User Tools

Site Tools


timeline

This is an old revision of the document!


Denmark Project Timeline

Back to the main Denmark project page

Fall quarter 2012

1 - implemented the trajectory stitching using simulated cost data 2 - implemented 3 different matching mechanisms for trajectory stitching 3 - presented results at end of quarter (Dec 12th)

Winter quarter 2013

1 - Engaged with Tommy (first meeting Jan 15th) 2 - decision to proceed with the “shopping cart” method using only the 8 years of (unstitched data) 3 - started quest to get actual cost data, and data dictionaries from Aarhus (Feb 5th)

  • requested 1-year bins on demographics file instead of current 3-year bins
  • started getting cost information for drugs and admissions (possibly DRG-classified costs)

4 - visited Denmark / Lars to actually get the data we wanted (we hoped to do it remotely, but nothing beats sitting side by side)

Spring quarter 2013

1 - started on shopping cart models 2 - Both patient-level and population-level validation runs for a test set of 1000 patients (validation, 5/6) 3 - Also tried, probabilistic approach to trajectory stitching (didn't work). 4 - decided to focus on a certain set of diseases going forward (diabetes / CHF / CAD / COPD) 5 - started clustering and template matching as other alternative methods (Ken Jung's work) 6 - focused all methods on one goal: find “bifurcations” in the 8 cost trajectories.

As of May 15th, we have narrowed the scope to: 1. simply identifying patients with some positive control conditions .. where we believe that bifurcations should exist. 2. finding patients of interest (with respect to their cost trajectory) 3. given 2, determining to first order (i.e., without trying to find intervention points) what drives higher costs in patients with breakpoints in their trajectories

Summer quarter 2013

1. Characterized cost patterns and concluded that outside of end of life, high costs are concentrated in acute episodes rather than chronic elevated outpatient costs. 2. Identified high cost subset of patients (defined as >= 90th percentile of total expense) and high cost episodes (defined as >= 90th percentile of annual costs). 3. Ran association rule mining (apriori) to identify simple combinations of codes that predict high costs (either definition).

Stanford-Aarhus meeting in late August 2013

timeline.1390370331.txt.gz · Last modified: 2014/01/21 21:58 by stamang