A cure for costly tobacco usage nondisclosure?

Verisk’s innovative voice analysis could save insurers millions in lost premiums

Verisk’s Tobacco Usage Propensity Model offers the life industry a two-prong solution: It can assist in substantially reducing the number of lab tests insurers order in the underwriting process, while simultaneously helping tackle the $10 billion problem of smoker nondisclosure.

In this Movers & Shakers Q&A, we interview Nick Irwin, director of life insurance underwriting products, about Verisk’s patent-pending voice technology that pairs with other data assets to help identify, with a high degree of accuracy, individuals who may have misreported their tobacco usage status when applying for life insurance.

What spurred the development of Verisk’s Tobacco Usage Propensity Model?

Tobacco usage nondisclosure has been a long-standing issue in the life industry, and many models have been developed historically that attempted to address this issue. Verisk has estimated that the life industry loses $10 billion of premium every three years from tobacco usage nondisclosure, and nearly half of all tobacco users fail to disclose their true status on their applications. In a survey we conducted with 15 insurers and reinsurers, tobacco use nondisclosure was confirmed as the respondents’ No. 1 issue.

Does Verisk’s Tobacco Usage Propensity Model use a science- and evidence-based approach?

Because our survey identified a clear need to address this nondisclosure problem, we invested quite a bit of resources in developing this technology. We started with multiday strategy sessions with our data scientists. When ideas began to surface of using voice to identify tobacco users, we brought on some medical doctors to further research the issue. They brought forth a number of peer-reviewed and evidence-based clinical studies that have documented smoking-based dysphonia — changes in the voice as a result of prolonged smoking.

Studies conducted in 2015 and 2017 both reported specific, significant variations in vocal patterns for smokers versus nonsmokers. In particular, the phonation of the sound “ah” is quite different for a smoker versus a nonsmoker. Our model gets that with a high degree of accuracy. Plus, we make quarterly updates to our model, and we are seeing more highly tuned performance with each iteration as we continue to collect more data to refine the model.

How does the Tobacco Usage Propensity Model work?

The model leverages our patent-pending voice technology combined with several other variables to estimate the likelihood that an individual is a tobacco user. Those who receive a high likelihood score can be flagged by an insurer for possible traditional lab testing or further investigation. The voice model is actually an ensemble of several models leveraging both classical audio analytics techniques that closely mimic how humans perceive sound and modern techniques leveraging convolutional neural networks. The model has been trained and validated on data from Verisk-commissioned focus groups as well as from data collected from pilot insurers.

Is it relatively easy for insurers to fold the model into their current processes?

Voice samples are already routinely collected through the telemed process at many insurers, so for most of our users, no additional data collection is required. We do offer a vocal attestation web app for carriers without telemed. And because our voice model measures physiological changes rather than pure socioeconomic correlations, it can also help lower exposure to regulatory risks or concerns such as those related to unconscious bias.

By unlocking the power of voice analytics, Verisk has paved the way for an entirely novel method of overcoming smoker nondisclosure.

How would insurers incorporate Verisk’s Tobacco Usage Propensity Model?

For insurers that have telemed, our model requires a simple API [application programming interface] integration. The recording of each customer interaction with the call center gets submitted to Verisk’s model for scoring. The resulting score is then used as part of an insurer’s underwriting rules engine to flag applicants for lab testing or further investigation.

For insurers without telemed, we offer a mobile web application. Applicants complete a voice attestation, and Verisk handles all requisite API calls within the app itself. Insurers
without telemed may find that implementation is even simpler. After consumers submit their applications, the applicants will receive a text message with a custom link. The applicants then click the link, which will direct them to a mobile web app, where they will record a 45-second statement, essentially certifying that information on their application is accurate to the best of their knowledge and they permit the usage of their voice for underwriting purposes.

Does the model benefit insurers and applicants?

Our model can help reduce the number of lab tests required as part of an accelerated underwriting program, which is especially important in the age of COVID-19, when obtaining labs is particularly inconvenient and, in some cases, distressing for some consumers. It’s a “win-win” for customers and insurers alike, as customers can bypass the inconvenient lab test process in most cases, and insurers may have greater protection against tobacco usage nondisclosure. Our model only identifies a small percentage of total life insurance applicants for laboratory testing, with the remainder often being allowed to bypass traditional laboratory testing requirements.

How was Verisk uniquely positioned to develop this model?

Verisk has hired a global team of some of the world’s leading audio analytics experts. Moreover, we have nearly 50 years’ experience in developing insurance products, including models, that comply with relevant statutes and regulations. We have an extensive legal, regulatory and compliance team to help review, monitor and evaluate changes in the regulatory landscape. Our models within the life program are evaluated for unconscious bias by testing performance on an annual basis by ethnicity, religion, gender and age, and are revised for findings from the internal review process.

For example, early in design, we found that males tend to disclose the incorrect smoker status more frequently than females do, so we built separate models for males versus females to help ensure that the performance didn’t vary materially by gender.

Are your clients seeing success with Verisk’s Tobacco Usage Propensity Model?

Based on our pilots, our model is estimated to provide an annual net present value benefit between $15 million and $30 million, or a 15% rate of return for an insurer that processes 100,000 individual life insurance applications per year through an accelerated underwriting program. Our model costs are only a small fraction — approximately less than 3% — of the economic value provided to our customers. The solution became available for pilot in October 2019. In September 2020, we made the tool available for use by our customers in a production setting. This solution is part of Verisk’s suite of end-to-end life software and analytical offerings, from underwriting to portfolio analysis to claims.

Discover how Verisk’s audio analytics and a short snippet of an applicant’s voice can reveal hidden tobacco use. Download Verisk’s white paper.

1. “Current Cigarette Smoking Among Adults in the United States,” Centers for Disease Control and Prevention, accessed on December 19, 2019.

2. Brian Lanzrath, et al., “Applicant Medical and Smoking History Nondisclosure in the Life Insurance Marketplace,” Contingencies, November/December 2016, accessed December 13, 2019.

3. Kavitha Hariharan, et al., Insurer Perspectives on Smoking Risks, Marsh & McLennan Advantage, 2019, accessed on January 29, 2020.

4. Tobacco fact sheet, World Health Organization, accessed on November 2, 2020.