Integrating Artificial Intelligence and IoT for Advanced Health Informatics

The advent of Artificial Intelligent (AI) and Machine Learning (ML) has enabled smart devices such as smartphones and tablets to detect and keep track of moving objects in three-dimensional (3D) space. These AI/ML models, such as the ML Kit Pose Detection API by Google and Augmented Reality Kit by Apple Inc., allow the mobile camera to capture a person’s motion to collect robust and repeatable data for functional tasks by accurately determining multidimensional kinematics across joints. With this technology, a cost-effective mobile gait analysis application was developed using a single camera on the commercial smart device. The Lower-Body Motion Tracking version 1.0.1 (LGait) application is designed to support the decision-making of clinicians quantifying mobility by calculating and analyzing the 3D kinematics of walking. The LGait app is designed to run on Apple (Apple Inc., USA) iOS mobile devices (iPhone and iPad). For capturing the body motion kinematics, the Apple ARKit-3 is powered by machine learning models running on the Apple Neural Engine chip using Xcode 11 IDE and Swift programming language. The LGait application provides all the significant features to support lower-limb mobility gait analysis. Two main features of the mobile application are the real-time 3D motion capturing and 3D gait joint angle calculation. Results of tests comparing kinematics acquired from a Vicon motion capture system to the LGait application show a compatible measurement. The proposed applications of the LGait application include the classification of mobility for clinical diagnosis and patient monitoring.

Human activity recognition (HAR) enables numerous application scenarios in ambient assisted living thanks to IoT integration. Healthcare is one of the most prominent use cases of HAR serving individuals who suffer from aging or disabilities as well as healthy people. Remote monitoring of daily activities within home environment may offer assistance in tracking adherence of patients to therapeutical procedures such as exercise monitoring. Within an IoT framework, the type of sensors used influences usability of the HAR system. Accelerometers introduce a noninvasive sensing option as opposed to cameras which intrude into users’ privacy. Recognition of target activities when they are performed among other activities brings about challenges. Typical multi-class classification employed in such recognition tasks necessitates training data collection for all activity types which can be encountered in the prediction stage. Due to unlimited variety of daily living activities, the number of activity classes for which training data should be collected is infinitely many. Expressing the recognition problem in terms of one-class classification (OCC) architecture can aid in resolving this bottleneck. In this chapter, we propose an OCC-based HAR architecture with IoT integration. In our OCC scheme, we utilize artificial data generation (ADG) to generate training data for the negative class based on the target class. In the proposed model, the target class is the only class for which training data are collected. The OCC scheme enables recognizing the target class when the other class is represented with artificially generated data. We present the results of an experimental study for our OCC model on a dataset consisting of ambulatory and static activities.

The need to automatically monitor the state and progression of chronic neurological diseases such as dementia, together with the emergence of state-of-the-art sensing platforms for the home environment offer unprecedented opportunities for automatic behavioural monitoring as a proxy of disease state. However, when these platforms have been deployed, data challenges, including the lack of reliable annotations, limit the applicability of standard machine learning techniques. This chapter specifically seeks to characterise behavioural signatures of mild cognitive impairment (MCI) and Alzheimer’s disease (AD) dementia. We introduce bespoke machine learning techniques accounting for partial annotations to produce behavioural metrics of key symptoms and use these on a novel dataset of longitudinal sensor data from persons with MCI and AD.

Obesity is a serious health problem that is on the rise at the global level. Recent studies suggest that BMI can be inferred from facial images using deep learning-based convolutional neural networks (CNNs) for obesity classification with about 85–90% accuracy. However, training and testing these deep learning models involves high computation and storage due to the involvement of millions of parameters. A recent trend is the use of lightweight CNN models to facilitate on-device computation in resource-constrained mobile and wearable devices. In this study, we evaluate several lightweight CNNs such as MobileNet-V2, ShuffleNet-V2, and lightCNN-29 for BMI prediction and obesity classification from facial images captured using smartphones. The comparative analysis is done with heavyweight VGG-16 and ResNet-50-based CNN models. These lightweight models when deployed on smartphones can act as self-diagnostic tool in weight changes and obesity monitoring. These tools can facilitate remote monitoring of patients, obtaining patients’ vital signs, and in improving the quality of care provided. Self-diagnostic tools would also help in keeping users’ health data private, safe, and secure.

In the last years, the Internet of Things (IoT) has pilot the vision of a smarter world into reality with a massive amount of data and numerous services. The development of smart sensorial media and devices is getting remarkable attention from academia, government, industry, and healthcare communities. IoT-powered systems produce valuable sources of information and can transform healthcare. With the increase of healthcare services in non-clinical environments, which use vital signs provided by sensors and devices connected to patients, the need to mine and process the physiological measurements is growing significantly. The utilization of IoT to support healthcare is possible thanks to the artificial intelligence (AI). AI techniques, like natural language processing, data analytics, machine learning, and its sub-category deep learning, offer immense opportunities including disease diagnosis and monitoring, clinical workflow augmentation, and hospital optimization. The synergy between the IoT and AI is promising to monitor state of health of patients and to move upon traditional healthcare structures. Accompanied by communication technologies, cloud computing, and big data, it led to the emergence of the Smart Health concept. The chapter exhibits a literature review conducted to determine the most important technologies, methodologies, algorithms, and models for smart health systems. In addition, the main benefits and challenges of smart health were explored.

In this chapter, we present an artificial intelligence (AI) and Internet of Things (IoT) platform.

