Making a success of AI in healthcare

By Shez Partovi

science healthcare

If you want to make a positive impact in healthcare with any data science project, keep two things in mind. First, put your customer and end-user at the center of everything. In the end, artificial intelligence (AI) is only as powerful as the human experience it makes possible because solving a problem using AI is about augmenting human expertise, not replacing it. Second, make sure you are addressing a real need. Naturally, these two go hand in hand, and you may find yourself going back and forth between the two. Avoid developing projects driven by dataset availability. Instead, focus on the problem to be solved and the experience you wish to create, then work backward and source the data you need.  

These simple rules are true in any industry, but especially so in healthcare where the primary objective is to improve people’s lives. If you fail to focus on the human experience and how AI will fit into an already complex workflow, you are at risk of inadvertently creating more work for healthcare professionals and generating little value. Even before the increased burden related to the COVID-19 pandemic, healthcare professionals were already overloaded. Now, we are hearing increasing signals that many are closer to burnout.

Use AI to relieve burden, not add to it

To ensure technologies such as AI and machine learning truly deliver value, you need to think carefully about how they are going to get seamlessly embedded into existing workflows in a way that allows healthcare professionals to gain the most from them. One simple way AI and machine learning can deliver value and gain rapid acceptance is to use such technologies to automate routine tasks that clinicians currently find time-consuming and repetitive. On the other hand, using AI as an “excuse” to introduce yet another application that clinicians have to learn and navigate would be the antithesis of automation and simplification.

Ensuring adherence to these foundational principles at all levels in a large organization like Philips can be challenging. One of the ways we are driving this at Philips is through our data and AI center of excellence, where teams from all our businesses and data scientists from across the organization can learn and advance best practices and develop frameworks, playbooks, and enablers. It leads to improved awareness, cross-fertilization of ideas across teams, and raising the bar.

Collaborate with end users, validate results

What often sets successful projects apart is close collaboration with clinicians throughout the project. For Philips, this means involving our clinical partners during the early stages of ideation and continuing to do so throughout the project lifecycle. For example, when developing Philips’ Radiology Smart Assistant—an AI-based solution that improves chest X-ray acquisition accuracy through instant feedback at the point of image acquisition—we collaborated closely with Cologne Merheim Medical Center in Germany. This collaboration enabled Philips to build features, such as AI-based patient positional accuracy, that Cologne Merheim Medical Center considered important for improving examination quality and enhancing image interpretation. The AI algorithm was trained on thousands of labeled PA chest images and clinically validated through scientific studies. In this case, the result of the AI-supported image quality evaluation was compared to evaluation by human experts for a representative set of images. It resulted in a solution that provides consulting radiologists with higher quality images, which can support decreasing additional costs, delays, and X-ray exposure associated with retakes. Another example where close clinical collaboration paid dividends is our award-winning solution for the fastMRI challenge—a collaborative research project that leveraged insights from Leiden University Medical Center and investigated the use of AI to make MRI scans up to 10 times faster. The researchers from Leiden, with the support of Philips, developed an algorithm with which it is possible to use eight times less data than normal and still reconstruct an MRI image that is almost as good as one using the usual amount of data.

Developing solutions like these requires us to immerse ourselves into the daily reality of the people whose lives we are trying to improve. That is perfectly illustrated by our collaboration, via our involvement in the Eindhoven MedTech Innovation Center (e/MTIC), with Catharina Hospital and Leiden University Medical Center in the Netherlands, where we are performing 360-degree on-site workflow analyses and conducting co-creation sessions with medical experts to look beyond the technology itself and understand the full context of its use. These collaborations allow us to uncover the experience drivers that support strong physician-AI interaction.

Take a multidisciplinary approach 

Beyond working closely with clinical partners, we also find that virtual teaming across our research and product development teams accelerates the translation of scientific research into solutions that benefit care providers and patients. For example, the insights we gained in the fastMRI challenge enabled us to productize fast MR reconstruction using deep learning.

However, you need more than scientific research and customer insights to develop an effective AI solution; you need an appropriately large training dataset that is high quality with robust annotation. Otherwise, you risk introducing potential bias in the resulting AI algorithm. To support our data science teams, we have created a catalog of available data sets within our organization, not only detailing their applicability but also any specific rights connected with the data, such as privacy and legal requirements, to ensure that we remain fully compliant with regulatory requirements.

We are also exploring how federated learning can help overcome some of the challenges of obtaining larger and more diverse datasets. Using federated learning, we take the training to the data and aggregate the model weights, rather than bringing the data to the model. As a result, the source training datasets can securely stay with the originators. 

Never stop improving

Any trained AI model will invariably require updates over time to ensure it performs well and improves. At Philips, we are building AI platforms that incorporate support for feedback loops and monitoring the performance or decay of our models over time, enabling us to identify when re-training might be necessary. We hope that in the future, regulatory bodies will find appropriate ways to enable continuous learning such that models can improve their performance either at a global level or with more local optimization.

Create insights not robots

We need to dispel the myth that AI and ML are going to replace clinicians. In reality, they are simply new tools in the tool-belt of researchers and healthcare professionals that will enable them to do what they do best by providing them with actionable insights. For patient managers, that could be predicting the probability that a patient will not show up for their appointment so that support can be offered to facilitate their attendance, or predicting ICU bed occupancy over the coming weeks to allocate appropriate resources. For clinicians, it could be predicting which patients will require an intervention in the coming hour or the probability that a patient will be readmitted to the ICU. For researchers, it could mean rapidly identifying potential drug candidates from databases containing data on tens of thousands of molecules. When you really unpack what AI and ML can do, it’s about providing those insights. But it’s what we do with those insights that defines our humanity.

About the author

shez partovi

Shez Partovi is Chief Innovation & Strategy Officer of Royal Philips, a position he has held since July 2021. Shez is an experienced clinical professor, neuroradiologist, global executive and entrepreneur with a track record of leading large health systems, cloud transformation, and artificial intelligence and machine learning in the healthcare, life sciences, and genomics industry.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.