Financial Sustainability

How using artificial intelligence enabled Flagler Hospital to reduce clinical variation

January 17, 2020 9:28 pm

One community hospital shares its lessons learned from engaging physicians and applying artificial intelligence to develop care pathways for treatment of pneumonia, with the goal of improving quality of care and reducing its cost.

The variation in how patients are diagnosed and treated accounts for a large portion of roughly $1.2 trillion in costs wasted each year on healthcare that doesn’t improve outcomes. Limiting this clinical variation would improve patient outcomes, reduce overall healthcare costs and enable hospitals and health systems to better handle financial risk contracts. These organizations can take a lesson from the success of Flagler Hospital in St. Augustine, Florida, in harnessing artificial intelligence (AI) software for this purpose .

Flagler’s clinical variation management (CVM) initiative was the result of a board mandate to reduce clinical variation. As a 335-bed community hospital, Flagler had neither the resources of a large academic medical center nor IT staff specially trained in machine learning. Thus, the approach it used is well within the reach of small hospitals.

A focus on care pathways

To minimize clinical variation, hospitals create care “pathways,” or plans based on nationally accepted clinical guidelines. But clinicians don’t always follow these pathways, in part because they perceive that the care maps are based on evidence from academic studies that do not necessarily represent the patients they see.

Flagler determined it could best manage clinical variation by developing optimal care pathways based on how its own best physicians treat patients. But it also understood such a project would require analysis of massive amounts of data from multiple systems, and traditional analytic programs would not be up to this task. An AI application would be effective, however, because it can correlate a vast number of variables and quickly detect patterns in the data.

To pilot its initiative, Flagler focused on optimizing the care pathway for pneumonia.a Applying AI was only part of what was required for success. Flagler’s story shows the steps taken for success in meeting its mandate.

Job one: Keep it simple

It took little time for the hospital to determine an AI application would be needed to meet the board’s 2018 mandate. After exploring many applications, Flagler adopted one that was being used in some large health systems and didn’t seem to pose major technical challenges.

Having received the green light for the pilot, Flagler’s medical informatics staff took the technical steps of implementing the solution. They wrote nine queries in structured query language (SQL) for five systems: the hospital’s electronic health record (EHR), enterprise data warehouse, and surgical, financial and corporate analytics platform. The SQL staff server experts had only to compose the 2,500 lines of code required to extract all this data. After the data was pulled, it was validated semantically and syntactically.

Fast healthcare interoperability resource (FHIR) standards were used to pull the aggregated data into the CVM software.b

Meanwhile, a workgroup of physicians, pulled from each of the hospital’s departments, was formed. This team selected the variables that would be critical in building an optimal care pathway for pneumonia, including “continuous” variables, such as cost and length of stay (LOS), and “categorical” variables, such as where patients came from and their comorbidities.

The CVM application used unsupervised machine learning to understand the structure of the data and find patterns in it that corresponded to the chosen variables. The algorithms automatically grouped patients into cohorts, according to their outcomes and treatments received. The application also showed the direct variable costs, average LOS, readmission rates and mortality rates for each cohort, along with the statistical significance of the data. Comorbidities also were factored into the program’s calculations.

Based on analysis of thousands of hospital records, the application highlighted the most prevalent care pathways in the hospital and associated outcomes and costs. The physician team was able to drill down into these models to see the steps and step sequences within each pathway accounting for the differences in results.

To provide the foundation for Flagler’s new pneumonia patient care pathway, the team selected the cohort with the shortest LOS, the least readmissions, the lowest mortality and the lowest cost. To implement the new pathway, they changed the pneumonia order set in the EHR and included evidence from consensus guidelines.

Physician compliance is imperative

Hospitals seeking to replicate Flagler’s success with CVM should understand it requires far more than simply acquiring a similar CVM application. Strong buy-in from physicians and nurses also is needed. Gaining this support starts requires efforts to involve frontline clinicians in the selection of the new care pathways, and to educate staff on the importance of reducing clinical variation to improve patient outcomes.

Flagler has seen a strong response from physicians, with a majority adhering to the pneumonia care pathway. Flagler attributes this positive response to two factors: The pathway was developed using best practices data from the hospital’s own physicians, and the final product reflects the professional views of the physicians, who saw the justification for each step the software recommended.

Although cost control is important to hospitals, it doesn’t necessarily carry the same weight with clinicians. For clinicians, the goal should be patient safety and better outcomes.

6 key success factors

Flagler used the following six steps in applying its AI solution to optimize care pathways.

