On Big Data, Machine Learning for Healthcare. Q&A with Don Woodlock

Q1. Prior to joining InterSystems, you were the vice president and general manager for the Enterprise Imaging division of GE Healthcare. What are the main lessons learned?

I learned throughout my career that there is absolutely nothing more important than customer focus.  To reinforce this, I like to keep all of the conversations of the company about customers and products.  I always like to run a business like a start-up or a garage business. When you are with your two co-founders starting something new, you are not talking about budgets, approval chains, performance appraisals, dress code, working hours – none of that.  You are spending 100% of your time talking about customer and market challenges and opportunities for you to make a difference. So even though I’ve been at big companies and run large businesses, I treat it all the same – spend as close to 100% of our time talking about customers, their opportunities and challenges, and how can we help,

Q2. What are the current key challenges associated with delivering improved quality, accessibility, and efficiency across healthcare?

The main issue is that our healthcare system isn’t a system at all.  A system, in engineering terms, is “a set of things working together as parts of a mechanism or an interconnecting network.”  I don’t know anyone who would describe our healthcare environment like that.  We have massive fragmentation of healthcare and of healthcare information.  For example, a typical senior Medicare patient in the United States will see seven different physicians from four different organizations every year.  Over the course of a few years, you have a real mess for everyone involved – any given caregiver doesn’t have the complete picture of their patients, and patients, especially those with chronic conditions, have a very complex system to try and navigate.  Our clinical quality challenges and our efficiency challenges fundamentally stem from this fragmentation.  Anything we can do to help unify the parts of this system are critical to improvement.

Q3. Using machine learning into clinical medicine holds promise for substantially improving health care delivery. What are the top applications of Machine Learning in Healthcare?

Folks that know me know I’m a huge fanatic of Machine Learning, so I will try not to say “world peace” in my response to this question.  Let me start with what I don’t like, which is applications that try and “out-doctor the doctor.”  I think that there are so many more interesting, achievable, and valuable use cases than trying to outwit one of the most complex knowledge workers on the planet in one of the most critical jobs imaginable.  So, I’m not too interested in that topic.  But the big three for me are ‘predicting the future’, easing the administrative burden of healthcare, and conversational user experiences.

On predicting the future, a hospital administrator wants to know which of the patients in her hospital now are going back as a readmission so she can focus her discharge planning activities, a physician office manager wants to know which patients will come in today so he can overbook intelligently, a billing manager wants to know which claims are not likely to get paid, a lab equipment department head wants to know which machine will break down next, a life science company wants to know which drug will get move to the next stage of approval, and so on.  Think of how directed your healthcare resources can be if you have a better sense of what will likely happen next.  The ML results in all these areas are very promising and are much better than the techniques we all use today, including simple risk models or often expert hunches.

I also like ML for reducing the administrative tasks of healthcare, including the physician documentation challenge.  A revolutionary approach to having the events of the patient and the treatment interactions flow into Electronic Medical Records as an offshoot of life instead of a massive data entry challenge is a big opportunity for ML.

And lastly, I like conversational user experiences. The ‘User Experience’ of healthcare is complex and clunky and voice can make that more natural – for patients, care-givers, and other users.

Q4. What are the implications of Big Data in Healthcare?

Around the world, we have spent a large amount of time and money digitizing healthcare.  This was a good foundational step.  But now, we have tons of data but not enough insight.  A physician once said to me “I am looking for a needle in a haystack, and you IT people just keep adding more hay.”  That’s the risk: that we feed onto the pile of data, but we don’t make it any easier to take of a patient and we don’t make it easier to understand the dynamics, effectiveness, and trends at a population level.  So big data is partially technical – how do we store it, scale it, manage it, government, democratize it appropriately, that sort of thing.  But the domain parts are the most interesting to me: how do you bring impactful information to the surface.  How do you allow what matters to jump to forefront when you are staring at your data.

Q5. Developers of AI for healthcare applications may have values that are not always aligned with the values of clinicians,  How do you think we should address the Ethical challenges, such as the potential for bias, and the questions about the accountability  between patients, clinical doctors and AI systems?

