Detecting Self-Reported Memory Problems in Transcripts of Interviews with Older Adults
Open Access
- Author:
- Lu, Xi
- Graduate Program:
- Informatics
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- July 07, 2020
- Committee Members:
- Pransenjit Mitra, Thesis Advisor/Co-Advisor
Ting-Hao Kenneth Huang, Committee Member
Shomir Wilson, Committee Member
Mary Beth Rosson, Program Head/Chair - Keywords:
- Discourse Analysis
Healthcare Text Mining
Alzheimer's Disease - Abstract:
- Alzheimer’s disease has an insidious onset. Throughout the disease’s development, there is a time when problems with memory are perceived by affected older adults but are barely detected by clinical screening measures. Researchers are thus studying the “self-reported memory problems” via the structured probing interviews then coding the transcripts manually, aiming to improve preclinical Alzheimer’s disease detection. However, manually analyzing interview transcripts is laborious and time-consuming: coding the transcripts of 21 hours’ worth of interviews would take more than 120 hours. To reduce the workload of researchers and thus make large-scale qualitative analysis possible, this thesis work presents the first exploration of identifying self-reported memory problems in interview transcripts automatically. Equipped with 96 inter-view transcripts collected from 48 cognitively intact older adults, this thesis work developed a stacking ensemble model that achieves an average F1-score of 77.3% and an average accuracy of 77.7% in the detection task, outperforming all the strong baselines. This thesis work also demonstrates a procedure that repurposes the expert-annotated transcripts for machine-learning algorithms.