PPATKickOffMeeting2012-01-12

From IMC wiki
Jump to: navigation, search

notes

Data

Currently only have clinicians' ratings as measure of adherence. Good/Average/Poor, but av/poor were collapsed for statistical analysis as so few. Could look at pill counts, meeting attendance as less subjective measure, but may be difficult as this kind of information

138 dialogues available, different patient in each one. About 30% average/poor adherence

Transcripts in Word format, but most annotations (incl all clinician repairs) are hand-written on paper copies - must be transferred to SCoRE. Rose may have assistants who can help with the boring tagging part

Once transcripts processed & cleaned up, Chris will send list of problems found back to Rose/Laura

Anonymity: Transcripts to be on SCoRE but password-protected; most already anonymized but we need to check each one as we annotate

Pat: do we need to annotate spans of talk rather than whole turns? Many repair classes will apply to only sub-parts of turns, with many different classes appearing in the same turn

Is there video/audio available? We could run ASR to generate automatic transcripts for later experiments. Yes - access needs to be arranged through Rose. Matt to look for batch audio extraction tool for MPEGs.

There are third parties present in some dialogues (e.g. relative, carer)

Approach

Some more simple measures might be useful: amount of talk per patient etc. Chris can generate these kinds of features easily if Rose/Laura can suggest them

Laura/Rose looking at agreement & disagreement - we could think about applying the same techniques to detect those later

Pauses, inter-turn timings might be important features. The transcripts won't have this to any accuracy, but ASR followed by alignment might be a way

Plan is to train classifiers to predict repair classes in the first instance; but should also look at predicting adherence category directly. That would also allow us to discover other features which might be predicting adherence, other than repair.

Features of doctor's talk in particular would be interesting to look at. Perhaps frequency of yes/no vs open-ended questions? Also doctor 1st-pos repairs correlated with patient satisfaction

Rose/Laura know this data well, and may have some other strong hypotheses about important phenomena - if can pass to Matt/Chris can think about how to code/extract them

Might also be interesting to look at patient questions other than CRs/NTRIs

Plan

Laura to ask Julian about presenting at IMC Reading Group, perhaps 2nd Feb?

Possible conference paper dates: CogSci probably too early, SIGDIAL good but not REF-able, ACL short paper date? Also need to think which medical conference.

Chris to work out SCoRE extensions to cope with transcripts, and visit Newham 23rd Jan