Leveraging AI/ML for Bionic Implants

Advances in technology have greatly benefited the field of?prosthetics?in the last few years. Today’s prosthetic limbs are made using space-age materials that provide increased durability and function. In addition, many prosthetics make use of bionic technology. These types of prosthetics are called myoelectric prosthetics.

Bionic implants refer to electronic or mechatronic parts that augment or restore physical functionality to a differently-abled person. The bionics industry has grown along four major application areas: vision, hearing, orthopedics and a small, motley group of implants that augment cardiac and neurological functions.

Visual neuro prosthesis, as vision bionics are sometimes called—are bioelectronic implants that restore functional vision to people suffering from partial or total blindness. Researchers and device manufacturers who are designing bionic eyes confront two important challenges: the complexity of mimicking retinal function and the consumer preference (and constraint) for miniature devices that can be implanted into the eye.?

Neuro prosthesis consists of a microelectronic array that is implanted in the retina, a wearable camera and an image processing unit. The camera, integrated into eyeglasses, captures images and transmits them to the portable processing unit, which wirelessly sends electrical signals to the implanted array. The array, in turn, converts these signals into electrical impulses that stimulate the retinal cells that connect to the optic nerve. It acts as the crucial link between the object and the optic nerve, bypassing the damaged photoreceptors. Machine learning is helping in scene simplification strategies such as highlighting visually salient information, substituting depth for intensity, object segmentation, combining saliency, depth and segmentation appropriately etc. Vision simulator algorithm takes downloaded versions of preprocessed images and interprets gray scale value of each pixel in a video frame as a current amplitude delivered to the simulated retinal implant.

Auditory Bionics, auditory brainstem implants and auditory midbrain implants are the three main classes of neuroprosthetic devices for people suffering from profound hearing loss. Auditory bionics create an artificial link between the source of sound and the brain—in this case, with a microelectronic array implanted either in the cochlea or the brain stem.?Auditory bionics is more mature as a technology than vision bionics, with a larger innovation ecosystem, more commercial products, and greater adoption globally.

Myoelectric prosthetics procedure for the implant requires one of the amputee’s peripheral nerves to be cut and stitched up to the muscle. The site heals, developing nerves and blood vessels over three months. Electrodes are then implanted into these sites, allowing a nerve signal from the residual muscles in the amputated limb to be recorded and passed on to a prosthetic hand in real time. Upon receiving the action potentials, the prosthetic amplifies the signal using a rechargeable battery. Detected signal can be used further into usable information to drive a control system. The detected signals are turned into movements using machine learning algorithms. Artificial Intelligence helps to identify motion commands for the control of a prosthetic arm by evidence accumulation.?It is able to accommodate inter-individual differences and requires little computing time in the pattern recognition. This allows more freedom and doesn’t require the person to perform frequent, strenuous muscle contractions. Reinforcement learning, a sub-field of machine learning, creates intelligent agents that decide how to perform a task by itself. The agent learns through a reward and penalty feedback, which reinforces its behavior to desirable results and is capable of performing cognitive tasks. The inclusion of AI technology in prosthetics?has helped thousands of amputees return to daily activities. While technologies that make bionic implants possible are still in their infancy stage, many bionic items already exist.

In the world of prosthetics, function is the key. Most amputees are constantly searching for the same level of functionality that they enjoyed before they lost their limb. With the introduction of artificial intelligence in prosthetic limbs, amputees are closer to their goals than ever before. Bionics having access to the relevant databases are capable of learning new things in a programmed manner which improves their performance.

Reinforcement learning is a breakthrough technology that brings model predictions to a cognitive level comparable to human actions; However based on our experience in implementing these models, it comes with its own challenges.

1.?????For a model to be empowered with knowledge comparable to that of an experienced doctor, it requires massive amounts of annotated data. This involves a lot of manual work where doctors have to look at years of data.

2.?????Model training and maintenance requires enormous computational power, thereby leading to high costs.

3.?????There is no direct input-output mappings in the labelled data sets. The model learns incrementally over a period and many combinations of inputs and outputs are possible. Therefore, reinforcement learning is a suitable solution for complex problems only.

4.?????Model testing is time and effort consuming as real world data and environments are difficult to simulate. When we try to validate our model results getting time from experienced doctors is not easy.?

5.?????Scaling neural network models in this case is also complex task for multiple hospitals, clinics and diagnostic centers.

Incorporating reinforcement learning for bionic implants into mainstream physician workflow is at its very early stages. This should be understood as a first step towards the ultimate goal of creating a retinal implant supported by deep learning–based image preprocessing. Such a device would require all processing to happen in real time at the edge. One solution could come in the form of low-power, low-latency neuromorphic hardware coupled with an event-based vision sensor. To meet the complexity of human reflexes, it is necessary to try out multiple algorithms and techniques. A combination of statistical and machine learning yields best results. Future iterations of this work may include end-to-end training of scene simplification strategies fitted to a specific implant technology or even an individual patient. Overall this work has the potential to drastically improve the utility of prosthetic vision for people blinded from retinal degenerative diseases. Reinforcement learning comes with huge potential to perform complex cognitive tasks and thus has the ability to innovate and transform healthcare, beyond expectations.

?About the Author

Raghuveeran Sowmyanarayanan is Artificial Intelligence & Analytics Leader for Healthcare at Cognizant and have been personally leading very large & complex Enterprise Data Lake & AI/ML implementations. He can be reached at [email protected]

#ai, #artificialintelligence, #machinelearning, #bionicimplants, #bioniceye, #bionicear, #bioniclimbs, #auditorybionics, #visualneuroprosthesis, #visionbionics, #myoelecticprosthetics, #reinforcementlearning, #neuralnetwork