11 June 2021
Over the last 150 years, our life expectancy has grown, from 40 years in 1850 to over 90 years today in some countries. This can be attributed to advances in medical science, improvements in public health, and equitable access to healthcare, especially for maternal and infant care.
What will the future hold for our world? Will we be overwhelmed by a ‘silver tsunami’ of retirees with poor health, or will we use the latest research findings to rejuvenate the elderly and extend their lifespan?
Our Longevity Futures is a show where I, Chris Curwen, speak to scientists, engineers, entrepreneurs, doctors, politicians, and community activists who are giving the world the hope that we can all live longer and better, and improve our health.
In today’s episode of Our Longevity Futures, we are delighted to speak to Nikhil Yadala, co-founder and CEO of Healome.
Healome is a service allowing users to store their healthcare information and bring structure to the unstructured data. Using artificial intelligence and machine-learning techniques, Nikhil and his team anonymise your data and enable healthcare practitioners and researchers to provide insights to enhance you and your family’s well being.
In this talk with Nikhil, we will discuss the impact of artificial intelligence on the future of healthcare, and how Healome is aiming to be the centre for personal health tracking and insights, in the future.
Here are some of the highlights for my conversation with Nikhil:
Chris: Could you tell us about Healome and the work you carry out there?
Nikhil: Healome is a direct to consumer application that let’s users take pictures of their healthcare prescriptions and diagnostic test reports on-devices, we digitize how various blood biomarkers are changing through time and we provide insights through various AI models into their health and detect any abnormalities in their health profile. This is the front ending first part of it.
And at the same time, Healome is also a platform for drug discovery companies and academic researchers to get insights into, various pathologies and dedicated deliver care to specific users based on the disease conditions that they are exposed to.
Chris: What kind of problems is Healome attempting to solve?
Nikhil: First and foremost Healome is a platform that lets users organize their healthcare information. In most of the developing countries, there is no concept of electronic healthcare record or EMR system.
People carry the healthcare prescriptions and physical hard copies in files. So if you go to a doctor, he writes something with a pen on a piece of paper and gives it to you. And by the time someone reaches the age of 40, they just have a bunch of papers lying around in order to show their healthcare information to someone that if somebody falls on conscious on drawer, and if he’s admitted to an ICU ward, there is literally no way for the doctor to learn about their previous health condition to get clues as to, why have they fallen unconscious?
So through Healome you can look at all of your information through your mobile application and search through them and track various records also, that’s one thing. And secondly, by practically looking at various signals in the blood biomarkers and other specific health information that we get from the users, so we prune for those signals to detect abnormal patterns at a very early stage so that we can hopefully change the trajectory of their disease progression.
Chris: What is Healome’s plan for obtaining the high quality data needed to best build and train your AI and ML algorithms?
Nikhil: It is definitely very important to have a complete health profile to do anything interesting in health care sector, because with incomplete information, we might fall into risk of delivering wrong results.
Such complete information need not, typically be necessary in typical applications, look at recommending your next youtube video or finding the third result in your search history with respect to your search history. We have two approaches in getting such high quality data.
One of them is a B to C (business to consumer) approach where users provide the data directly to us. They comprehensively take the pictures of their records and we accumulate and digitize them on device. Second, we create an incentive mechanisms for the users and also certain features like being able to track costs, and being able to track the medical expenditure, getting notified as to when to take medicines. So all of this information, which leads to some metadata helps us annotate the actual data and this sort of creates a complete health profile of the user rather than just having the blood test.
The other approach we have is it B to B (business to business) approach where we partner with diagnostic test centres who already store the diagnostic test reports with them. And we directly take them and we directly give that user an ability to search for the lab reports from our application, which would be fetched from the diagnostic test centres. So in this way, we have access to a large volume of data. It is more representative of the various segments and demographics of a population and diseases. It it helps us to actually have good quality information with us to train any learning algorithms on top of that.
Chris: How are you managing to keep down the costs and time taken to retrain your neural networks in the presence of new data?
Nikhil: One of the important things to make sure that the data is continuously coming in for us to continuously keep iterating is to have proper incentive mechanisms in place for each and every stakeholder that is over there.
So for users, they always need to be able to organize their health information. Second definitely keep becoming better with time, as we have access to more longitudinal data from any given person, because now we have more things to look into their history. So definitely it’s in the interest of the users to keep using applications in order to get better results through time.
And it’s not just that, as a user, the incentives that they get with respect to using the application also include having discounts with our partner diagnostic centers when they get more tests done, to get more insights into that health profile. So it sort of encourages them to get more preventative care and it helps us to, keep us connected. And second stakeholder is of course the academic research labs or the drug discovery companies. Most of the drug discovery companies are working in that next gen precision non personalized medicine. Most of these are in the developed countries or the North American or the European regions are the kinds of datasets that they get access to are typically the Datasets from their own countries, which is not a good representative of all the population in the world.
So what happens is when they’re designing personalised treatment plans, these treatment plans would not consider the diet patterns, lifestyle patterns, stress patterns of people across various regions of the world. So these treatments cannot be immediately used to everybody else. Particularly when companies are trying to deploy ML driven systems into their decision making of it is in their best interests to have very good dataset that is representative of all these diverse things, so that their models are not biased.
So being a direct to consumer application we will be in a position to let them get access to as diverse data as possible. Like just by the number of population in the Asian countries, even for any rare genetic disorder, we would help a lot of people with the disease.
