A brain that is half artificial intelligence, half human

First Look – AI Foundation Success Factors

The Problem: AI: Long on Promise and Short on Proof

This is a statement by Eric Topol, cardiologist and geneticist at Scripps Research and author of the 2019 book Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. AI is being touted as the savior of drug discovery, clinical decision support, diagnostic accuracy, diagnostic image evaluation, and administrative process efficiencies. One of the best quotes I’ve seen to date is by Frida Polli, cofounder and CEO of Pymetrics, “[AI] is like teenage sex. Everyone says they’re doing it, and nobody really knows what it is.”

Key challenges for AI are related to the lack of good data to train machine learning systems[1] as well as the need for AI solutions to be interpretable[2]. Key databases of drugs or patients are rarely shared, and often have poor data structures that do not improve machine or deep learning AI environments when shared. AI solutions must also support “interpretability,” or the ability to interpret machine learning algorithms relative to the accuracy of their output and outcomes. In many cases, AI solutions are designed with “black boxes” that prevent users from evaluating the clinical impact of the AI guidance.

The Solution: Collaboration Between Providers and AI Solutions Vendors

While the Ascension and Google partnership is receiving a lot of negative press regarding the access of de-identified patient data from a provider for creating effective digital and AI solutions, this is likely the most effective way to drive successful AI solutions for healthcare[3]. Provider organizations have been storing clinical information for years and building large enterprise data warehouses that have been used to generate predictive analytics on operations and patient populations. These organizations have been creating effective data structures to enable these analytics. Provider organizations also have large diagnostic image databases that are not being effectively used in many cases. The same applies for clinical trials and drug data bases for pharmaceutical companies. Collaborations between those with the data and images and the companies that are driving sophisticated AI solutions will be critical for demonstrating the true potential and value of AI.

Representative collaborations between AI companies and providers have been established not only with Google, but with Imagen (Mayo, Stanford, and others), Proscia (Johns Hopkins and the University of Pittsburg), and Qventus (working with leading academic, public, and community hospitals across the US). We believe that these collaborations between providers and AI solutions vendors will drive new advances in treating cancer, chronic disease, and genetic disorders. If providers are not already involved with viable AI solutions vendors, they should seek out relationships with these vendors in order to mitigate their risk with AI investments.

The Justification: Mitigate AI Investment Risk

Providers who seek out collaborations and partnerships with advancing AI solutions vendors are most likely to realize an ROI on their AI investments. These business relationships will serve to not only help advance the AI solutions but also to help build AI experience and knowledge within the provider organizations. The partnerships can also provide additional benefits to the provider organizations for the enterprise data warehouses that they developed. This approach also provides the foundation for the organization’s clinical experts to create the policies and processes for evaluating AI interpretability. Interpretability of AI is likely an extension of medical informatics.

The Players: Long-Term Viability Will Drive AI Solutions Winners

While the list of AI solutions across many healthcare environments is too numerous to list, a key consideration for any provider organization must be the long-term viability of the AI vendor. Vendors such as Google, Amazon, Microsoft, and Apple will likely develop or acquire successful AI solutions. Substantial VC or PE funding of AI vendors may not be a good indicator of long-term viability.

Success Factors

  1. Providers will need to create strategies to access large normalized data environments in order to create a solid data foundation for AI success from existing or shared enterprise data warehouses.
  2. Providers will need to develop “interpretability” policies and evaluations for AI solutions that are used in their organizations and led by medical informatics teams.
  3. Ensure that all data used for AI is de-identified and that patients have opted to allow their data to be used.

Summary

Last week, in a KLAS meeting, Doug Tolley made the following observation: “Artificial intelligence today is more artificial and less intelligence than we had hoped.” The industry needs to be honest and open on how to advance AI without hype and how to use evidence-based medicine to demonstrate advances that truly deliver outcomes that help reduce healthcare costs while driving higher levels of outcomes and patient safety. While many of the current successes are based on machine learning solutions for administrative payer transaction efficiencies, the real potential of AI lies in deep learning for drug discoveries, diagnostic accuracy, clinical decision support, and diagnostic image evaluations.

Several of the large digital or IT companies such as Google, Amazon, Microsoft, and Apple already have a significant repository of healthcare data. These companies will only be successful with their AI strategies if they involve providers to help them better structure the data while opening their black boxes for analysis, evaluation, and ongoing algorithm tuning. Let’s hope that our industry can quickly move and work together to improve AI capabilities that improve and advance patient care.



[2] Interpretable Machine Learning in Healthcare, Muhammad Aurangzeb Ahmad, Carly Eckert, and Ankur Teredesai, KenSci Inc., Manuscript received July 31, 2018; revised July 31, 2018