1. Educate staff about the importance of CVM. Organizations should hold training sessions for physicians and nurses explaining the purpose of CVM and how it will affect clinicians, and underscoring how following care pathways that have been shown to produce the best results can help clinicians deliver higher-quality care. Instructors should emphasize that physicians will retain control over the development of optimal care pathways and can depart from them when they feel it is necessary.

2. Identify similarities in outcomes and their relationships to care processes. All care is defined by events. The sheer volume of events in any episode of care (a lab test, a drug, a puncture, rehabilitation, office visits) is immense, and is almost entirely captured in EHRs and other systems. But for clinicians to really understand the variation in the care provided to different patient cohorts, the events’ sequence and timing also must be visible. These data points should be pulled from the source systems and integrated with the events to yield a true understanding of care pathways. Using machine learning, hospitals can rapidly identify which pathways are most common, the efficacy of the pathways based on patient outcomes and cost and the statistically significant differences among pathways.

3. Identify best care practices. An AI-based CVM application can identify candidate pathways based on the outcomes of patients’ treatments for a condition or procedure. But it should be up to physicians to select the best pathway. The hospital should urge its physicians to form a committee charged with this task. Using the pathways automatically generated from the data on patient cohorts, the committee can easily compare the differences to identify the elements of an optimal care pathway in terms of patient outcomes and other variables such as cost, LOS, mortality and readmissions. Clinicians should apply their own knowledge of medicine to the results to determine whether a care variation is contributing to the best outcome, thereby helping the team build the optimal care pathway.

4. Implement the new care pathways. The identified best practices serve as templates to be disseminated hospitalwide through integration with the EHR. Lab and medication order sets may be modified in the EHR to help facilitate clinicians’ adherence to the care pathway. Physicians also should receive alerts providing reliable recommendations on how to deliver the right care to the right patients at the right time. But they should also be allowed to deviate from the care pathway when their judgment indicates a patient would benefit from a different approach. Physicians must have the power to build on a care pathway by adding and subtracting events based on their clinical knowledge. Clinicians also should receive feedback from the AI software if the deviation from the care pathway enhanced or reduced the pathway.

5. Measure adherence with the care pathway. Measuring clinicians’ adherence to care pathways is essential to ensure the hospital is realizing the full benefits of the CVM initiative. But tracking compliance manually can be a huge operational challenge, because nurses must review charts representing every aspect of each patient’s hospital stay to record the requisite series of data points. Hospitals therefore need to be able to determine adherence automatically. Ideally, a hospital’s AI application for CVM should have this capability, but the hospital alsomust make full use of it. The monitoring tool can be used for measuring compliance with both new and existing care pathways, thereby providing a means for improving quality scores. Users should be able review the steps taken to adhere to specificguidelines, by physician and department, and generate comprehensive reports comparing physician performance across the board.

6. Provide for continuous improvement. AI-based applications are designed to improve continuously as they learn from new information. Once the data feed from all relevant systems has been established, the algorithms will continue to work on new data and refine their ability to identify best practices. As physicians incorporate new medical treatments and diagnostic tests, they should use the application to determine whether these techniques are improving patient outcomes. Once optimal care pathways have been developed for most of the common conditions, physicians will be able to open patient charts and immediately access prescribed best practices for each condition.

Computers and the art of medicine

Physicians are understandably skeptical about following advice from a computer. After years of experience with alerting systems prone to providing false alarms, many physicians are inclined to ignore these systems.

This reality underscores the importance of Flagler’s use of data on its own physicians’ practices in its efforts to identify the best pathways with the best outcomes at the lowest costs. It ensures physicians will perceive that the pathways have a practical, real-life basis.

Aided by the latest technology, Flagler has managed to deliver care more effectively and cost-effectively, without breaking the bank or hiring data scientists, using an approach that is within reach for most community hospitals. In the short term, such an approach can help these hospitals reduce LOS and readmissions, and thereby increase the bottom line. And in the long run, it can help them improve outcomes and be better prepared to manage financial risk.

Flagler’s care pathways to date

  • Pneumonia
  • Septic Shock
  • Congestive heart failure
  • Chest pain
  • Chronic obstructive pulmonary disease
  • Myocardial infarction
  • Coronary artery bypass graft
  • Total hip and total knee replacement


a. Future projects would focus on total hip and total knee replacement, sepsis, chronic obstructive pulmonary disease, heart failure and diabetes, among other areas.

b. For information about FHIR, which is a widely used protocol for integrating disparate systems, see Bresnick, J., “4 basics to know about the role of FHIR in interoperability,” Health IT Analytics, March 22, 2016.


googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text1' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text2' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text3' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text4' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text5' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text6' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text7' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-leaderboard' ); } );