Artificial Intelligence for Balance Control of Elderly People
Designing automated balance control system for elderly people is one of the key project of our Compassionate AI Lab. Here, Dr. Amit Ray discuses about one of the recent project of AI using deep learning algorithms for automatic balance control of elderly people. He explains how machine learning algorithms can be used to study and improve the dynamical properties of postural stability of elderly people. The project focuses on how image recognition, human-body joint dynamics, and path navigation methods of artificial intelligence can be used to eliminate the imbalance, fall and injury of elderly people or for physically challenged people.
Compassionate Artificial Intelligence can be used for helping elderly people in many ways. Here, we discuss about one of our recent project of using AI & deep learning techniques for automatic balance control. The machine learning algorithms are used to improve dynamical properties of postural stability. In this project AI based machine learning algorithms are used to find the insights into the person specific postural strategies for older adults in order to adapt to the postural challenges during sleeping, standing, turning and walking. To study the body movement behavior of elderly people accurately, it is necessary to observe and record their movement trajectory and joint movements quantitatively and precisely in three dimensions.
Why Balance Control is so Important?
Falls of elderly people and weak patients are considered to be an important public health problem due to the risks of injury and mortality. Imbalance is one of the main cause of falls. Good balance is essential for upright stance and for conducting daily living activities for elderly people.
Elderly patients, especially with back-pain decreased their reliance on ankle strategy and hip strategy during balance control. With age, the body undergoes changes in the brain, bones, muscles, and joints that lead to changes in the physiological spinal curvature. Changes in the projection of the center of gravity negatively affect body balance control.
It is projected that, by 2050, one-fifth of the world’s population will be ≥ 60 years. According to the report of the World Health Organization, approximately 28~35% of people aged 65 and over fall each year and 32~42% for those over 70 years of age. Falls presented the major mechanism for hospital admission (59.9%) in the elderly patients . Fall-related injuries have a great impact on the quality of life of old people and have huge medical and social costs. According to one study , Seventy three (89%) of all incidents occurred at home and nine (11%) at streets. Eighty three (90.2%) patients fell from the same level, four (4.3%) from stairs, and five (5.4%) from more than one meter high . Ground Level Fall (GLF) has becomes a leading mechanism of injury in the elderly. Often, it causes hip fracture and it is fast becoming a significant cause of mortality among the elderly people.
Balance Control Mechanisms
There are two types of balance control: static balance control and dynamic balance control under perturbed external conditions. Balance control during walking and turning is especially dynamic in nature, involving coordinated adjustments in posture (hand, head and trunk stabilization) and foot placement and foot pressure manipulation from step to step. The ankles produces the force to support the body weight and work must be done to lift and propel the body. Mostly, these demands are met by the muscles that produce force while minimizing mechanical work. The pair of knees support nearly the entire weight of the human body, and help the body perform different mobility functions in the locomotion. We have noticed that recording and controlling the dynamics of ankle joints, knee joints and hip Joints are very critical in modeling the automated balance control system.
Neuroscience of Balance Control:
Balance is a participant’s ability to control the center of mass of the body within the base of support and to compensate under external disturbance, unevenness and uneasiness. When healthy humans stand unperturbed on a solid surface in a brightly lighted surrounding, most of the input to the central nervous system (CNS) is received from the somatosensory system. The vestibular, somatosensory (including proprio-ceptive and cutaneous inputs) and visual systems are involved in the complex process of maintaining upright balance in humans. Afferent nerves carries sensory information from the periphery of the body to the central nervous system. A deficit in any one of these systems or in the integration of information from these systems could affect balance. Somatosensory function declines with age, and such changes have been mainly associated with diminished motor performance and balance.
Sensors for Balance Control
Two types of sensors can be used: non-wearable floor sensors and body-mounted or wearable sensors. In floor sensor-approach, movement parameters are measured using pressure sensors, force sensors and moment transducers fitted on the instrumented platforms.
Advances in hardware systems have allowed for free movement in space, with wireless data transmission. Numerous body-worn sensors are required for a full-body motion analysis and control. However, we want the system to use minimal body-worn sensor setup. In our setup, the system uses different sensors, including the skeletal-tracking feature of kinect sensors. The kinect tracker often captures unnatural human poses such as discontinuous and vibrated motions when self-occlusions occur. Typically, kinect sensors includes a video camera, an IR emitter and IR depth sensor, a microphone array for speech recognition and a tilt motor to track the movements of the user.
The main devices for this project are wearable sensors, foot pressure sensors, floor sensors, video cameras, Functional Near-Infrared Spectroscopy and smart phones. The wearable sensor-based systems are used to get the physiological measures such as joint angels, joint forces, and muscle activity. Our image processing based fall detection method uses spatio-temporal context tracking over three-dimensional (3D) images.
Currently the use of three-dimensional camera-based visual posture tracking system is progressing fast. However, the main challenge for visual tracking is to handle large appearance changes of the target object and the background over time due to occlusion, illumination changes, and pose variation. Numerous algorithms have been experimented with focus on effective appearance models, which are based on the target appearance. To improve the accuracy, here, we tried to combine non-vision-based methods and vision-based method in a judicial way.
The dynamics of ankle joint, knee joint and hip Joint are very critical in modeling the automated balance control system. Dynamics, including the calculation of joint forces, joint angles are very important. Knee-joint dynamics during gait. It is important to measure knee angle during walking , rotating and seating. The 19-point walking model is shown below.
Dynamics of the body center point is another critical factor. Dynamic foot pressure measurement is important for postural analysis. Pressure measurement sensors are used which records all the relevant information needed to analyze the foot’s behavior.
Artificial Intelligence Algorithms for Balance Control
Model Free and Model Based Algorithms
Deep Learning and Deep Reinforcement Learning for Balance Control
Over the last decades, several deep learning architectures have been developed including: Deep Neural Network (DNN), Recurrent Neural Network (RNN), Convolutional Neural Network (CNN), Deep Autoencoder (DA), Deep Boltzmann Machine (DBM), Restricted Boltzmann Machine (RBM) and Deep Belief Network (DBN), Deep Residual Network, Deep Convolutional Inverse Graphics Network, etc.
Read more : Deep Learning Past Present and Future
There are several reinforcement learning techniques, e.g., Monte Carlo Search, Temporal Difference (TD), and State-Action-Reward-State-Action (SARSA), which describe various aspects of the model-free policy evaluation and improvement process.
We are experimenting with various artificial neural network models for designing the balance control system. We have examined the neural network based reinforcement learning application of Deep Q-network (DQN), Double DQN (DDQN), Deep SARSA, N-step Q-learning, Deep Deterministic Policy-Gradient, Continuous DQN, Asynchronous Dueling network DQN, Prioritized Experience Replay, Asynchronous Advantage Actor-Critic, and Actor-Critic with Experience Replay.
Currently, we focused on Deep Q-Learning with Recurrent Neural Networks (DRQN) based artificial neural network architecture. Recurrent neural network architectures have been used in tasks dealing with longer term dependencies between data points. We examined several architectures for the DRQN. One of them is the use of a RNN on top of a DQN. It is showing promising results. The principles and algorithms are discussed in the book Compassionate Artificial Superintelligence AI 5.0. We observed DRQN is more effective for the data set we are using.
3. Geriatric fall-related injuries 2016