Medicosnext
Suresh Manandhar, PhD
CEO and Chief Scientist
Wiseyak
AI need not be 100% accurate but we need AI systems to have 0% (or close) false negatives.
In Wiseyak we are focused on data quality and data interoperability… For this reason, we are trying to obtain accurate medical data and ensure that doctors and healthcare workers can input the data with ease.
My suggestion to aspiring AI scientists is to become independent thinkers… Understanding the fundamental concepts is important. Do not believe everything you read in papers. Be critical and cultivate your own judgment.
Language models possess this generalization capacity and it can multitask. ChatGPT is not magic; it’s just that we have trained them for a varied number of tasks.
Prof. Suresh Manandhar, CEO and Chief Scientist of Wiseyak, started his career in computing as a Systems Engineer with the National Computer Centre in Singha Durbar in 1983 where he worked primarily on low level assembly coding on mainframes. He also worked on designing a Nepali language encoding system suitable for data processing while prevailing Nepali language computer fonts and encoding systems were designed for document processing applications. This encoding system was designed long before Unicode that incorporates similar principles.
He completed his Master’s in Artificial Intelligence (AI) from the University of Essex, England, in 1987. After completing his Masters, he returned to Nepal and co-founded PCS (Professional Computer Systems) and worked on implementing Nepali language data processing. While in PCS, he implemented Nepali language keyboard into MSDOS. “The government used our software during the national census,” he shares.
To further enhance his understanding of AI, he chose to do his doctorate on the subject and went to the University of Edinburgh. He completed his PhD in AI, especially focusing on the use of AI to process the syntax and semantics of different languages. He completed his PhD in 1993, graduating in 1994. From 1992, he was employed as a Research Fellow at the Human Communication Research Centre (University of Edinburgh) where he worked on improving the work conducted during his PhD.
AI has tremendous potential within healthcare… AI systems can be employed to provide a second opinion to help the doctor.
He took up a permanent faculty position at the University of York in 1996 and joined the AI Research Group. He worked on a number of branches of machine learning including symbolic, statistical and deep learning. He supervised over 20 PhD students and published over 120 research papers and articles. He worked at University of York for 23 years becoming the Head of the AI Research Group before returning to Nepal in 2019.
While at the University of York, around 2000, he played a pivotal role in co-founding Lexicle. Lexicle built the first virtual character embodied conversational agent that could take typed questions and respond using voice and body language. “It was like the older version of ChatGPT that we have at present,” he mentions. The Lexicle system was featured in the BBC Click program around 2002 and was used by FirstDirect Bank (part of HSBC Bank). “We were too early to market. Even the internet was new at that time. Our system worked with 8K modems. And despite 2 rounds of funding, we did not succeed in making the company profitable”, he says.
In this edition, we spoke to Prof. Manandhar on various aspects related to artificial intelligence and also about the education system in Nepal to get an insight on how we can further improve the quality of education in the country. Excerpts:
From your personal experiences how would you gauge the development of artificial intelligence in the days to come in Nepal?
I have been involved in the field of artificial intelligence (AI) since my Master’s as well as during my PhD and over the years I have seen that AI as a discipline itself has changed a lot. Even though I have been working in this area for so long I myself have had to learn and reinvent myself. What has helped me is that I have been interested in science and technology since childhood and I am always willing to learn and explore, so I have been able to adapt to the changing world of AI. Even now I spend a couple of hours a day in the evenings on research, which is important if I want to stay updated with the rapid changes that are taking place with AI.
I completed my PhD on knowledge graphs and mathematical consistency of such formalisms. This is still very relevant today for translating from unstructured NLP to structured knowledge graphs. During my time at the University of York I worked on symbolic machine learning (also known as inductive logic programming) and then I moved towards statistical and Bayesian machine learning. Then, subsequently, I worked on deep learning which is the latest technological trend. Like all disciplines, machine learning is also going through different phases and evolving rapidly. And we see that past research is being revisited. For example, there is an emerging strand of research on extending deep learning models to do symbolic reasoning.
There is a tremendous interest within our young generation of engineers towards machine learning. And, there is ample opportunity for the young engineers to either apply existing machine learning tools or to become AI scientists. We already see this trend happening with growing adoption of AI technologies within the IT industry in Nepal. The IT industry together with AI technologies will become a dominant driver for the economic growth of Nepal.
We often hear about deep learning these days. For a layman how would you describe it?
