Additional Information on Claims Dataset

[robi86]

For explainability purposes some recent papers have shown that stacking embeddings learned from claims/EHRs of different care settings can be valuable. Intuitively this also makes sense as the distribution and co-occurrence of diagnoses codes is different in the hospital versus the clinic. As such, can you provide additional information on the claims dataset used to learn these embeddings including the following:

• Were these claims for commercial insurance or from public insurers (e.g., Medicare/Medicaid)?
• Were these claims from a mix of care settings, inpatient only, or outpatient only?

[pandeyp]

Hi @robi86, can you explain how we can stack/sum claims if a member has multiple claims lines that have multple ICD10 diagnosis and procedure codes in claim lines. I am trying to generate word embedding representation on a member-by-member basis looking at all the claim lines available for a member.

[temiwale88]

@pandeyp. This is obviously late but might be helpful: Maybe try generating the embeddings (which are real-valued vectors) for all ICD10 diagnoses and procedure codes, i.e., per claim line (or for all claims per member), then summing (sum-pooling), averaging (mean-pooling), or max-pooling over all the vectors / embeddings? Any of these pooling strategies will generate some representation of that patient's service and utilization history.