Rise of AI Conference 2019 · Berlin

Deep Learning to Save Lives

Monishwaran Maheswaran · Computer Scientist, Inventor, Author, Mathematician, Quantum Algorithmist

At Rise of AI 2019 in Berlin, Monishwaran Maheswaran, a 12th-grade student who describes himself as a computer scientist, inventor and mathematician, presents two self-built deep learning projects aimed at healthcare. He traces his own path from teaching himself programming languages as a child to building an English-vocabulary app, an early chatbot, and, after encountering an accident involving a visually impaired person, a camera-and-earphone device that detects hazards such as crossroads and potholes and describes them in natural language. He argues deep learning's biggest impact will be in healthcare because of the volume of data the industry produces, and presents Project Health, a wristband that tracks tremor activity to detect Parkinson's disease years before symptoms appear, and Project Life, which analyzes smartphone behavioral data such as sleep, steps and web searches to detect emotional states linked to suicide risk. He states his own aim is to provide affordable, universal healthcare. A Q&A follows about funding, data sources and ethics approval for this kind of work.

Key takeaways

  • Maheswaran built a camera-and-earphone device that detects hazards like crossroads and potholes for visually impaired users after learning of an accident.
  • He argues healthcare will see deep learning's biggest impact because of the scale of data the industry produces.
  • His Project Health uses a wearable tremor-tracking band to detect Parkinson's disease years before visible symptoms, with over 92% accuracy.
  • His Project Life analyzes phone behavioral data such as sleep, steps and web searches to flag suicide risk and alert family members.
  • He says social projects like his do not require large capital, relying instead on mentors and institutional collaborations such as with NIMHANS in Bangalore.

Detailed notes

From childhood tinkering to a device for the visually impaired

Maheswaran says he started opening up home appliances around age three, learned three programming languages by around age eight and built an application that measurably improved his English test scores, and was introduced to machine learning around age 11 or 12, when he built a simple chatbot that could solve basic mathematical problems. He says his view of why AI matters changed in tenth grade after he learned of an accident involving a visually impaired person; noting that most visually impaired people rely on rudimentary aids such as guide sticks or guide dogs, he built a device combining a camera and an earphone that scans the environment for dangers like crossroads, potholes or traffic signals and communicates them to the user in natural language. He also mentions collaborating with MIT on content for an EdX course on computational thinking and data science, and working on a mathematics research project at Harvard in 11th grade.

Why deep learning, and why healthcare

Maheswaran describes deep learning as loosely inspired by how the human mind processes information and as especially effective at finding patterns and cause-and-effect relationships in data. He argues its largest impact will be in healthcare because of the sheer volume of data the industry generates, citing a figure of about 153 billion gigabytes produced in 2013, growing at roughly 48% a year, which he projects would reach about 2,314 billion gigabytes by 2020. He illustrates the value of larger datasets with a number sequence example, showing that more data points make it possible to correctly identify a pattern, in this case the Fibonacci-like sequence 1, 2, 3, 5, 8, that a shorter sequence would not reveal.

Project Health: detecting Parkinson's years before symptoms

Maheswaran describes building Project Health after his grandfather died following a late diagnosis of Parkinson's disease. He notes that neurological disorders like Parkinson's typically progress for ten to twelve years before symptoms such as hand tremors become noticeable, and that early tremors are often mistaken for something else, such as too much coffee. His device is a wearable arm band that measures tremor activity; as the disease progresses, tremor intensity increases, creating a measurable cause-and-effect relationship that he says the system can track with an accuracy above 92%, and that can also indicate whether a prescribed medication is working if tremors fail to subside.

Project Life: behavioral data and suicide risk

Maheswaran turns to mental health, citing close to 800,000 suicides a year worldwide and stating that by the end of his talk about 22 people would have died by suicide; he describes suicide as the second leading cause of death for people aged 15 to 29. He argues that unlike diseases with visible external symptoms, depression and suicide risk are hard to detect because a person may not outwardly show distress, and existing supports such as hotlines and mental health workers are limited in availability, cost and hours, while social stigma discourages people from seeking help. His response, Project Life, is a phone application that tracks behavioral data such as sleep patterns, step counts, calories burned and web searches, looking for correlations and sudden changes, for example a person who normally sleeps regularly and stops sleeping after losing a job, or web searches for methods of self-harm, in order to flag risk early and alert family members or loved ones.

