The questions and solutions beneath have been edited for clarity and length.
The period of customized medicine at all times seems to be just around the corner but by no means fairly right here for the plenty. Can you explain what “personalized medication” means and what’s wanted to make it a reality?
Personalized medicine, also recognized as precision medication, refers to medical therapies, predictions, or even preventions which would possibly be carefully tailored to a selected particular person. It has been round for quite a while but not very broadly. An early example of precision medication that has been around for almost twenty years is breast most cancers therapy in which patients receive chemotherapy based mostly on sure gene expression. In fact, genotyping cancer cells to discover out whether or not a sure chemotherapy works is amongst the most advanced examples of personalized medication at present.
But the promise of personalized drugs has been a lot bigger and broader. The hope is to use not only genetic information but additionally to incorporate data from a wide selection of sources: completely different biomarkers, the patient’s historical past, pictures, and information from implantable units that may repeatedly monitor sufferers. The hope is to make use of all this information and apply it to a broader set of circumstances past most cancers and chemotherapy.
What are the tools we need to get there? There are, in fact, things on the medical facet. But on the engineering and data science side, we need to generate very high-quality data that can be utilized for many, many people, maybe over years. We want gadgets and imaging methods that can generate this knowledge, and we have to embody many individuals in our information assortment. Finally, we’d like algorithms and data science approaches that can handle the complexity of this knowledge.
How will artificial intelligence or machine learning play a task in this?
As I talked about, the hope is that we collect information from many various sources and from many different individuals. People of different ages, genders—you name it. We then have to correlate this complex set of information to completely different circumstances and likewise, perhaps, predict whether or not a patient is at greater threat. This is a place the place machine studying and AI can play an enormous role.
Another space in which AI and machine studying could be very useful for precision medication includes embedding machine learning into wearable and implantable gadgets. Very quickly we may prepare algorithms to the baseline of an individual patient and likewise to different sufferers. We might then use the gadgets to foretell problems or detect dangerous conditions.
Can you give us a brief overview of your analysis and whenever you began incorporating AI methods?
Generally, in my lab, we’re interested in biomedical devices—for instance, sensors and drug delivery techniques or navigation tools for precision surgical procedures. We build tiny devices that may be implanted. They’re usually wi-fi and battery-less, and we attempt to make them as noninvasive as attainable. I started working on this area very soon after I arrived at Caltech in 2007, however I all the time wanted to work within the domain of neural interfaces.
One of the reasons was that one of my sisters suffers from epilepsy. From very early on, I observed that this was tough for her. Drugs had unwanted effects and typically were not successful in controlling her situation. That was one thing very near my heart, and I at all times wanted to work in this domain. I was very lucky, about six years ago, that I might begin engaged on early seizure detection, or prediction, for patients who’ve epilepsy.
About one-third of patients with epilepsy can’t reap the advantages of medication. There is a subset of this group for whom, if we predict the seizure early enough, deep brain stimulation can cease it. That was certainly one of my first tasks that involved large amounts of data. We had been taking a glance at neural knowledge from electrodes under the cranium, intracranial electrodes. The aim was to predict a seizure based on the signature of the neural knowledge. As you probably can imagine, this may be very affected person dependent. Fortunately, there are large publicly out there data sets on epilepsy. First, we constructed systems that concerned conventional sign processing, but we shortly realized that we wanted an adaptive system that would be taught from the data units and turn out to be more personalised for each affected person.
That was a turning level. We began utilizing studying algorithms, and we may efficiently present that we could practice a network to very reliably predict the seizure early sufficient to stop it.
This was the beginning of utilizing AI in my group, however now we have extended that effort to new brain-machine interfaces. We are additionally adding machine studying algorithms for heart monitoring to wearable gadgets.
What issues are you hoping to tackle with brain-machine interfaces?
One of the tasks we’re very enthusiastic about is in collaboration with Professor Richard Andersen [James G. Boswell Professor of Neuroscience and director and leadership chair of the T&C Chen Brain-Machine Interface Center], who has pioneered brain-machine interfaces and has been working in this area for many years. For sufferers who’ve a spinal cord injury, the objective is to implant electrodes in certain areas of the brain and then use decoding algorithms to predict their intention as they give thought to shifting a robotic arm or transferring a cursor on a screen.
