CONNECTED HEALTH CONFERENCE • SAVE THE DATE • October 16-18, 2019 • Boston, MA
by Simone Orlowski, PhD, Research Scientist; and Sunetra Bane, MPH, Human-Centered Design Specialist, Partners Connected Health Innovation
You’ve heard the buzzwords – Artificial intelligence (AI), machine learning, predictive algorithms, natural language processing, big data – and you may know that scientists are creating algorithms for robotic surgery, image analysis and to determine correct drug dosages.
While great strides are being made to envision and develop predictive machine learning algorithms, there is a significant lag in the adoption of these in real-world care delivery. In our work, we are guided by some key principles that can set up an algorithm for successful adoption: these guiding principles must be considered early and repeatedly throughout the formative stages of the design of machine learning algorithms.
First, many organizations focus resources towards building out a specific use case based on research and development priorities. Very few actually invest the upfront time and energy to match these R&D priorities with a true clinical need. With the wide range of demands in healthcare faced by a multitude of clinical and non-clinical staff with specific and specialized roles, a one-size-fits-all approach will likely benefit no one.
Recently, the Connected Health Innovation team at Partners HealthCare explored the use of machine learning algorithms to aid in hospital readmissions. A range of frontline staff currently contribute to this effort based on their specific roles; therefore, strategic decisions should be made about which of these staff should use, and benefit from these tools, and why. For example, we spent significant effort understanding the landscape of readmission reduction, across a range of workflows and care settings, to determine which individuals prioritize which data points and have the ability to act on them based on their roles.
Additionally, we often find that even though a true clinical need has been identified, a predictive machine learning algorithm to meet this need may require data that is not collected routinely. A related challenge is obtaining the right data at the right time.
The added value of a ML algorithm is based on the large number of variables, sheer processing power, and predictive abilities that far exceed traditional statistical techniques. However, this value-add must be balanced with the ability to access the required data when needed. For example, an algorithm may be designed for use during an inpatient stay, however, the data needed to make the prediction may only be available after the patient has left the hospital.
Finally, for a predictive machine learning algorithm to be adopted, it needs to provide actionable value above and beyond standard clinical care practices for it be worthwhile for users to change their existing behaviors. In our recent work in readmission reduction, we found that an ideal tool would help users identify modifiable factors to change, and a clear pathway to implement those changes. A tool that merely confirmed for users what they already know would not be a must-have tool. In a busy workday, and a fast-moving and competitive readmission landscape, we must be in the business of designing essential tools, not nice-to-haves.
Taking time at the beginning of the algorithm design process to gain this knowledge uncovers unforeseen barriers to change. Thankfully, good design can temper the disruptive tendencies of innovation in new technologies, especially in healthcare. By consulting with end users at the outset of the process and collaboratively designing tools to fit into their workflows, you will have built a tool that meets a core user need and assists clinicians in providing optimal care, thus improving the healthcare experience for both physicians and their patients.
Gain firsthand experience with these concepts during an immersive, two-hour workshop at the Partners Connected Health Innovation-sponsored breakout session on October 18, “From Hype to Reality: How to Make AI Work for You” at this year’s Connected Health Conference. Dr. Orlowski will also address this topic during the panel discussion Thursday afternoon, “Driving Engagement: Recognition, Rewards and Incentives in Digital Health.”