A very insightful response from Fred Trotter. He lays down the options in a very nuanced and clear fashion. Much appreciated.
Tuesday, November 17, 2009
Thursday, November 12, 2009
Major Depressive Disorder
Perlis, Smoller, et al.,
MInutes courtesy Patience Gallagher.
· Longitudinal Classifier
o Roy and Victor will finalize parameters of algorithm
o Victor will then report (and provide a visualization of):
§ % depressed, % well, # of all notes, etc
o To query Crimson: Victor will provide a list of medical record numbers within the two groups of interest (responsive and resistant) to Lynn Bry who can then report how many samples are currently available within Crimson
§ Parameters may need to be readjusted, depending on response from Crimson
· Discussion of Validation
o The issue with last week’s approach:
§ The algorithm is based on the text, so the first level of validation should not use information (i.e. clinician’s extensive knowledge of a patient) that is not in the text.
o New validation plan:
§ STEP ONE:
· GOAL: Determine if an expert clinician’s classification (based on notes only) is the same as the result of the algorithm
· Pull successfully classified notes
o All of a patient’s notes will be reviewed
o Clinicians will be blinded to the results of the classifier
o The sample of notes will reflect the results of the algorithm: “Random but representational”
o To keep the validation clean, it will be a “case-control” model - Treatment resistant vs. Responsive
o For now, only the electronic medical record will be used.
§ The consensus was that patient charts would be more annoying than beneficial, and the electronic records probably have sufficient information
§ Additionally, the information from the paper records is not integrated into the algorithm
§ May look at paper records down the road
§ STEP TWO: Compare list of patients that Roy knows are treatment resistant or responsive and run their notes through the classifier
· (Tianxi says this will be beneficial as it will give more power to the classifier)
§ STEP THREE: Use the Quick Inventory of Depression Severity (QIDS) as an external source of validation
· Many patients have QIDS scores in their charts – can determine if the algorithm classification is consistent with performance on the QIDS – this would be a cross-sectional measure
o The output of this approach would be: “Among patients classified s depressed, the mean QIDS score is ____”
§ NEXT STEPS:
· Do first level of the validation over the next few weeks.
o Victor will pass the notes to Roy.
o Roy will be the sole clinician reviewing the notes
· Manuscript:
o Victor and Tianxi have provided their input to Roy
o Roy will integrate this information and then re-distribute the manuscript to the group
· PV
o Update from Victor on obesity:
§ Compared the BMIs of this data set with all other patients and the distribution of the MDD sample is very similar, but right shifted compared to majority’s BMI distribution
· This makes sense! - Being depressed (and on antidepressants) leads to weight gain