OBJECTIVES: Reducing and preventing avoidable hospitalizations for chronic and acute conditions through model-based interventions is a key interest at health plans to improve outcomes and lower healthcare costs. By using survival analysis, it is possible to assess different treatment options that can significantly reduce a member’s risk.

METHODS: BCBSLA used a machine learning model based on hybrid Cox Proportional-Hazards and neural networks to analyze a member’s risk of hospitalization (ROH). The duration comes from the first occurrence of high-risk score, and the censor describes if the member never became high-risk for hospitalization. The established threshold was 10%, meaning that the treatment must reduce risk by at least 10% or it will not be considered. Once researchers calculated the hazard ratios for all treatments, they chose as the optimal treatment the one that yielded the most favorable ratio while also satisfying the threshold requirement. Otherwise, “No Treatment” becomes the optimal treatment.

RESULTS: Results consisted of the members and their recommended treatment, their baseline hazard, and their hazard ratio for the recommended treatment. The model recommended a treatment of some type to 90% of members, meaning that almost 90% of members can benefit from a treatment that can reduce their risk of generating a high ROH score by 10%. The remaining 10% of members did not receive a recommended treatment. The average hazard ratio for those recommended treatment was 60.67%; this means that the recommended treatments reduced a high ROH score by 39.33% on average.

CONCLUSIONS: This model shows promise in recommending actions and treatments for members, while also quantifying the potential benefits through an easily interpretable statistic.