Artificial intelligence—the simulation of human intelligence by machines—is rapidly becoming a key enabler for businesses to deliver consistent, high-quality, and efficient outcomes. Healthcare organizations across the value chain are making significant strides in embedding AI capabilities in areas such as diagnostics, medical imaging, and lifestyle management.
One healthcare process that could potentially be improved through the application of AI is prior authorization (PA). PA is a core administrative process in which payers require providers to obtain preapproval to administer a service or a medication as a condition of coverage. The goal of PA is to ensure members receive the most appropriate care for their medical needs in alignment with the latest medical evidence and guidelines. PA can prevent wholly inappropriate service utilization or, more commonly, ensure that first-line treatments are attempted before escalating to more invasive or risky therapies.
The PA process can lead to straight approval, a recommendation to either start with a less expensive option (called step therapy) or to pursue an alternative treatment path, or a denial of coverage. Payers consider PA a core component of their utilization management programs, which aim to determine the appropriateness of medical care and to reduce fraud, waste, and abuse in medical spending.
While the healthcare industry has made strides toward standardizing and automating PA,
the process continues to be an administrative challenge. Payers’ clinical staff must devote considerable time to reviewing PA requests. Meanwhile, doctors and staff report spending 13 hours per week on PA;
many clinicians believe it undermines their clinical judgment and can inhibit timely care.
Our analysis suggests that AI-enabled PA can automate 50 to 75 percent of manual tasks, boosting efficiency, reducing costs, and freeing clinicians at both payers and providers to focus on complex cases and actual care delivery and coordination. This, in turn, may improve the healthcare experience for both clinicians and insurance plan members.
In this article, we discuss the state of automation in PA; how AI may be able to transform the practice; the potential benefits this would bring to payers, providers, and members; and challenges
the industry will need to overcome.
While the healthcare industry has made strides toward standardizing and automating PA, the process continues to be an administrative challenge.
The benefits of AI across industries and value chains
Led by the high-tech and telecommunications industry, the automotive and assembly industry, and the financial-services industry, companies around the world are embracing AI across their
value chains (Exhibit 1).
While AI adoption in healthcare lags more digitalized sectors, the technology has promise for many applications in the midterm to long term, including predicting the onset of disease; facilitating decisions in radiology, surgery, and post-acute care; and using AI-enabled chatbots to schedule doctor appointments or perform other operational tasks.
Reducing administrative overhead, which currently accounts for 25 percent of total US healthcare spending, is an important objective of AI innovation. PA is a core administrative process that primarily consists of manual information exchange and reviews by payer clinical staff (Exhibit 2). We have observed that in the most drawn-out cases, the PA process has high administrative overhead for both payers and providers. Moreover, our analysis revealed that a high rate of manual decision making can lead to inconsistent clinical determinations given the variability in clinicians’ training and in interpretations of medical policies.
As a result, payers and providers often disagree on the value and practice of PA (see sidebar, “Divergent views on the prior authorization process”).
To improve efficiency, reduce provider mistrust and dissatisfaction, and eventually improve both provider and member experiences, some payers have begun the PA automation journey. Electronic prior authorization (ePA), for example, accelerates
the exchange of information between payers and providers, as shown in steps one through three in Exhibit 2. ePA digitalizes workflows
and results in faster turnaround times. More than 60 percent of requests were completed in less than two hours with ePA, compared to zero requests submitted by phone or fax.
Nonetheless, ePA can still involve significant manual work for the payer, including gathering medical evidence and cross-validating information with patient records. Software vendors are helping to solve this problem with growing support from
payers and providers by integrating ePA platforms into electronic health record (EHR) systems and workflows.
That said, this integration is seldom easy, and ePA is still an emerging trend with room for growth and improvement: provider adoption of ePA is less than 50 percent, with about 20 percentage points fully electronic, about 45 percentage points partially electronic, and the remaining 35 percentage points fully manual.
Enabling PA automation with AI
AI has the potential to create the next generation of PA workflow design. In concert with digital and workflow management technologies, AI may help organize information from EHRs, emails, policies, medical protocols, and other sources, vastly reducing low-value, time-consuming tasks involving searching, collating, and cross-checking information that people have traditionally done manually.
Two AI technologies are integral to this effort: computer algorithms, which render decisions consistently based on a set of defined procedures, and natural language processing (NLP), which extracts, interprets, manipulates, and assimilates unstructured or structured spoken or written data. In simpler cases, AI significantly increases automated decision making, while in more complex cases it can aggregate and present information to a clinician to make a decision.
In an AI-enabled workflow design, there are two key components: a triage engine and an automation engine (Exhibit 3).
The triage engine
The triage engine relies on data from multiple sources, including member eligibility and benefits information; member clinical and pharmacy claims; historical authorization requests, including related clinical decisions, appeals, and outcomes; and, where available, elements of EHRs shared by providers.
Triage is based on rules that are dynamically generated using an ensemble of classification algorithms. It determines the complexity level of the request using whatever data are provided, along with the member’s clinical history and the experience the payer has gained from previously processing similar requests. There are four complexity levels based on the variety of data available and how much human judgment is required to make appropriate clinical and coverage determinations:
- Low: These requests require only members’ past
claims history (medications, comorbidities, and so on), information on the PA request form, and the provider’s history of PA requests (such as the number of requests, the approval rate, and the responsiveness to payer requests). Other clinical
and coverage determinations include whether the requested treatments are proven and established and whether step therapy (beginning medication with the most preferred drug therapy and progressing to other therapies only if necessary) would be appropriate. The provider is notified if the request is not approved, whether for lack of information or for any other reason. If applicable, the provider will receive instructions on supplying the supporting information necessary for automated approval without manual review.
