The ultimate guide to predictive analytics in P&C insurance claims

Predictive analytics have opened a world of possibilities in the ways marketing, underwriting, and claims management are executed and managed today. Driven by sophisticated artificial intelligence (AI) technologies like deep learning, neural networks, and machine learning that mimics human thought, the field has gained a nearly magical quality. This guide is intended to demystify the topic and bring it easily within operational reach to claims professionals.

Background

Predictive analytics attempt to establish relationships among variables or characteristics in order to predict future outcomes. But the process has long been hampered by access to limited forms of data in legacy systems or overwhelmed by the massive effort typically involved in trying to make connections among a wide number of variables. These obstacles have largely been overcome in today’s predictive analytics, which rely on technologies that use sophisticated algorithms to identify patterns in large data sets to establish relationships.

Used in the context of claims management, predictive analytics can segment or triage claims, prioritizing potentially high-cost claims early in the process for cost containment or fast-tracking low-cost claims for settlement. What had been a sometimes hit-or-miss and often labor-intensive claims identification process can now be much more data-driven and efficient with the use of predictive analytics.

The tools

AI uses computers to mimic the human thought process to solve problems. Machine learning, a type of AI, allows computers to understand patterns in data and perform tasks. Often, the chosen task is to predict a future outcome, which is called a predictive model. A predictive model learns the relationships between input and output using historical data. In the case of claims management, a predictive model can learn the relationship between different features like body parts, attorney involvement, and/or location to predict an outcome such as the cost or severity of the claim. The models can also predict a variety of other outcomes like risk level, surgery, or attorney representation. These types of predictive insights enable adjusters to intervene throughout the claims life cycle.

Because so much data in claims systems is unstructured text data, natural language processing (NLP) and deep learning are used to interpret this data. By using NLP, the computer can intelligently extract relevant information from long text data such as adjuster notes to feed downstream predictive models. Deep learning is a branch of machine learning, which attempts to mimic the behavior of the human brain. Deep learning excels on tasks that depend on very large data sets such as NLP, computer vision, and speech recognition, among others.

The data

Claims departments have avoided exploring the application of AI over concerns about missing or incomplete information in their data. But with the use of advanced NLP techniques, information can be accurately extracted from the unstructured text to create high-quality structured data sets. For example, hard-to-access information like comorbidities, which have the potential to cause a claim’s costs to rapidly escalate, can be extracted using these algorithms.

A comprehensive knowledge of predictive modeling techniques and the claims life cycle is critical to the development of accurate models. Equally imperative to development is the support of claims professionals, actuaries, data engineers, and data scientists, who can work together to create and evaluate a predictive model and use their abilities to improve claims performance.

AI effectively bridges the gap between noisy, messy data and predictive analytics in insurance.

Output

Predictive modeling output is often an estimated probability, dollar amount, or score. The exact output varies based on the objective and the stakeholder. Some models can also provide insight into the features that drive the prediction itself, providing context to the user. If a model is perceived as a black box, its output can seem less credible to users and its predictive insights are less utilized.

Uses

There are a wide variety of uses for predictive modeling in claims departments. The list below is therefore not exhaustive but provides several key applications.

• Claims segmentation
  Predictive analytics can identify high-cost claims for intervention shortly after being reported, the assignment to an experienced adjuster, settlement opportunities, or other cost containment strategies. Low-cost claims can be fast tracked and expeditiously closed, saving claims administration expenses. In many property and casualty (P&C) lines, the top 5% to 10% of claims represent more than 80% of claims costs. By segmenting claims, predictive analytics can improve claims triage, promoting a more efficient and data-driven allocation of claims resources.
• Legal analytics
  Using the unstructured data from defense attorneys’ invoices, predictive analytics can identify high-performing defense firms on a risk-adjusted basis, giving carriers the ability to make data-driven decisions in assigning outside defense counsel.
• Medical benchmarking
  In workers’ compensation, predictive models can also identify the likelihood of overspend on medical treatment. Overspend occurs when workers’ compensation payers overutilize and overpay for medical services for similar injuries, relative to the more rigorously managed group health.
• Data enrichment
  With access to unstructured data, AI also delivers a deep level of granularity into a carrier’s claims data, a form of new intelligence to drive strategic decision-making. AI enables executives to answer questions using unstructured data that was not possible with only structured data and to see emerging claims trends as they are occurring such as during the pandemic.
• Fraud detection
  Predictive analytics can also help automate the identification of suspicious claims for a special investigative unit (SIU). Rather than relying on adjusters to manually flag claims, AI can proactively identify claims as new information is gathered on the claim.
• Subrogation opportunity
  Predictive analytics can identify subrogation opportunities. Rather than using manual processes to identify recovery opportunities, predictive models can automatically flag claims that exhibit characteristics of a subrogation opportunity and provide an estimate of the amount of recovery.
• Attorney representation and litigation potential
  Claims with attorney representation and/or litigation potential have both high cost and risk. Predictive models’ early identifications of these claims present opportunities for insurers to resolve claims quickly and avoid uncertainty and unnecessary defense costs in cases with no liability.

