Over ten years ago Pascal Metrics embarked on a journey to use data to improve the safety and reliability of patient care by collecting clinically validated EHR-based adverse event outcomes. Today, Pascal holds the largest such data set worldwide, which has enabled unique patient safety R&D using fine-grained adverse event outcomes data — starting with a first-ever study in 2012 applying machine learning (ML) and advanced artificial intelligence (AI) techniques to such outcomes data.

The results of this unique research were published today in the Health Affairs in an article entitled, “An Electronic Health Record-Based Real-Time Analytics Program for Patient Safety Surveillance and Improvement,” This piece focused on “Using The EHR” for advanced analytics as part of Health Affairs’ patient safety-dedicated issue sponsored by the Gordon and Betty Moore Foundation. We will discuss this article or more detail in a future blog post.

The most important takeaway for the field is that, despite the great promise of technologies such as ML and advanced AI techniques patients will not be kept safe by AI alone. Indeed, sound epidemiology must remain foundational, with patient safety and quality standing to be improved with scientifically valid and clinically credible use of AI — or what Pascal calls “AI-assisted” patient safety and quality improvement.

As leaders in the field and industry respond to the hope of AI, the solution is neither a luddite reaction to the use of advanced data science-driven technology, nor is it to be overly enamored with the promise of AI to the exclusion of bedrock principles of epidemiology. Both are required in a comprehensive system which includes components that range from fundamental all-cause harm measurement to powerful next-generation prediction.

Along this vein, the Health Affairs article referenced concentrates on two components: a Patient Safety Active Management (PSAM) system to measure, monitor, and manage all-cause harm and a Safety Predictive Score (SPS) to predict global harm and, ultimately, specific harm. These are, respectively, genericized references to Pascal’s Risk Trigger® Monitoring system and the Global Safety Risk (GSR) Score. A more extensive list of components in a comprehensive system are:

  1. Patient harm detection, including measurement monitoring
  2. A patient safety & quality improvement management system
  3. Patient harm prediction, including of both global harm and specific harm
  4. Expert education, training, and coaching, based on real world experience in translating real-time information into intervention & improvement
  5. Safety outcomes data, i.e. clinically validated EHR/HIT-based adverse event outcomes data set
  6. Patient Safety Organization (PSO), a U.S regulatory workflow and environment supporting all of the above

To be clear, ML and AI are not solutions per se. They are horizontal technologies to improve the processes and products used by people to deliver patient care. If trained with high quality outcomes data, ML and AI stand to extend and enhance healthcare performance results. Contrariwise, applying deep learning to huge volumes of data without common and sufficiently fine-grained definitions of patient harm will fail to fulfill expectations.

Indeed, adopting a comprehensive system to identify and reduce patient harm and related cost should include all of the above. Further, leaders planning implementation and execution are well served to keep operations:

  • Evidence-based
  • Regulatorily-aligned
  • Collaboratively-enriched
  • Expert-led
  • Value-oriented

Start with the evidence and end with value. Successfully applying machine learning and AI to patient safety and quality improvement is a journey. For most hospitals, achieving a scientifically robust and clinically useful measure of harm is the first step. Recognizing that progress is inevitably incremental is required while rejecting organizational inertia to “run in place” with traditional methods.

For example, seeking to apply AI when still relying on event/incident reporting — which identifies only about 5% of total harm — is akin to applying calculus before mastering arithmetic. Undoubtedly, it’s “both-and,” but proper sequencing is necessary.

The journey deeper into next generation patient safety & quality improvement continues…”Know thy harm.”