OBJECTIVES: To demonstrate the feasibility of using natural language processing (NLP) and machine learning to predict whether a member contacting the call center of their medical insurance carrier was likely to be dissatisfied.

METHODS: Two models were evaluated. The member-level model looked at Blue Cross Blue Shield of Louisiana (BCBSLA) members who called BCBSLA customer service at least 1 time during the 2016 calendar year and used January 1, 2017 through March 31, 2017 as a target for predicting member complaints. The inquiry-level model used inquiries from the 2014-2016 calendar years and looked to see if an inquiry had a complaint within 30 days. Twelve-month claim and member relationship history were used to generate a member and an inquiry complaint propensity (MCP and ICP). Scores were used to categorize members into high-, medium-, and low-risk groups. Three cohorts for analyses were defined: (1) patients actively engaged in case management and disease management (CM/DM) programs, (2) patients enrolled in CM/DM but do not participate, and (3) patients not enrolled in any CM/DM program. Several models were tested by varying covariates and interactions.

RESULTS: Of 336,000 members who contacted BCBSLA customer service in 2016, 9,000 (2.6%) made a complaint within the next 3 months. Of 2,000,000 inquiries made (2014-1016), 34,000 (1.7%) made a complaint within the next 30 days. Both the MCP and ICP had positive predictive values. Members with a high-score from the model had a 1 in 5 chance of complaining again compared to a 1 in 39 chance of a random complaint. The best fitting model accurately predicted member and inquiry complaints based on prior data.

CONCLUSIONS: While not currently in practice, NLP predictive modeling tools have a future in the payer-provider-patient realm because they can decrease administrative and medical costs and increase member satisfaction and engagement.