All of this data that’s getting generated is neither stored anywhere, or taken insights from. It’s just completely, I would not say it’s even decentralised because it’s not even digitised currently. So in this way I believe the drug discovery companies have incentives to actually interact with the users and use then benefit the users.
I think that’s how we try to go about it. And second one is avoid the avoid the ML systems itself. As I said, it’s very important to keep tuning our model from time to time based on the feedback that we are getting from the users and also from the doctors. And we need to keep improving on the metrics.
In general, the algorithms that we typically use there is a possibility of continuous learning. So once a model is trained if you have some more data points to append to it, we don’t need to continuously discard the model. We take the chain model as an initialisation point and try to again, retrain it with the new dataset and get a newer model now, which is more accurate than the previous model.
It’s very important to have this awareness a priority while building the tech stack because the data pipelines that are built need to be in such a way that such a training mechanisms are easy to perform by the data science team. So they’re trying to take care of that.
Chris: What are three things that you believe are inevitable for the future or AI, longevity and healthcare?
Nikhil: First and foremost thing there is going to be very good improvements with respect to accelerating the pipeline of drug discovery by using AI and machine learning algorithms. If you think about the problem of drug discovery, it’s basically a problem of such in chemical space.
We want to design certain molecules with certain characteristics, with certain physical and chemical properties. These are just constraints. So when we can draw parallels to the problem of such like how such Indians tackle it, so you get a queery, which is just four to five words often, and that’s not even to medically correct most often. Within a few milliseconds, the search engines are able to collate the best ten articles for you by searching all over the internet from petabytes of the internet index. All of this is happening within few millisecond searching through the entire internet, which means that there have already been software systems developed to the extent that they can do this inference very quickly. And also with very good accuracy of the results.
Now an LP is of course, something that so it’s basically in natural language and we know the syntax of natural language. So it’s definitely a much more easier problem than understanding the language of the body and the genetics. And all this, but I think all the learnings that as a field, we have accumulated in delivering software systems in the problem of search would eventually be applied to the problem of drug discovery as well.
Another thing with respect to healthcare, it’s bit farfetched, but with all that is happening these days, I think that going forward, it will be a common thing where people maintain ledgers of their healthcare data, and the information of both how a piece of info about how that piece of data has flown from point A to point B would be minted onto a blockchain in a decentralised manner. It basically gives a proof of how on which all parties have access to a piece of data. And also it gives the ownership of the data to the users themselves. So I think it’s eventually going to go there.
We already have a lot of applications on a hyper-ledger’s both like news articles and also with the recent NFT, maintain non fungible tokens for so I think going forward the next step to apply all of these technologies is to healthcare.
Now I also like one other thing I Think is going to happen is we are going to have a much better understanding of how to tackle aging as a disease. Currently the biomarkers that track to measure, like even the methylation clocks or those which look good are even more deeper things., they’re not easily accessible to everyone. I think going forward, we’ll be in a position to at least approximately measure how good is that an intervention is doing with respect to increasing the health span or reversing the age?
One thing I always keep thinking about being from a computer science and technical background: we are going to reach longevity, escape velocity in our lifetimes, most probably. When that happens, most of the computer systems today are all about making sure that it is speed and making sure that it has very less latency in solving. To the extent that we are even often find with deploying software systems, which are just approximate solutions to that exact solution, but the exact problem, just because it can be computed in a much faster way, on a distributed way and with less latency.
I think that then we reach the longevity escape velocity. I’m not just talking in terms of healthcare in general, broadly about software systems and utilization of AI itself. We need not care much about speed and agility of development in this companies, which use AI. Currently every company or startup that uses machine learning they have all these timelines set up who gave you to deliver to certain accuracy, with the problem that you’re tackling by this month, that forces them to be more agile and that forces them to have more restrictions in the space of models that they can actually experiment with.
When speed, and time is no more problem than speed is also no a problem. I’m more interested to look how software systems and companies in general would evolve to that situation.
Chris: What does the roadmap for Healome look like over the next 5/10 years?
Nikhil: So currently we are working in a phased approach.
In the first one, the goal is really to help users with preventative care. So people often get their blood tests done and all these biomarkers, that are not only the blood biomarkers, but also other digital biomarkers that are being collected.
All of this, shouldn’t be used practically to identify onset of any abnormality, so that people can priori get treated and prevent the disease instead of trying to cure that disease. It should be as simple and as common as opening basically using some social media app, it is so simple to just open the app and write something.
It would be great to have access to preventative care as simple as that, you just keep learning about what’s happening with your body. Simple suggestions are with respect to diet patterns, or lifestyle changes. Most of the things can be prevented if they are understood at an early stage and also identifying that trigger points to certain conditions would be easier. So this is the phase one goal.
Going forward it should be easy for academic research labs and drug discovery companies to request a specific kind of information. Let’s say people age 23 to 30 with a specific type of cancer and co-morbidities, how has a specific treatment plan been affecting their overall health and biomarkers? This longitudinal data and also the diverse data would be used by companies to basically in their research to enhance their insights into the disease, pathology and progression.
Also there is one more thing that I currently are trying to explore. How do we incentivise people? Everybody agrees that we need to exercise and we need to take care of our health. But not many are in a position to do it or do it in general. So we’re trying to also explore other incentive mechanisms which you haven’t implemented yet. Getting to know, your biological age and being able to boast about it.
We really appreciate Nikhil taking the time out of his busy schedule to come and talk to us. So a massive thank you from Chris and everyone on the Gowing life team.
Please join us for future episodes and tell your friends and family about us!