The core functions of this platform are an AI data processing pipeline and an IoT data processing pipeline. In the pipelines, all different types of application-specific data are processed. For applications where an AI is needed, e.g., face/object/scene detection/classification/recognition, an AI engine is presented. For applications where large-scale searching is needed, a search engine is presented. For applications where most data are sensor data, the IoT pipeline is used. These two pipelines are parallel to each other with data communication mechanism. In the data processing core part, they work independently processing different types of data. But on the boundary and interface, they share many supporting functions including the web/mobile app API, user management, and device management.

Technological embeddedness into everyday life and interconnectivity between omnipresent devices, termed the Internet of Things (IoT), have spurned a dedicated research venture in the field of mental health. Recognizing that mental health issues are on an alarming rise, affecting the individual and the society in a progressively multi-faceted nature, and that existing human resources are not sufficient to tackle the crisis, decision-makers have turned to technology to see what opportunities it may offer. More than ever, this endeavor has gained importance due to the COVID-19 pandemic, whose consequences not only severed the already fickle live human contact between the professionals and their patients but also onset a broad mental health crisis stemming from the virus’ impact on health and the implemented measures to control it. The role that IoT-enabled technology plays in this new landscape of digital mental health can be roughly divided into two complementary processes: assessment and intervention. Assessment concerns monitoring, learning about, and recognizing a person’s mental health issues through their physiology, behavior, thinking, emotional and cognitive states, and the context they live in. Intervention follows, and it conforms to the specifics of an assessment, attempting to effect attitude and behavior change in a person. Technology, especially artificial intelligence, enables assessment and intervention to be tailored very specifically to the individual. Omnipresent devices—e.g., smart bracelets—allow increasingly more accurate assessments, which allow not only better interventions but also interventions that can be delivered momentarily—e.g., with an intelligent cognitive assistant on a smartphone—with the continuous interchange of both as the biggest leap forward. Due to the research field still being young and thus not systematized into a coherent framework, even lacking an overview of methods, trends, and directions, this book chapter serves as an early attempt to codify this highly interdisciplinary relationship between technology and mental health.

Diabetes has become one of the major causes of deaths in the world, and diabetic eye complications causing blindness and low vision have greatly increased. The International Diabetes Federation (IDF) (International Diabetes Federation, https://​www.​diabetesatlas.​org/​en/​sections/​worldwide-toll-of-diabetes.​html) reports that about 1 in 11 adults (463 million people) worldwide has diabetes, and 1.6 million deaths are directly attributed to diabetes each year. It also estimates that, by 2035, there will be 600 million people with diabetes, and by 2045 the number will be 700 million.

Diabetic retinopathy (DR) is a complication of diabetes that affects eyes: it originates from the damage of the blood vessels of the light-sensitive tissue of the retina and is among the primary cause of blindness.

Considering the number of patients affected by diabetes worldwide, it is straightforward that an affective screening of potential number of patients affected by DR is of paramount importance. While the primary method for evaluating diabetic retinopathy involves direct and indirect ophthalmoscopy, artificial intelligence (AI) has been on the rise in the eye care sector. AI uses sophisticated algorithms to analyze a vast amount of clinical data in order to provide effective diagnostic insights with the final aim of accomplishing tasks with minimal involvement of human beings. AI is undoubtedly a major frontier in the general healthcare domain. AI tools provide low-cost and effective solutions in supporting early and accurate diagnosis, facilitating the work of specialists, allowing the release of low-cost solutions for effective (self)-diagnosis, and allowing to select specific treatments. Diabetic retinopathy can be revealed by analyzing fundus photograph datasets of patients and therefore is a disease to which AI tools can provide effective support. This chapter describes the state of the art of AI-based DR screening technologies, some of which are already commercially available.

The Electroencephalogram (EEG)-based emotion detection relies on extensive training data. Generalization from training to testing is accomplished by collecting enormous labeled samples during training, leading to a time-consuming and laborious calibration procedure. In the last few years, numerous papers have proposed different semi-supervised and transfer learning approaches for reducing the demand for training data. However, most of them are iterative methods and require substantial training time, which is unfeasible in practice. To address this issue, we propose Fast Online Instance Transfer (FOIT) for enhancing the affective brain–computer interface (aBCI). FOIT heuristically selects auxiliary data from historical sessions and (or) other subjects, which are subsequently combined with the training data for supervised training. After that, a multi-classifier ensemble makes the predictions on the test trials. During the training, since FOIT is a one-shot algorithm, it avoids time-consuming iterations that satisfy the demand for fast response of BCIs. Experimental results show that FOIT (˜35 s) significantly decreases the time cost than iterative methods (˜45–900 s). Meanwhile, FOIT still maintains the accuracy, even improves ˜1–14% of accuracy in some settings. Our method provides a straightforward, fast, and practically feasible solution for enhancing the effectiveness of EEG-based emotion recognition, allowing for various choices of classifiers without constraints. The code is available at https://​github.​com/​JC-Journal-Club/​FOIT.

Internet of Things (IoT) refers to a network of interconnected things that can communicate, sharing data and information over the Internet. IoT encompasses several different types of devices such as smartphones, smartwatches, smart fridges, etc. Besides, IoT became pervasive in several areas over the years, especially in healthcare systems known as the Internet of Medical Things (IoMT). The IoMT encompasses blood pressure and heart rate monitors and specialized implants, such as pacemakers, etc. IoMT enables the connection of devices that generate significant traffic of distributed, heterogeneous, and dynamic data. To analyze the data produced by IoMT devices before becoming obsolete, it is necessary to develop efficient and customized data mining solutions. Association Rules (AR) are a simple and effective unsupervised learning methodology used to extract actionable knowledge from these dynamic data that can be used at the edge of IoMT, or even directly embedded into the IoMt devices.