. Finally, autoencoders, trained with data segments containing ground contact transition matrices from acceleration series from labeled steps are used to reconstruct the computed transition matrices at each pre-detected point. A similarity index is used to assess if the pre-selected point contains a true step in the 30-40 steps per minute speed range. Our experimental results, based on a database from three different participants performing similar activities to the target one, are able to achieve a recall = 0.88 with precision = 0.50 improving the results when directly applying the autoencoders to acceleration patterns (recall = 0.77 with precision = 0.50).

Assessing Gait Impairments Based on Auto-Encoded Patterns of Mahalanobis Distances from Consecutive Steps.

Insole pressure sensors capture the force distribution patterns during the stance phase while walking. By comparing patterns obtained from healthy individuals to patients suffering different medical conditions based on a given similarity measure, automatic impairment indexes can be computed in order to help in applications such as rehabilitation. This paper uses the data sensed from insole pressure sensors for a group of healthy controls to train an auto-encoder using patterns of stochastic distances in series of consecutive steps while walking at normal speeds. Two experiment groups are compared to the healthy control group: a group of patients suffering knee pain and a group of post-stroke survivors. The Mahalanobis distance is computed for every single step by each participant compared to the entire dataset sensed from healthy controls. The computed distances for consecutive steps are fed into the previously trained autoencoder and the average error is used to assess how close the walking segment is to the autogenerated model from healthy controls. The results show that automatic distortion indexes can be used to assess each participant as compared to normal patterns computed from healthy controls. The stochastic distances observed for the group of stroke survivors are bigger than those for the people with knee pain.

Time-Elastic Generative Model for Acceleration Time Series in Human Activity Recognition.

Body-worn sensors in general and accelerometers in particular have been widely used in order to detect human movements and activities. The execution of each type of movement by each particular individual generates sequences of time series of sensed data from which specific movement related patterns can be assessed. Several machine learning algorithms have been used over windowed segments of sensed data in order to detect such patterns in activity recognition based on intermediate features (either hand-crafted or automatically learned from data). The underlying assumption is that the computed features will capture statistical differences that can properly classify different movements and activities after a training phase based on sensed data. In order to achieve high accuracy and recall rates (and guarantee the generalization of the system to new users), the training data have to contain enough information to characterize all possible ways of executing the activity or movement to be detected. This could imply large amounts of data and a complex and time-consuming training phase, which has been shown to be even more relevant when automatically learning the optimal features to be used. In this paper, we present a novel generative model that is able to generate sequences of time series for characterizing a particular movement based on the time elasticity properties of the sensed data. The model is used to train a stack of auto-encoders in order to learn the particular features able to detect human movements. The results of movement detection using a newly generated database with information on five users performing six different movements are presented. The generalization of results using an existing database is also presented in the paper. The results show that the proposed mechanism is able to obtain acceptable recognition rates (F = 0.77) even in the case of using different people executing a different sequence of movements and using different hardware.

Assessing Walking Strategies Using Insole Pressure Sensors for Stroke Survivors.

Insole pressure sensors capture the different forces exercised over the different parts of the sole when performing tasks standing up such as walking. Using data analysis and machine learning techniques, common patterns and strategies from different users to achieve different tasks can be automatically extracted. In this paper, we present the results obtained for the automatic detection of different strategies used by stroke survivors when walking as integrated into an Information Communication Technology (ICT) enhanced Personalised Self-Management Rehabilitation System (PSMrS) for stroke rehabilitation. Fourteen stroke survivors and 10 healthy controls have participated in the experiment by walking six times a distance from chair to chair of approximately 10 m long. The Rivermead Mobility Index was used to assess the functional ability of each individual in the stroke survivor group. Several walking strategies are studied based on data gathered from insole pressure sensors and patterns found in stroke survivor patients are compared with average patterns found in healthy control users. A mechanism to automatically estimate a mobility index based on the similarity of the pressure patterns to a stereotyped stride is also used. Both data gathered from stroke survivors and healthy controls are used to evaluate the proposed mechanisms. The output of trained algorithms is applied to the PSMrS system to provide feedback on gait quality enabling stroke survivors to self-manage their rehabilitation.

Human Movement Recognition Based on the Stochastic Characterisation of Acceleration Data.

Human activity recognition algorithms based on information obtained from wearable sensors are successfully applied in detecting many basic activities. Identified activities with time-stationary features are characterised inside a predefined temporal window by using different machine learning algorithms on extracted features from the measured data. Better accuracy, precision and recall levels could be achieved by combining the information from different sensors. However, detecting short and sporadic human movements, gestures and actions is still a challenging task. In this paper, a novel algorithm to detect human basic movements from wearable measured data is proposed and evaluated. The proposed algorithm is designed to minimise computational requirements while achieving acceptable accuracy levels based on characterising some particular points in the temporal series obtained from a single sensor. The underlying idea is that this algorithm would be implemented in the sensor device in order to pre-process the sensed data stream before sending the information to a central point combining the information from different sensors to improve accuracy levels. Intra- and inter-person validation is used for two particular cases: single step detection and fall detection and classification using a single tri-axial accelerometer. Relevant results for the above cases and pertinent conclusions are also presented.