DeconDTN: Deconfounding Deep Transformer Networks for Clinical NLP

点击翻译按钮获取中文摘要

Abstract This proposal relates to ongoing efforts to develop automated methods for the detection of linguistic manifestations of cognitive changes in Alzheimer’s Disease (AD). These methods have the potential to alleviate the personal and societal burden of AD, by reducing time to diagnosis. Delayed AD diagnosis has adverse effects on care planning and family relationships, and has been estimated to cost the healthcare system close to $8 trillion dollars. Language reflects cognitive status, and contemporary neural network models have been shown to discriminate between transcribed speech from patients with AD and that from healthy controls with promising accuracy. However, most of this work has been conducted in the context of recorded responses to picture description tasks, which are not suitable for repeated, continuous or passive monitoring for linguistic indicators of AD-related decline. In contrast, our recent work has identified the task-agnostic linguistic construct of semantic coherence as a sound basis for machine learning models for detection of language from people with AD in casual conversations, with classification performance in this context exceeding that of Bidirectional Encoder Representations from Transformers (BERT) based classifiers, the state-of-the-art for text-based AD detection. Unfortunately, recent work has shown that automated coherence estimates are vulnerable to bias, with lower estimates for speech from people identifying as black irrespective of diagnosis. Interrogation of coherence- and BERT-based classifiers reveals they dramatically underperform in this group also. For the parent award for this supplement proposal (R01 LM0104056), we will develop methods to debias deep transformer networks to mitigate for the confounding variable of provenance in multi-institutional datasets, culminating in the release of a toolkit to Deconfound Deep Transformer Networks, the DeconDTN suite. In this proposal we will apply these methods to deconfound coherence- and BERT-based AD detection models for the confounding variable of race and/or ethnicity, and evaluate the effects on model performance across groups. In addition, we will evaluate the utility of fine-tuning semantic coherence models and models used to generate automated transcripts on text and speech from the recently-released Corpus of Regional African American Language (CORAAL). We hypothesize that both improvements in these underlying models, and explicit deconfounding for race/ethnicity will reduce the performance differential across groups, resulting in equitable models for language-based AD detection.