We can now predict the price of something for tomorrow with the help of deep learning. It could be the price of the US dollar, or the price of the stock market, or even tomatoes. To predict that price, typically we study the factors that affect the price of these goods. There could be other factors also that could affect the price but we do not consider them. For example, the truck drivers may strike and the price of the commodity might rise but we do not consider such things. We just focus on the primary and main factors to create a prediction model. Next, we focus on the combinations which need to be brought into attention. In conventional machine learning we inform the model that these are the things which need to be considered. We create a mathematical model that employs all the variables including combinations. Then we fit data into this equation. This is standard statistical machine learning. In deep learning we do not have to do the mathematical modeling part. We just need the input and output as the equation part will be done by the machine itself. The main advantage of deep learning is that the domain knowledge needed is relatively less.
Complicated problems can be solved with limited human engineering. For instance, by just showing a picture we can ask them to identify whether it is a dog, a horse, or a human. By just using the picture of the skin we can detect whether a person has skin cancer or by using the X-ray image of the chest we can also detect various infections and diseases. Of course, training data is required but human involvement is reduced and complicated tasks can be done easily.
How does ChatGPT work? For a layman how would you describe it?
ChatGPT is a type of language model also commonly known as a large language model (LLM). A language model is typically trained to predict the next word in a sentence, like a fill in the blank question where the AI system learns to fill in the blank. The context could be, the first few words of a sentence, or, it could be a whole sentence in a different language. This general approach can be employed for many NLP tasks such as machine translation from one language to another. Typically, we have just one model for one task. For example, we will have one translation system that translates from English to Nepali. However, we can also combine multiple tasks into a single model by providing an instruction or a prompt. So, we can have a single model that can understand, ‘Translate into Nepali, the sentence :: Mount Everest is the tallest mountain’ and ‘Translate into French, the sentence :: Mount Everest is the tallest mountain’. By training a single model on hundreds of different tasks, the deep learning model learns to generalize to unseen tasks. So, it may be able to understand, ‘Translate into Bengali the sentence :: Eiffel tower is the tallest tower in France’ even though we never trained the model on translating from English to Bengali (but the model was trained on some other Bengali tasks e.g. translating from Bengali to Nepali). Similarly, if the model is trained to write letters to your bank then it will be able to write a letter to the principal saying that you are sick, even though the model was never trained on that specific task. This is what we call generalization capacity. Language models possess this generalization capacity and it can multitask. ChatGPT is not magic; it’s just that we have trained them for a varied number of tasks. We trained it for 100s of tasks and based on this it was able to do 1,000s of other tasks automatically. Language models are now able to write computer programs based on instructions in English. They are becoming increasingly good at understanding images, videos, and music. They can generate music based on the user’s preferred style and sing given the lyrics.
How best can we utilize artificial intelligence in healthcare?
AI has tremendous potential within healthcare. Since trained healthcare professionals are either too busy or unavailable (e.g. in a remote location). AI systems can provide an initial diagnosis to detect early signs of a disease. This has a huge impact both on the health outcome for the patient and population health in general. AI systems are now routinely applied for image-based diagnosis such as MRI scans, X-ray, microscopy, etc. Such applications of AI will help in reducing the cost of service delivery and offsetting the lack of trained healthcare professionals.
This technology will of course be useful within a city setting as well. Doctors in cities can use AI as an assistant. This will save time and will help spot a possible misdiagnosis. AI systems can be employed to provide a second opinion to help the doctor. Also, AI systems can learn continuously from human feedback to improve its performance.
In Wiseyak we are focused on data quality and data interoperability. To train any AI, the most important aspect is that we need quality data. If the data quality is not good, neither humans nor AI can diagnose anything. Since the last couple of years, we are focusing on high quality data capture as a prerequisite for AI driven healthcare. We know that AI is evolving at a very fast rate and we are confident that we will be able to integrate the latest AI algorithms. For this reason, we are trying to obtain accurate medical data and ensure that doctors and healthcare workers can input the data with ease. Our WiseMD platform can record and store any type of medical data in a standardized form following HL7/FHIR standard. It permits AI algorithms to be easily added taking advantage of the latest developments in AI.
Our data platform also permits integrating medical data from diverse sources. Nepal is grappling with a challenge due to the development and deployment of multiple types of Hospital Management Information Systems (HMIS), resulting in a significant hurdle for data sharing. Presently, the inability of one HMIS to accept data from another complicates the landscape. However, Wise MD could serve as a solution by facilitating seamless data sharing among these systems.
Do you think artificial intelligence will one day complement doctors or will it substitute them?
