What if AI could save CM & UM 40% to 60% of their time by automating data collection and aggregation?
Inextricably Intertwined Value Propositions – Care Management & Utilization Management
Strong care management programs can be a tailwind propelling the effectiveness of health plans’ utilization management programs. Care management and care coordination are foundational to patient-centered medical home model of care (NCQA 2020) – a strategy shown to be effective in managing utilization and thereby creating value in health systems via reduced costs, ED visits and hospitalizations (Veet et. al. 2020).
Conversely, ineffective utilization management can counteract care management efforts by inciting more visits, tests and avoidable acute events — all additional unnecessary expenses had the decision been right the first time.
AI & NLP create the shared knowledge base, derived from the entire digital informational universe on members, providers and clinical practice — serving CM & UM team members what they need to know, when they need to know it.
In this article, our 3rd in the series, we will review Utilization Management & Care Management’s critical functions augmented by NLP & AI for substantial ROI. We’ll look at:
- Determining whether a member’s case meets Level-of-Care Criteria;
- Matching evidence guidelines and rules to a member’s case;
- Connecting a member’s case and pathway to care coordination; and
- Augmenting concurrent review and retrospective review with automated decision support
Using NLP and AI to Determine Level of Care (LOC)
Payor utilization review nurses perform highly repetitive tasks that can be augmented and automated by AI & NLP. One of the most repetitive functions is determining whether a case meets the level-of-care criteria for a particular treatment. The payor nurse takes in data, finds and applies the criteria rules and makes a decision to either approve, deny or refer for physician review.
LOC criteria are well-researched with a strong evidence base. What diagnosis, symptoms, behavior or functioning warrants a particular service? What are the expected outcomes from the treatment? To demonstrate this, hospitals and physicians submit documentation read by the payor UR nurse – searching for this information.
In addition to the criteria, there are protocols – actions required to be taken in certain identified situations to ensure a good diagnostic-treatment fit. Certain combinations of symptoms, patient history, behaviors, and pre-existing conditions determine a particular pathway of appropriate treatment and care setting.
LOC criteria and protocols are two areas where it’s highly unlikely that a payor will see justifiable variations or deviations. The high specificity, regularity, volume and standardization make these tasks strong targets for applying AI and NLP. Inpatient stay vs. Observation. Post-acute setting vs. home health vs. home. These decisions are made with such greater frequency and with such voluminous data that machine evaluation proves exceptionally accurate and efficient.
Ultimately, identifying and aggregating all the information for the case including policies, guidelines and supporting information represents 40% to 60% of UR nurses’ total time (Schatsky et. al. 2015). Machine learning models can be automatically applied to all the documents in a member case and key data surfaced that indicate level of care for the UR nurse. Instead of the payor UR nurse facing a stack of PDFs, AI & NLP can deliver a summarized extraction of the documentation:
- Diagnostic results
- Standards of care
- LOC criteria
- Match to applicable rules / guidelines
- Recommended decision
The UR nurse may confirm all this information and adjust the documentation accordingly. Taking the nurse’s feedback and final decision into account, the machine’s models will improve over time.
The same models that identify symptoms, diagnostic results and conditions then serve as the basis for matching the case to probable protocols and pathways.
Matching Member Cases to Treatment Guidelines
Practice Clinical Guidelines – What is known to be the best care procedures given the current state of knowledge. Practice guidelines inform LOC criteria and approval policies but do not alone account for authorization criteria. Frequently, payors will utilize the guidelines established by external independent institutions such as the NCCN for oncology. Great evidence substantiates the efficacy of guideline adherence. To continue with oncology as the example, members whose treatment deviates from NCCN guidelines results in significantly higher utilization, costs and mortality risk (Rocque et. al. 2018; Williams et. al. 2019).
Guidelines can frequently be ambiguous — and this is where an AI-driven decision support can be truly valuable. First, by automating the easiest decisions, the UR nurse will have more time to spend on the more challenging ambiguous cases. Second, by embedding AI decision support models that learn and train themselves within the UR process – the ROI in subsequent years will be substantial as improvement over time enables greater precision and gradually decreases ambiguity. Even without the AI component, decision support augmenting UM has demonstrable ROI with both financial and clinical outcomes (Powell et al. 2018, Baron & Dighe 2013)
Downstream Reviews & Appeals
The appeals process varies from payor to payor, but whatever the level, the information gathered during all previous steps will be invaluable to substantiate a rapid and accurate decision.
Care Coordination & Utilization Review (UR) share many aspects, including:
- Pre-authorization – done at initial point-of-service for low, brief LOC
- Prior Authorization – particular service for a particular amount of time
- Concurrent Review – determine if a service needs to continue
- Retrospective Review – determine whether the record justifies the service already delivered
The use of aggregate information collected from these reviews is a key part of utilization management and is also used for quality management.
Care coordination or care management focuses on the long-term recovery of the individual to cope and function independently to manage symptoms, disease etc.
Overlap occurs in many tasks — such as aggregating information, reading documentation and extracting key information to determine eligibility, criteria and guideline application. Comprehensive risk assessments are frequently utilized in clinical guidelines and protocols as well as the care management process. Each of these functions can be performed on the front end by a machine and sent downstream to both CM & UM.
Benefiting from the longitudinal holistitic member view discussed in the first article of our series, machine learning models can make connections and identify relationships between disparate variables that frequently escape human observation.
Especially in the age of value-based care contracts and ACOs, cross-functionality in models, applications and identification of potential high risk scenarios delivers improvements to efficiency and member health.
Make All Communication Machine-Readable & Predictable
With a great portion of documentation provided by physicians still in PDF, machine learning driven OCR can process PDFs — even handwritten and image based PDFs — with great accuracy. Automatically extracting this text data, applying models and delivering into the ideal environment will save UM & CM significant time.
Even greater benefits can be achieved by automatically transcribing the audio of CM & UM phone calls, then utilizing those for documentation, verification and pattern indicators.
At Digital Reasoning, we’ve built proven universal ingestion data pipelines to create holistic views for analysts in the healthcare, banking, law enforcement and government sectors.
Outcomes That Matter Most
We see tremendous benefits of AI & NLP applied to augment and strengthen UM & CM:
- Faster time to decisions
- Higher performance consistency, lower deviation from guidelines
- Increased productivity
- Increased care management capacity
- Better health outcomes
- Lower costs
Predicting Denial or Approval
Collectively, this information and the results of all of these models play into a comprehensive Denial Predictor. Predicting the approval or denial of a case by the UR nurse, physician, or other medical reviewer can filter out the complex cases and accelerate decision-making time (Aruajo et. al. 2013). We will explore this cognitive computing application and opportunity in greater detail in the next article of our series.