After engaged on the seizure prediction techniques and when the Chen Institute for Neuroscience started at Caltech, there was an opportunity to start collaborating with Richard. In discussions with him, we realized that there is not any small implantable device right now. Patients are linked to massive wires and computer systems, and it’s a system that’s not mobile. They have to take a seat in the clinic to use it.
So, one goal was to see if we may implement the system in a tiny chip. Another was to enhance its performance. The reliability and the efficiency of the system has a lot more room to improve. We’ve been working with Richard and his group for nearly three years now, and we have made lots of progress.
How does the brain-machine interface work?
We have penetrating electrodes that, remarkably, can pick up small electric indicators as neurons fire. Neurons are continuously firing. We have two to three arrays that each have roughly a hundred electrodes implanted within the mind to collect that knowledge. The analog data then goes by way of amplifiers to a module that will get digitized, and the info may be very noisy. It’s a huge quantity of information, so the next step is plenty of pre-processing and filtering of the information. Then we go to a feature-extraction unit, in which we use a threshold-crossing approach to determine whether a neuron has fired. In other words, if neural exercise goes above a given threshold, we depend it as one spike (like a spike on a graph).
We then measure what quantity of spikes we have in a given period. That’s type of the traditional means of defining the characteristic that maps the measurements to neural exercise. This firing rate on this given amount of time is distributed to a decoder. The decoder decides what direction, for example, the patient wants to maneuver a cursor.
How has your approach to tackling this downside evolved over time?
It has occurred very naturally. It has by no means been because everyone’s doing AI or machine studying. Step by step, we’ve gotten to a point the place our group realized, “OK, now the most effective strategy is to make use of learning algorithms.”
Where do you see the sphere stepping into 10 or 20 years? Where will the technology be?
There continues to be so much to be done. We need higher electrodes which are less invasive. There can additionally be an analog front-end system that we mainly have to amplify, filter, and digitize the signals. This is still not at a place the place we will integrate everything and create a really low-power energy-efficient system. There’s a lot of room left to miniaturize things, to make the technology less invasive, to make it extra robust towards micro movements, encapsulation, and electrode degradation.
Another area the place extra work is needed is training. At the beginning of each session, plenty of brain-machine interface systems need coaching. The affected person needs to train the system so it could regulate to them. Our aim is to try and provide you with systems that may basically be taught in real time—using online studying or unsupervised learning—to modify themselves routinely as patients use the device. That’s an enormous challenge.
I additionally assume it’s fairly important is to create methods which are low cost and that people can use everywhere for a big selection of tasks.
What do you suppose are essentially the most troublesome tasks for AI in coping with medical applications?
I’m not a physician, however I have talked to many physicians about this. In drugs, the stakes are very high. It’s life and dying, and we want very dependable algorithms if we wish to depend on AI. As is the case for medical devices or drugs that require FDA approvals, algorithms are also starting to have requirements from the FDA.
There are sure purposes, such as brain-machine interfaces, during which a mistake is not an enormous deal. But if you need to really transfer towards personalised drugs and get predictions or ideas from AI, we want to prove that the system is reliable. For that, as I mentioned before, we’d like large knowledge units from many, many individuals. We want studies to prove that the system is strong and giving us useful data.
In what different areas may artificial intelligence–powered devices enhance treatment?
There are areas the place we imagine it will make an enormous distinction. I have a collaboration with Professor Wei Gao [assistant professor of medical engineering, Heritage Medical Research Institute Investigator, and Ronald and JoAnne Willens Scholar], who’s using variable sensors and measuring different biomarkers in sweat. He is attempting to combine completely different biomarkers to judge things that are associated to mental health, despair, nervousness. These are areas that generally folks with severe cases have problem explaining. Or perhaps they do not appear to be as open about it, or tracking it’s troublesome for them. So, positively for points associated to psychological health, this is ready to be important.
How keen have college students been to get involved on this research?
I find that students are actually fascinated in this domain as a end result of it’s an area that they will relate to. They see the influence. In reality, I began as a more traditional electrical engineer at Caltech. I worked on high-speed data communication systems, computing methods, and so on. I truly have continued that analysis, however gradually I get increasingly students who are thinking about medical devices. They’re so passionate as a result of in addition they find multidisciplinary analysis rewarding. They learn from different groups in chemistry, biology, neuroscience. So far, it has been unbelievable.