Q&A highlights

Asked how socially oriented projects like his are funded, Maheswaran says they do not require large capital; he gathers information from mentors around the world and collaborates with institutions, citing a partnership with NIMHANS, a university hospital in Bangalore, where he interviewed patients about how to improve their care. Asked whether he uses social media data, he says he currently does not, both because platforms restrict access to user data through their APIs and because users may be reluctant to grant permissions; his current data points are sleep, calorie burn, step count and web searches. Asked about ethics committee approval and informed consent, typically required for healthcare research, he clarifies that his approach does not involve directly interviewing or monitoring diagnosed patients, but instead analyzes everyday behavioral patterns of people who have not sought help, developed in consultation with a university hospital's psychology department.

Quotes

Machines and humans are not two independent identities. We are codependent and we have co-evolved. Technology is not the problem, rather technology is the solution.

Monishwaran Maheswaran

Questions answered in this talk

How can a wearable device detect Parkinson's disease early?

According to Monishwaran Maheswaran at Rise of AI 2019, his Project Health uses an arm band that measures tremor activity; because Parkinson's symptoms typically appear ten to twelve years after disease onset, tracking tremor changes over time can flag the disease before it is otherwise noticed, with an accuracy he states as above 92%.

How can smartphone data help predict suicide risk?

Monishwaran Maheswaran explained at Rise of AI 2019 that his Project Life tracks behavioral data such as sleep patterns, step counts, calorie burn and web searches for sudden changes or correlations, such as a normally regular sleeper losing sleep after a job loss, in order to flag emotional distress and suicide risk and alert family or loved ones.

Why is depression harder to detect than physical illness, according to this talk?

At Rise of AI 2019, Monishwaran Maheswaran said depression and other mental health conditions often show no obvious external symptom, unlike a visible illness such as a common cold, and that people hiding distress combined with social stigma make it especially hard for others, and even for humans generally, to recognize these patterns.

Mentioned

deep learning, Project Health, Project Life, Parkinson's disease, suicide prevention

Keep exploring

More from Rise of AI Conference 2019

More on Health & Science

Transcript

Auto-generated transcript, not edited. · 3,357 words

Show the full transcript

good evening everyone my name is Mona chuan and I love to explore technology and more specifically I like to think of my programming computer programming as a superpower so I could create anything I want before I begin I would like to give my heartfelt gratitude to the rise of AI 2019 committee for inviting me and also bringing me over here thank you moving forward I would like to discuss a little bit little about myself first so that you guys can get familiar with me so I'm currently studying 12th grade so I would be going to university coming fall in coming August so I'd have to explore technology most specifically my passion is to build and innovate machines so that they could think and also solve our problems more creatively so I started exploring technology from a really young age so at about eight three eyes to open up home appliances you know just to look inside and see how they function and it was really interesting a couple of years later about age eight I

was twenty nearly three programming languages and I built an application that will help me improve my English vocabulary and it was so good that the English test scores were substantially increased from the past test when I was about you know 11 or 12 I was introduced to machine learning and I built my first AI a precaution and it was a simple chat board that we interact with me and also soul in a simple and mathematical problems but my perspective about machines and wide why we need to use machines or about AI or about human-computer interaction completely changed when I was in tenth grade so when I was in tenth grade I came across an incident that will completely change this perspective so I came across an accident of a visually impaired person and you know this prompted me to look into the lives of the visually impaired and most visually impaired people use rudimentary instruments to walk around their environment for example guiding sticks or you know you have guided dogs that are trained to

train to move the visually impaired people around their environment so I do believe this was one other reason for the accident itself so I taught myself why not use artificial intelligence you know to help guide the visually impaired so I came with the solution I came up with the solution basically the device consists of a camera and a year phone the camera scan the environment and look for any dangers like crossroads or potholes or even traffic signals and signal that back to the user using natural language itself I also collaborated with several universities all over the world or in fact a really good at a course that I would recommend for or for beginners is by MIT X on EDX called introduction to computational thinking and data science I did collaborate with MIT to provide content for this course and it has been tailored well so I'm pretty sure you would enjoy it when I was in 11th grade I I got a chance to go to Harvard University and work on a mathematics research

project so in the end I did you know because mathematics is the language of nature and if you see deep learning machine learning artificial incidents if you strip down everything mathematics is the core it is behind everything so I do suggest you to you know look look at math and deep learning as well I have put forward two innovative projects that I had built a couple of while back for this conference if you notice the the universe the human brain cells and the internet all have something in common they are all made of similar structures and I do believe decoding the human boomin brain is a key to solving world problems from poverty to climate control and this decoding of this human mind is possible through machines see computers are really efficient at finding patterns we can use computers to find the cause and effect relationship of any ting in the data set and we can do that using deep learning technology so a lot of other speakers did give a good introduction to deep

