Project Objective

As a revenue cycle management vendor, the client wished to deploy predictive modeling to assess the likelihood of claims getting invoiced, understanding drivers of claim denials, and deploy mitigation strategies, thereby ensuring efficient claims management for seamless patient experience and maximized revenue recovery for healthcare providers.

Approach

To train the machine learning model, we used data about patient admissions, census data, and mortality rates to model the likelihood of claims getting invoiced. The dataset was further enhanced by incorporating mortality rates of various diseases, providing additional context that could influence the invoicing of claims. Moreover, engineered features such as the season, day of the week, and time since the last visit were added to capture historical patterns in the dataset.

We created two separate models for the inpatient and outpatient cases, which allow us to better understand the different characteristics and patterns. For instance, inpatient cases often involve more complex procedures and longer stays than outpatient cases. In addition, the differences in the care processes can lead to significant differences in costs. These cost differences can influence the likelihood of a claim being invoiced. Having two different models can account for this contrast, and we can draw better conclusions. The relative advantage of our approach is visible when comparing the importance of various features for the two models. On one hand, in the inpatient model, age of the patient, average delay between two visits, and whether the patients commute using public transport are the most important features in determining whether a claim will get invoiced. While in the outpatient model, the average delay between visits, ratio of claims approved, and whether a household makes over $200K annually are the most important features in deciding claims. It would not have been possible to understand these nuances with a single model for both the cases.

Outcome

The inpatient model achieved a training accuracy of 82.16%. The model’s AUC-ROC score of 92.8% indicates its strong ability to distinguish between claims that are likely to be invoiced and those that are not. Similarly, the outpatient model demonstrated a training accuracy of 81.64%. Its AUC-ROC score of 77.3% suggests that though slightly lower than the inpatient model, it still effectively differentiates between claims that will and will not be invoiced. These results underscore the models’ robustness and their potential to enhance revenue cycle management by accurately predicting claim invoicing.

Way Forward

Further to this, we will also be developing a patient-responsiveness-to-outreach model to identify the best channels to engage with patients for the high risk claims to ensure fast resolution of their claim denials and increase conversion rates. Additionally Finarb will also develop a robust MLOps pipeline that facilitates seamless model training, hyperparameter tuning and deployment.