Atomic patient groups

The main concept of my thesis is to group patients of constant treatment complexity. I have a returning idea of creating atomic groups, a base granularity. Atomic groups would be the smallest units of the grouping process.

Atomic groups would be defined by a set of procedures, probably one or two. Every atomic groups should have some minimal size, let's say, about 50 patients.

Each atom could carry a regression-based model, explaining additional effects inside the atom. Especially, influence of the weight on the mortality. Each patient would carry the source atom signature and the predicted (according to the atom's model) mortality.

Summing up:

  • Find popular pairs of procedures
  • Find single popular procedures
  • Create a list of rules that assign patients to atomic groups
  • Following operations would be performed against the atoms

Patients’ weights are important

Basic idea of my thesis is to find groups of patients with constant complexity of treatment. How to find those groups is still an open question. K-means was very fast to find a given number of clusters. However, I have no idea, how much clusters should there be. Manual group creation according to combinations of procedures looks reasonable. However, the combination of procedures might be not enough to provide a constant treatment complexity. Binomial regressions show that patients weight can have significant influence on the result.

Continue reading “Patients’ weights are important”