The FDA’s approval of an electrocardiogram (EKG) that enables atrial fibrillation detection right from a patient’s watch band is just one example of how the digitization of medical devices, a part of the Internet of Things movement, is leading product development and innovation in medicine. However, while medical devices built on a connected services platform include components for data storage, security, accessibility, and mobile applications, along with advanced analytics, successfully implementing artificial intelligence to drive actionable intelligence remains a challenge from an execution perspective.
According to Gartner, 85% of data science projects fail. Successful integration of data science into medical device development requires a rethinking around the role of data science in product design and life-cycle management.
Viewing Data Science as a Product
While data science is rightly defined as the process of using mathematical algorithms to automate, predict, control or describe an interaction in the physical world, it must be viewed as a product. This distinction is necessary because, like any medical product, data science begins with a need and ends with something that provides clear medical utility for healthcare providers and patients.
It is erroneous to restrict the realm of data science to just the designing of algorithms. While data scientists are good at fitting models, their true value comes from solving real-world problems with fitted data models.
What it takes to develop a medical device algorithm
A successful algorithm development process in data science includes business leaders, product engineers, medical practitioners, and data scientists collaborating to discover, design and deliver. For instance, a typical data science integration with a medical device product would include many of the following activities:
- Identifying the medical need
- Identifying proper data variables
- Developing the right analytic models
- Designing analytic algorithm integrations
- Performing testing and verification
- Deploying beta versions
- Monitoring real-time results
- Maintaining and updating algorithms
Considering data science as a product or feature of a product provides organizations with a different paradigm for execution focused on a tangible outcome. Data scientists are trained to develop accurate models that solve a problem, but the challenge many companies face is operationalizing those models and monetizing their outputs. Furthermore, conceptualizing data science as a product will ensure companies focus on its implementation, rather than just its development.
Advanced analytics must be a part of the process and not just an afterthought. Designing intelligence (even AI) into a connected medical device first depends on whether the data is being used to make a real-time decision or report on the outcome of a series of events.
Most companies don’t realize the different layers of advanced analytics that create actionable intelligence. They may include:
- Simple rule- and complex rule-based analytics
- Asynchronous event rules
- Complex event processing, and
- Unsupervised learning models
By understanding these layers, companies can move quickly into developing mature analytics that have an impact from day one.
As a company matures its analytics system from descriptive and diagnostic to predictive and prescriptive, it should also evolve to include strategic opportunities to provide business value, including automating decisions that can be delegated to a smart decision-support system.
Successful integration involves viewing advanced analytics as an architecture and not as a single solution to be implemented. The best way to make sure that you are successful in analytic development is to follow a continual process of discovery, design, and delivery.
For instance, data science architecture may begin with a business question, requiring you to determine if you have the right data and can actually leverage that data in the existing IT system. If you don’t answer this basic question, you will have challenges fully vetting the analytic opportunities available to you.
Common Challenges in Data Science Execution
Data science execution is often impaired by common missteps, like incongruence between customer and business needs and solving technical problems when it’s too late to have a positive impact. Another significant mistake from the business side is treating data science like a one-time accomplishment and not realizing it is a continuous process. Also, there is often an unwarranted fixation on tools rather than skills and capabilities.
To use a common metaphor,
data science is not a single moon shot, but laps around a track.
Ultimately your goal is to run progressively faster around the track.
An equally major drawback hindering execution is artisan thinking where design is seen as the ultimate end to the data science process. In fact, the most desirable approach is a modular system with emphasis on consistently maintaining and improving what has already been designed. This is particularly true for medical devices where innovation and changes in technology are continuing to better support and enable patients and practitioners.
Developing products inside of analytics
Successful integration hinges on clearly identifying the data science process control. To design and support connected equipment analytics, data science should include clear steps such as,
- analytics incubation
- analytics validation
- analytics enablement
- analytic consumption
- analytic maintenance.
While the first two steps are where a data scientist will play a vital role, the subsequent three steps are what will ultimately lead to successful implementation and require strong organizational cross-functional support.
Why businesses fail at integration
Businesses fail at successful integration for a variety of reasons, ranging from
- insufficient investment in project management
- the inability to update or replace analytic components
- downplaying security threats
- not having alternative plans or an exit strategy.
For instance, the update of a pacemaker requires the analytic component to be upgraded as well. Similarly, while data cannot be patented, the process can be, and this will keep the analytics secure.
Often, failure to get buy-in across the organization and the greater focus on technology rather than business strategy also causes integration to fail. However, it is vital to remember that connected products and data science are still very new and change management is a common theme and challenge when it comes to success.
For data science to become a business discipline, it also should be scalable. This includes measuring all processes and reducing iterations. It also involves socializing metrics to identify advantages and inefficiencies.
A successful data science practice stems from the seamless coordination of people, process and technology. And, successful integration is dependent on identifying business goals and building analytics into connected devices. Only then can a business strategy evolve from the development of products with analytics to developing products inside of analytics.