Contrastive Methods

[Oufattole]

We want to support general contrastive learning window definitions.

There are some common patient specific latent space temporal structures we may desire (as shown in Figure A).

• Consistent - that a patient has a constant representation if we look at two different windows of time for them. This can be local (only for adjacent windows) or global (for any two randomly selected non-overlapping windows)
• Continuity/interpolation - that if we take two non adjacent windows for the patient, the window between them should be approximately an average of those two windows.
• Ordering - that if we reverse time, or shuffle windows and input those to a model, the representation will be very different. This can also be local (adjacent windows) or global (no adjacency constraint).
  We can also have multi-patient properties:
• Label-Based Neighbors - patients with the same label should have similar representations

Additionally, we want to allow users to select global and local windows a