learning so I'm just gonna keep it a little bit vague so deep learning is basically inspired by the workings of the human mind and we are trying to kind of you know assimilate how the human mind processes information on a computer right now millions of devices used to be planning technology in fact the very application that I used to come to this event uses deep learning to find fastest route to this conference all the deep learning would have immense impact in the field of vast variety of fields I believe its biggest impact would be in the field of healthcare and why do I say this mainly because of the huge amounts of data the healthcare industry produces alone in 2013 nearly 153 billion gigabytes of data were produced as this astronomical rate of 48 percent annual growth in 2020 we would have 23 14 billion gigabytes of data that's huge amounts of data so what could we do with this data there's so much data but you know because the since healthcare industry produce huge

amounts of data we can use this data to find and find correlations in this data for example symptoms and why why does the symptom appear when this disease is progressive and we can do that using computers as I said earlier computers a really efficient of finding patterns to illustrate this further let's have an example let's say you have a number 1 2 3 - and - how many of you think the next two digits would be 4 & 5 okay so I see more than 50% but the actual answer is 1 2 3 5 8 you see even if I had given one digit say 5 Yura gets it correctly that it's 1 2 3 5 8 so the greater the amount of data we have the greater the amount you know agree the accuracy of finding that pattern and this project called Project health that I would like to introduce you to is basically on the on the on the idea that you know you could could diagnose diseases early even before the symptoms

appear so my grandfather was suffering from Parkinson's so and his death was due to the late diagnosis of this disorder so I told myself is there any way to decrease the diagnosis time across the world we have access to healthcare that's very limited and also expensive so even these two points have to be kept in mind while constructing this kind of instrument another problem with neurological disorders is the slow progressive nature of these diseases most neurological disease progress at a slow pace so for say the actual onset of the disease would be ten to twelve years prior to the actual symptom appearance so you have 10 to 12 years of time gap when the when I see a person for example is suffering from Parkinson's so when he's in the late 30s he might he might get some tremors in his hand so but he might you know mistake it for you know he has taken too much coffee but it was actually the disease that's showing the symptom so after a couple of years

it gets worse but and after 10 or 12 years it moves to a point where it's really hard to control it so if the person would have known that he is suffering from Parkinson's 10 to 12 years prior to the actual onset or the symptoms it could have been curbed and probably prolong his life so the idea is decreasing diagnosis time and also provide better patient care and feedback so just just like I said let's go back to the Parkinson's example so all the person has to wear is a band around this arm and it measures its tremor activity the amount of movement in his hand so if if the disease is progressing the tremors would get worse so you this clear cause and effect relationship between the symptom appearance and the disease greater the amount of disease progression greater the tremor also it could also provide in a feedback see a doctor does provide medication to him and if the tremors don't subside even know for sure that this medicine is not having it's

not really effective for at this point of time it can do all this with an accuracy of greater than 92% another big project that I believe would have an impact in the lives of people it's project life right now a big problem exists in the field of mental health care I'm talking about suicides nearly 800,000 people come in suicides every year it burns me as a sadist but by the end of my talk nearly 22 people would have taken their lives more was suicide is the second leading cause of death for 15 to 29 year olds as any youngsters so why is suicide so hot curve is it invisible diseases say like common cold could be easily detected because they issue an external symptom that's very obvious right you have runny nose so if a person comes to you with a rhino's you even know for sure that this person is suffering from common cold but that's not the case for major depressive disorders like depression right most most diseases like this they don't show

their actual manifestation of externally a person who is cheerful outside may not actually be happy inside he might be just burying a mask or facade a person for example who is happy might sleep for 10 hours but when he is in a different kind of emotion Street he might not sleep for 10 hours you might see 4 4 he might 3 4 - or not sleep at all major problem in detecting these kind of neurological mental health problems is that they are not easily detectable and we humans are not really good at recognizing these patterns as well but machines are really good at recognizing these patterns the parent fiying geniuses so why not use a machine to detect the emotional state of a person right now we do have solutions to curb suicides for example we have hotlines and mental health workers but in our world we don't have some of these hotlines don't run 24/7 and some of these hotlines you know you need a call at a specific time period on a specific

day to receive emotional support and also we don't have much mental health workers and what if the person doesn't ask for help to answer the question we need to answer ourselves why won't the person ask for help one of the biggest problem for a person in seeking help is the social stigma surrounding mental health disorders more access to mental health care is expensive so how do we curve these problems the answer lies with machines as I said earlier computers are really efficient at finding patterns and we humans exhibit behavioral changes outside when there is change in our mind if a person is sad he might not show it externally but he might you know show it indirectly for example might not sleep that much or you might not go for a walk we carry devices that record all this human behavioral data for example smart watches smart phones but most of this data is not he used effectively so we can track and analyze these changes and find correlations between the cause and effect so