For the next 5 years, AI is going to be there to complement the doctors and to reduce the time and fill the gap where you will need highly specialized and skilled doctors. AI partially fulfills the need of specialists. AI need not be 100% accurate but we need AI systems to have 0% (or close) false negatives. So, it’s OK to have a few normal diagnoses to be considered as abnormal. Since, further diagnosis is always carried out to confirm any abnormality. But we do not want abnormal cases to be considered as normal. Within the next 5-10 years, there is going to be robotics surgery and AI doctors doing routine diagnosis using natural language. AI is used for drug discovery, precision personalized treatment, mental health, surgery assistant and many more.
What are the challenges of integrating AI into the health and education sector globally and in Nepal particularly?
There are many challenges actually but the main challenge in healthcare is data like I mentioned earlier. To use AI, we need a computer and then of course we need data and till the time we use paper and pen for the data, it will be difficult to integrate data into AI, and subsequently AI into healthcare. So, I think that’s the biggest problem. AI can be employed to translate the conversation between the doctor and patient and automatically convert this to a medical record. This will make certain adoptions easier and faster. So, the main thing is everything needs to be digitized and this is the main problem. AI is not the problem but digitization is a challenge. We need behavior change from analog pen and paper-based health records to digital health records. This is not a technological challenge but a social one. Democratization of AI within healthcare and equitable digital healthcare will require us to adapt ourselves to an open standard based digital technology and it will be difficult. Healthcare is the most siloed industry sector where only a small number of global providers dominate the market in the electronic health record space.
How has the medical fraternity accepted AI in Nepal?
The major problem in our country is that hospitals are only focused on the revenue cycle and that is especially true for private hospitals; maybe not all but most are. I think we are still quite far in terms of digitization but once the digitization is done AI adoption will be fast. One beautiful aspect about us Nepalese is we can adapt to changes fast. Mobile banking can be taken as an example. The banking systems within the banks may not be fully digitalized but from the user’s side it has been digitized. Once we see a few successful hospitals using AI and if people demand it then the adoption will be fast.
For young people who want to study AI, what would you recommend to them as a student, research fellow or as a professor?
That is a very deep question. It depends. Talking about school, it’s not just AI, but for any science subject the teaching methodology needs to be changed. In Nepal we have a book culture where a student requires 10-15 books and is asked to rote learn everything. We need to get rid of this mentality. For instance, in the United Kingdom, kids do not go to school carrying heavy bags or they don’t memorize text books. They hardly carry one or two books, and mostly leave their books at school. They just bring homework back home and most homework is done in school itself. They teach students to become independent thinkers. However, in Nepal we are taught to rote learn to pass the exams. This has become a generational issue as the teachers themselves come from that background and they teach the same method to the students. When I was in Nepal, I used to study course books but I was less interested in those. I used to go to the American library and to the British Council and read books that I was interested in. This helped me to become an independent thinker. If I hadn’t done this, then today probably I would have been a different person. In that way I am self-taught. This culture is very important and is missing in our school children. The primary difference is we ask our students what X is forcing our students to memorize. We also need to teach our students the ‘why X’ and ‘how X’. This is the fundamental change that is needed.
Even in our engineering colleges we have the same methodology for our students. They are taught to follow a rigid recipe to build something. They are not taught to critically question the recipe itself and think about improvements to the recipe. So, it’s difficult for them to think like a scientist.
Our university education is compounded by the affiliation system. Since there are many affiliated colleges and most teachers in the affiliated colleges are less qualified compared to their university counterparts, the exam questions are designed to achieve a high pass rate across all colleges. This is a big problem that I see at the college level IT/CS/AI education here in Nepal.
So, my suggestion to aspiring AI scientists is to become independent thinkers and avoid cutting-and-pasting from online repos (Github, etc.) without truly understanding what you are doing. Of course, one can start by imitating and copying, but that should only be used for getting started. Secondly, understanding the fundamental concepts is important. This includes understanding the mathematical foundations. Do not believe everything you read in papers. Be critical and cultivate your own judgment. Do your own projects, make many mistakes and learn from these.
Conclusion
In summary, Prof. Suresh Manandhar’s interview provides valuable insights into the trajectory of artificial intelligence in Nepal, the versatile applications of deep learning, and the pivotal role of AI in transforming healthcare. His emphasis on the significance of high-quality data and digital integration in the healthcare sector underscores the challenges and opportunities for effective AI implementation. As the landscape of AI continues to evolve, the interview offers a glimpse into how AI can act as a complement, augmenting the capabilities of healthcare professionals. Additionally, Prof. Manandhar highlights the imperative for a paradigm shift in teaching methodologies, advocating for independent thinking and project-based learning to cultivate a new generation of AI enthusiasts and experts.
