I have a list of patient accounts based on their discharge date. I have various inputs related to each patient such as their financial class, insurance information, demographics, claims information, etc. I am ultimately trying to better understand the complex patterns that may be taking some patient insurance companies longer to pay. First, I want to understand what rules the random forest algorithm determines from the dataset because there’s potentially thousands of inputs after one-hot encoding which causes the underlying patterns to be too much for the average user to interpret without using machine learning. An example use case, would be to help identify why similar type accounts (similar procedures and insurances) would have one account with a payment in 5 days and another “like” account in 27 days. The issue is my dataset could have multiple payments for each account and that introduced bias, I believe.
So I have 2 questions. Originally I thought about clustering the like accounts and then trending the clusters with their average payment age over time. But because of the high dimensionality of the dataset, even using tSNE or PCA, I couldn’t get it to work. So I’ve moved to using a decision tree based approach to at least help understand the patterns.
How should I handle the duplicative entries for each account in my dataset? If I only look at the first payment for each account then I would love the ability to see payment plan type entries or when insurance companies are paying in installments. Similar issues would exist if I only look at the last payment age for each account.
Does anyone have any advice? I’m also open to an entirely different approach given my stated goal is to ultimately help understand why similar accounts have such varying payment ages.
Any and all help is welcome. Thank you!!!
