Data Science Professionals Need To Have Domain Knowledge: Sreetama Das, GSK

From being a researcher in Computational Biology & Structural biology/ Biophysics to a Senior data scientist, Sreetama Das has gained a wide knowledge base and worked in diverse industries like manufacturing and healthcare. Analytics India Magazine caught up with Das, who is currently serving as a Senior AI ML Engineer at GSK, to understand her insights on AI-based solutions in digital health.

AIM: Given your experience in a broad career path from data science to AI and machine learning, what do you think will be the future of the tech-medical space?

Sreetama Das: To briefly mention my background, I have worked with data science and machine learning applications to biomolecules as part of academic research during my PhD, and then worked in diverse industries like manufacturing and healthcare. This experience has shaped my view of what I think the field will evolve to be.

Several areas in the tech-medical space use machine learning for improved outcomes – for example, research and development for drug discovery or repurposing, optimising clinical trials, manufacturing and quality control, digital health sensors, digital pathology patient triaging, to name a few. The data could be numeric, structured nicely in tables, or huge images or even messy text – that is what makes problem statements in healthcare exciting but challenging. Advancements in computer vision and natural language processing will improve solutions or even solve problems that were not feasible earlier. In addition, solutions focussing on explainability will find more acceptance. The overall trend is to move towards faster development, efficient manufacturing, better quality control, and easier access to health monitoring, disease detection, and patient treatment outcomes. In short, AI-based solutions will be assisting to improve lives.

AIM: There have undoubtedly been significant advancements in the field of digital health. Is there a particular situation that prompted you to conduct additional research?

Sreetama Das: Digital health has seen significant advancements in developing digital sensors for health monitoring and detection of diseases. I will cite an example from a project I was part of at my earlier organisation (Robert Bosch Engineering and Business Solutions, India) since I have been with GlaxoSmithKline for only a short time. Back there, the team developed a novel, non-invasive haemoglobin monitoring sensor based on photoplethysmography. We collected data from many participants, both healthy and frail, and trained machine learning models. As a result, our device performed better than some existing non-invasive hemoglobinometers, which tended to overestimate haemoglobin levels. The project required a lot of research, and our results have been published in several reputed scientific journals.

AIM: Enterprises began migrating away from on-premises data centres and applications toward cloud and SaaS-based solutions. How feasible is this transition to the cloud in light of the numerous security concerns and ransomware attacks occurring?

Sreetama Das: I am not an expert in this area, but I am aware of private commercial cloud and hybrid cloud offerings that combine the advantages of cloud with the security of on-premise solutions and are used to address regulatory concerns. Moreover, we are starting to see the use of blockchain in commercial cloud solutions and upcoming technologies like federated learning and its use in med-tech. So, I think the transition is feasible with well-thought-through strategies. 

Download our Mobile App

AIM: In what we call a “combination of AI and Digital Health”, what are some near or so groundbreaking developments that carry future potential as well?

Sreetama Das: One of the groundbreaking developments in recent times would be the development of AlphaFold (by Deepmind) which has revolutionised the science of protein structure solution and has important implications for drug design. The software provided near accurate theoretical models of protein 3D structures, which are presently obtained by time-consuming and expensive experiments or sometimes represented by not-so-accurate theoretical models. Implementation of AlphaFold generated models will dramatically reduce the search space, thereby reducing the time and expense in the drug discovery process.

Other advancements include deep learning in several areas, for example, in the classification of chest X-rays of normal flu versus COVID-19-infected cases with high accuracy or in early detection based on voice abnormal respiratory sounds. In addition, image and video analytics are finding applications in monitoring physical activity or senior patients for falls or other issues. Finally, precision personalised medicine is also an area with significant potential.

AIM: In general, why do you believe there are widespread misconceptions about artificial intelligence in healthcare, and why do they persist?

Sreetama Das: I believe there are several sides to this question. Firstly, when there are breakthroughs in artificial intelligence algorithms, many articles in general forums are often very flattering, without trying to explain how they work, the input requirements or where they may fail. As a result, there is a disappointment when such tools do not generate the expected results, either due to inappropriate application or improper data. It is important to remember that gathering enough good data is still a challenge for several healthcare problems. Also, some problem statements may not be feasible to solve and require modification after multiple rounds of discussion with the stakeholders. Moreover, healthcare is a sensitive topic, and people are often apprehensive (and sometimes rightly so) about accepting ‘black-box’ solutions or the ‘fairness’ of such solutions. 

Hence, raising awareness about artificial intelligence and addressing concerns – data requirements, understanding how machine learning works and what is feasible, and model explainability – will help remove persistent misconceptions.

AIM: Could artificial intelligence eventually replace clinicians?

Sreetama Das: Human biology is highly complex – we still don’t understand everything. Clinicians make decisions by looking at many different aspects and based on their experience. It is hard to design ‘general’ machine learning solutions that can work similar to an experienced clinician – only solutions to very specific AI can develop well-defined problems and where a lot of data are available. Since healthcare delivery directly impacts people’s lives, the future would have AI solutions “assisting” the clinicians in their decision-making rather than replacing clinicians altogether.

AIM: What resources, such as books and journals, would you recommend for aspiring professionals in this field? 

Sreetama Das: A lot of useful information is often found in blogs on Medium. I refer to GitHub for codebases. Other sites to look out for will be papers with code and company blogs for recent developments (e.g., Google, Facebook, Microsoft). There are also many free videos and course materials (e.g. MIT Courseware) for aspiring professionals to get started.

AIM: What would you recommend as an efficient career direction for professionals aiming to make it big in the digital health space? 

Sreetama Das: I will answer this question based on my journey. Data science professionals need to have some domain knowledge to develop appropriate solutions. Since digital health is a vast space, it is important to be aware of that and identify the topics the professionals are interested in and understand. 

It also helps to be flexible and adaptable to change since different areas may gain traction at different times in a professional’s career. Finally, we need to be lifelong learners and acquire relevant knowledge when the need arises.

Lastly, it is important to focus on the foundation for such careers. Good coding skills, understanding of key concepts, and good communication are important. Also, please remember that such projects are often based on teamwork, with different skillsets within the data science team (analyst, ML engineering, MLOps, etc.). So, it is important to talk to peers and learn more about these aspects.