I think there is a special burden on the AI community for transparency and explainability in healthcare models.  I think it’s irresponsible to throw a lot of data at a model, train it, see that it performs well from an aggregate accuracy point of view, not really understand how it is working, yet throw it into production. The importance of the decisions that may be guided by ML and the risk of these models focusing on the wrong features, like the background of an image instead of the foreground, as in the famous wolf vs. dog classification story, make the black box model a non-starter in healthcare.

Sophistication in visualizing models, feature importance, local explainability that can be shown to users so they understand why the model is making a particular prediction, and discipline in model building and testing that looks for bias and unintended consequences are all key.  It is also true that models aim to reinforce your training data. So, the wide variation in the practices of healthcare, some good, some bad, must be sorted through also in the model-building process.  Just because something was practiced in the past doesn’t automatically mean you should reinforce it with ML going forward.  The training data needs to be carefully curated to make sure it indeed reflects best practices that you want to reinforce, is broad in terms of representative populations and practices, and is free from bias.

Q6. Giving access to your health data means giving access to your private and proprietary information which can be used against you or in your favour. The handling of data strongly depends on the ecosystems you are in (e.g USA, China, Europe).  What is your take on that?

The values of citizens in different parts of the world and the values of individual patients need to be fully supported.  It causes complexity for us software companies that provide products in many countries, but so be it.  We in particular have very sophisticated patient consent functionality that enforces the local regulations and wishes of the patient – and it’s a centerpiece of our system. The advances in the patient de-identification space are a very important part of helping with this, as so many use cases—whether it be ML, trend reporting, population analytics – often don’t need identifiable patient information to be useful.  So, we can leverage technologies that can remove information and scrub clinical notes comprehensively to enable big data and analytics with less risk.  The second advance that I like is the federated learning concept – models that can be partially trained at the source of the data and only the weights sent somewhere centrally is another great way to get value out of the data without the risk and exposure of always having to bring it together.

Qx. Anything else you wish to add?

I would just say that we are the beginning of an amazing era in healthcare.  In so many parts of the world, we have made the move from paper to digital and that has had its primary benefits.  But this next stage of getting value out of the data that we are painstakingly collecting is upon us.  It reminds me a bit of radiology, one of the domains of my previous jobs.  We rolled out PACS systems, Picture Archiving and Communication Systems, 15-20 years ago and completely converted that department from film and lightboxes to digital.  What you see on TV with doctors putting X-rays in lightboxes doesn’t actually occur anymore – we sold all of our lightboxes to those TV studios.

Today, radiologists sit at large multi-headed high-resolution workstations and do their work there.  This has had enormous workflow and clinical benefits to patients and physicians.  And all throughout that era, Machine Learning was never mentioned once.  I really can’t recall it as a topic back then at all.  But now that everything is in digital form, ML is making huge impacts in radiology, perhaps the largest in healthcare actually, and it’s one of the most promising areas of breakthroughs.  While we were all digitizing radiology, we really didn’t know that down the road, the secondary benefits of doing this would be groundbreaking.  That will hit the rest of healthcare, which has more recently been automated. From my vantage point, we have digitalized, we have improved workflow, but the best is yet to come.


Don Woodlock, Vice President of HealthShare, InterSystems

Don Woodlock has had a 30-year history in the healthcare software industry.  As Vice President of HealthShare, he is responsible for strategy, product development, commercialization, implementation, and overall customer success in this segment.  The HealthShare solutions empower care providers and connect care communities around clearly presented, comprehensive, and actionable health information.

Prior to InterSystems, Don has held many senior leadership positions including VP and GM of GE Healthcare’s Enterprise Imaging business and earlier their Chief Technology Officer.  Prior to GE Healthcare, Don worked at IDX Systems Corporation, where he led the development of many of IDX’s flagship products including the successful Managed Care product and EDI Clearinghouse offering.  

Woodlock holds a BS degree in Electrical Engineering from MIT.

You may also like...