Benefits

Predictive analytics provide tangible cost savings in the form of:

• Lower claims indemnity and medical severities by intervening earlier in problematic claims
• Lower claims administration expenses by eliminating wasted spend on low-cost claims
• Lower defense costs as caseloads can be directed to more efficient outside defense counsel and away from less efficient firms over time
• Faster cycle times by focusing on early resolution
• Less volatility in claims development, resulting in more stable projections

Operational benefits include:

• A more data-driven decision-making process that helps automate the claim’s assignment process by allowing claims managers to more efficiently direct claims to the appropriate resources
• Provide more information, guidance, and validation for long-term claims management strategies, validating processes or initiatives that the company has put into place
• An improved claims settlement process from AI’s ability to use unstructured data to identify shortcomings in data coding, possible areas where information appears underreported, or inconsistencies in coding, each a signal to a claims team of a need to review certain procedures
• Greater efficiency, especially compared with manual processes, from the ability to auto-process low-cost claims or assign potentially high-cost claims to experienced adjusters who can direct the appropriate cost containment measures to these claims early in the cycle when they have the most impact
• Increased confidence in the ways in which data is interpreted because of the use of AI technologies that consistently apply statistical frameworks in the same way across large data sets to develop relationships among variables
• Increased visibility into claims data that can provide better insight into company-specific factors that drive costs
• On-demand access to a granular layer of data that can reveal changes in claims trends as they emerge

Other considerations

Predictive models look for patterns by comparing characteristics of outstanding claims with those of closed claims. An underlying requirement in the model’s development is the need for a robust database that spans multiple economic and insurance cycles and is appropriately balanced with respect to product lines and the age of the claims used in the model. While AI can overcome many shortcomings in a carrier’s claims data, the data used by the modeler to develop the predictive analytic platform needs to have integrity.

Incorporating demographic, meteorological, or other external data in the model can help to improve the performance of a predictive model in some product lines, but the improvement is minimal for a line like workers’ comp, where the characteristics of claims—what type of treatment the claimant received or is scheduled to receive—is more germane to the predictability of a claim’s outcome than, for example, the location of a claimant’s residence.

A predictive model’s accuracy is important, but it is only one measure of a model’s value. Models that have been overparameterized by tightly fitting the model to historical data can produce highly accurate results at first, but as new and different claims data becomes part of the database, the model’s accuracy typically erodes over time. Developing a way to track the model’s performance after implementation against predetermined key performance indicators (KPIs), for example, is instead a more useful way of assessing the model’s performance. A good model will typically improve over time as the carrier’s claims data collection process improves and information from "false-negative" outlying high-cost claims is fed back into the model.

It is also important to consider the level of support that is provided during and after implementation. Access to claims professionals who were part of the development team can provide advice with a model’s deployment and the best ways to use and track the model’s output.

Developing the caliber of expertise in AI technologies to develop a predictive model can tax the resources of even the largest carriers. New algorithms and hardware are continually emerging. Keeping pace with these advances while maintaining a competitive edge in a changing risk market can be a tricky balance for the most agile carrier.

Predictive analytics have steadily become an integral part of many carriers’ operating toolbox. Once suitable for only the largest carriers, they have made their way into claims operations where one person may wear many hats. While predictive analytics will never replace the expertise that claims adjusters bring to the settlement process, they have become a driving force in carriers’ efforts to reach new levels of efficiency and competitiveness.

About Nodal

Milliman’s Nodal is a predictive model for early claims intervention and cost reduction.

Nodal uses advanced AI technologies to identify high-cost and low-cost claims soon after reporting, allowing for efficient triage of claims and allocation of resources that maximize the use of staffing and cost containment strategies. Milliman’s team of actuaries, claims professionals, and data engineers has developed an end-to-end solution that is fully supported through implementation, deployment, and assessment.