- Mid: These requests require the data noted above and, where available, EHR data, such as radiology results. When information is missing, AI algorithms will use the assembled information (family clinical history, allergies to certain drug classes, and so on) and make automated decisions.
- High: These cases need even more data and detailed analyses of patient history and potential outcomes, along with all the basic and EHR data. High-complexity determinations could be needed for skilled nursing facility discharges or for neurosurgical or neurological procedures.
- Very high: These cases—organ transplants or neonatal surgeries, for example—invariably require a manual evaluation.
The triage engine speeds up decision making and improves the provider experience in a number of ways. It makes suggestions in data fields by populating potential answers to questions, making it easier for providers to submit PA requests. As the data fields are populated, the engine validates the information for consistency with the member profile based on past claims. Should more data or a peer-to-peer review be required, the engine can also use intelligent scheduling and optimize communications channels (email, web application, or voice call) to reach providers. The engine can also optimally assign cases to medical director reviewers based on their availability, specialization, clinical expertise, and provider feedback. Finally, it can track progress of the PA decision and update the requester accordingly.
The authorization automation engine
Once a request is categorized by the triage engine, the PA automation engine—a combination of NLP and classification algorithms—provides the likelihood of success for different outcomes and speeds decision making. For example, the engine might determine that there is a 90 percent likelihood that a certain procedure would add little benefit to a patient, or it could provide guidance on when a clinician’s review is necessary. This would then be validated through a manual review by a medical director.
The PA automation engine applies the following elements to approve the request or to present the consolidated information to the payer’s clinical staff:
- automated eligibility determination by integrating the authorization request with the insurance plan’s eligibility database (depending on how these data are stored in a payer’s database, NLP may be required)
- clinical rationale based on historical PA decisions and past clinical outcomes, enriched with members’ historical claims data, comorbid conditions, medications, presence
of a specific diagnosis, family history, and payers’ medical policies (by sharing anonymized data among themselves, payers could further enrich the data pool, thus improving PA accuracy)
- well-rounded provider profile based on information such as provider utilization history, member outcome metrics, provider characteristics, and member experience metrics
- annotation and highlights of relevant information for review, including, for example, family history and explanations of data inconsistencies
- visualization of patient journey information, including dates of procedures and diagnoses
- suggestions for written responses, including requests for additional information, step therapy and alternative treatments, and supporting clinical literature
- integration with case management workflow and messaging platforms
The triage and PA automation engines work together to automate decision making, which could result in significant savings per case (Exhibit 4).
AI-enabled workflow requires fewer steps
Once established, AI-enabled workflow could reduce the number of steps in the PA process compared with current manual workflows (Exhibit 5). The new workflow will include the following elements:
- A PA request will enter step one after the triage engine classifies its complexity level. The request will be integrated with member eligibility and insurance plan details, and then dynamic rules aligned with the complexity level will make a decision. Algorithms will ensure any approval decision at this stage is final or near final with a high level of confidence. A sample of decisions will be reviewed for quality assurance and
- In steps three and six, NLP will be used on structured and unstructured clinical text and transcribed notes in the member’s EHR (with drug names converted to composites for ease of use in algorithms and to avoid brand bias) along with the payer’s own review and historical audit notes. This output will then be processed by an ensemble of algorithms for decision making that is at least as accurate as manual processes. The AI-based decision-making capability may be further enhanced by using member interaction data captured directly at the call center, through email, or, where available, from a payer’s mobile app used by the member.
Throughout the PA process, payers will be able to use NLP to extract information from EHRs, clinical notes, and medical policies in four ways:
- Benefits and cost-sharing information can be parsed to extract structured information. Algorithms can extract key terms used in PA decision making.
- Unstructured clinical notes and EHRs will serve as input data aggregated from numerous sources into a database.
- Unstructured text can be organized into structured data, with algorithms parsing information about diagnoses, procedures, symptoms, evaluations, and the like into a structured tabular format.
- Clinical terms can be converted to ICD-10
diagnosis codes by algorithms based on ontologies that are widely referred to in medical coding.
Challenges to overcome
While AI will certainly have a transformational effect on PA, several challenges will first need to be overcome. First, payers will need unfettered access to EHRs, which requires both strict compliance with data privacy regulations and a considerable design effort to ensure interoperability among various EHR software applications and platforms. At an operational level, industry participants will need to work together and define standard guidelines for attachments, data templates, and data exchange protocols as another important prerequisite for AI-driven PA automation. Various codes currently are applied inconsistently in communicating PA status and in the clinical documentation required to make a determination, which creates a greater administrative burden for providers and limits automation potential. Likewise, the approaches and regulations related to storage, control, and ownership of EHRs will significantly affect the degree of automation that can be achieved in the PA workflow.
When using AI self-learning features, it is also important to ensure that training data sets do not contain unforeseen biases that could result in unintended or inappropriate decisions. This includes closely monitoring PA decisions made for minority population segments and member groups of lower socioeconomic status, such as those on Medicaid. Ongoing review of AI models for biases will be important going forward. For example, metrics and reporting will be needed to continually track health equity in the decisions that are made.
While automation can improve efficiency and provider and member experience, workforce implications must also be taken into account. Automation of PA will likely free up staff, particularly nurses and clerical personnel, including those performing data entry. These staff members could be transitioned to higher-value activities such as care management—scheduling follow-up appointments, closing clinical care gaps, and so on—that directly improve quality of care and member experience.
Highly experienced clinicians will remain the ultimate PA decision makers, but AI can provide important decision support for both payer and provider clinicians while improving efficiency and enhancing the provider and member experience. By automating most PA decisions, payers can leave the most complex and sensitive decision making to highly experienced clinicians. To fully realize these benefits, the industry will need to work together to define a new series of standards to facilitate data exchange and to establish additional protocols for interoperability and integration across systems.