Automatic assessment of motivational interview with diabetes patients

Wei, Xizi (2022). Automatic assessment of motivational interview with diabetes patients. University of Birmingham. Ph.D.

Preview

Wei2022PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (5MB) | Preview

Abstract

Diabetes cost the UK NHS £10 billion each year, and the cost pressure is projected to get worse. Motivational Interviewing (MI) is a goal-driven clinical conversation that seeks to reduce this cost by encouraging patients to take ownership of day-to-day monitoring and medication, whose effectiveness is commonly evaluated against the Motivational Interviewing Treatment Integrity (MITI) manual. Unfortunately, measuring clinicians’ MI performance is costly, requiring expert human instructors to ensure the adherence of MITI. Although it is desirable to assess MI in an automated fashion, many challenges still remain due to its complexity.
In this thesis, an automatic system to assess clinicians adherence to the MITI criteria using different spoken language techniques was developed. The system tackled the chal- lenges using automatic speech recognition (ASR), speaker diarisation, topic modelling and clinicians’ behaviour code identification.
For ASR, only 8 hours of in-domain MI data are available for training. The experiments with different open-source datasets, for example, WSJCAM0 and AMI, are presented. I have explored adaptative training of the ASR system and also the best training criterion and neural network structure. Over 45 minutes of MI testing data, the best ASR system achieves 43.59% word error rate. The i-vector based diarisation system achieves an F-measure of 0.822. The MITI behaviour code classification system with manual transcriptions achieves an accuracy of 78% for Non Question/Question classification, an accuracy of 80% for Open Question/Closed Question classification and an accuracy of 78% for MI Adherence and MI Non-Adherence classification. Topic modelling was applied to track whether the conversation segments were related to ‘diabetes’ or not on manual transcriptions as well as ASR outputs. The full automatic assessment system achieve an Assessment Error Rate of 22.54%.
This is the first system that targets the full automation of MI assessment with reasonable performance. In addition, the error analysis from each step is able to guide future research in this area for further improvement and optimisation.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):