if something's bothering a person and he's not willing to share that information we can know for sure that this person is in this emotional state and probably need to talk to him or even providing support and we can do that using deep learning technology the V human brain sauce problem is very efficient and we trying to kind of simulate that kind of efficiency on a computer and that is very challenging since we carry these devices see for example there is this person and whenever he's happy he you know he goes for a walk to his job which is nearby it is home but when he is you know in a different emotion Stacy pres he might not walk he might take a cab to this change in the amount of calorie burned for the person so that that one data point is changed right so there is a correlation between his emotion state and the amount calorie burned so we can analyze all this behavioral data to predict the emotion state you find these changes

and predict the problem even before it becomes worse since we carried these devices everywhere it's easier for us to be to build an application around these data points and we can do that using deep learning technology so it's basically like an application that runs on your phone that collects information about sleep data and your amount of Calabro and how much steps you're taking the kind of web searches you're doing and based on that it can predict what kind of a message emotional state you're n begin let's go back to sleep time example so a person who regularly sleeps at 10 p.m. and gets up at 8 a.m. whenever he does it regularly he or she does that regularly and that is completely changed or there's a sudden change in this behavior when there is a difference in his emotions let's say he gets fired from his job so you might not sleep that much he might have thoughts about how do I pay my mortgage how do I pay my loan how do i how

do I sustain so he goes on in his mind and he might not sleep much he might not sleep at all and this would get worse and at some point he might snap we can also use behavioral data like workout routine and verb suggests a person who's on the verge of committing suicide might actually search on the web for most efficient way to die so all this data included we could predict with the reasonable accuracy whether this person is going to commit suicide or not we can also provide help and also alert family members and loved ones see machines and humans are not two independent identities we have we are codependent and we have co-evolved technology is not the problem rather technology is the solution my aim is to provide affordable healthcare and that is one of the reasons why I do what I do my aim is to provide universal health care to everyone all over the world thank you thank you very much is there some Q&A you would like to ask those

great presentations thank you a lot of the social projects it's it's hard to found it like you know there is I'm not sure whether there is money in it how you and I believe it social projects are very needed like who help humans especially the depression examples at the show so how do you found how this kind of projects can be founded okay so you mean how can these kind you mean funded how do you get the money for it okay so the thing is you don't really need capital to actually do something that's one thing so what I do is I gather information I get a lot of information I have mentors all over the world and they guide me and I collaborate with companies all over the world so you don't need capital you can you can you know you can do you can get information from them you can say that you know a 1-1 a good example is I collaborated with NIMH hands so it's a it's a University Hospital in

in Bangalore and and you can you can interview patients you can you can ask them how could I improve your life you know it's all about conveying the need for this kind of projects and you don't need exactly capital if you mean you mean like you need service and everything but it's again you know you can you can do that that's a small budget right so can it can come out of your product so do you also use social media like the acting of the person on Facebook for example okay so the thing is uh you know it's all depends on the user so he might be a little more you know skeptical about giving permission to each and every object app on the phone right so right now I haven't scaled it up to social media but all it can do is it can search for web web results of web searches because you know some of these social media companies you know they don't really in a cell of data and you know

some of these api's are not present where you can access that kind of user data from from those kind of companies social media companies but you know web searches are really you know one way you know you could get information from from the user directly so right now or all kind of data points that I have is a sleep data then you have the calendar calorie brand and then how many steps you have taken and web searches that's up now the first question that was asked so as far as I know for any healthcare research you need article committee approval yeah and there are a lot of checks and balances in places and they do require funding I I actually don't know of any research that doesn't require funding because you need access to the data and as far as I know you cannot directly walk up to the patient and ask how they are doing for that also you need informed consent and right exactly so I didn't get an idea from your talk

of how you actually aim to do what you presented so can you give us sure sure okay let me rewind back okay so what how did I go about this project let me tell you so what I did was once I got the idea right so I reached out to several University Hospitals and I collaborated I mean I talked to Sarah professor and what their idea about suicide was and you know a psychology department and you know once everything you receive right then you can you know you can you you don't need to it I do believe that you need you do need to go to ethics department whether you need you know specific patient data but this this this this kind of information that you get you don't need really need to you know it's kind of like it's more decentralized I would say because these kind of you know mood behaviors right you it's kind of an external that happens every day so and every time so you it's kind of like a

24/7 process so you can't really go an interview patient of for you know whether you say he's depressed or not right she can't go to a patient who's who's who's actually suffering from depression and you know monitor his thing then then you kind of have an imbalance because this Bush person has actually come out and he has told that he is suffering from depression and he is seeking help but what I'm concentrating right now is a patients who don't you know people who don't you know seek help so you find correlations between some of these you know information and data points and then you know construct something that's value out of that

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.