Detecting late-life depression in Alzheimer’s disease through analysis of speech and language Kathleen C. Fraser1 and Frank Rudzicz2,1 and Graeme Hirst1 1Department of Computer Science, University of Toronto, Toronto, Canada 2Toronto Rehabilitation Institute-UHN, Toronto, Canada kfraser,frank,gh @cs.toronto.edu { } Abstract It is also important to detect when someone has both AD and depression, as this serious situation can Alzheimer’s disease (AD) and depression lead to more rapid cognitive decline, earlier place- share a number of symptoms, and commonly ment in a nursing home, increased risk of depression occur together. Being able to differentiate be- in the patient’s caregivers, and increased mortality tween these two conditions is critical, as de- (Thorpe, 2009; Lee and Lyketsos, 2003). pression is generally treatable. We use linguis- Separate bodies of work have reported the util- tic analysis and machine learning to determine whether automated screening algorithms for ity of spontaneous speech analysis in distinguish- AD are affected by depression, and to detect ing participants with depression from healthy con- when individuals diagnosed with AD are also trols, and in distinguishing participants with demen- showing signs of depression. In thefirst case, tia from healthy controls. Here we consider whether wefind that our automated AD screening pro- such analyses can be applied to the problem of de- cedure does not show false positives for indi- tecting depression in Alzheimer’s disease (AD). In viduals who have depression but are otherwise particular, we explore two questions: (1) In previous healthy. In the second case, we have moderate success in detecting signs of depression in AD work on detecting AD from speech (elicited through (accuracy = 0.658), but we are not able to draw a picture description task), are cognitively healthy a strong conclusion about the features that are people with depression being misclassified as hav- most informative to the classification. ing AD? (2) If we consider only participants with AD, can we distinguish between those with depres- sion and those without, using the same picture de- 1 Introduction scription task and analysis? Depression and dementia are both medical condi- 2 Background tions that can have a strong negative impact on the quality of life of the elderly, and they are often co- There has been considerable work on detecting de- morbid. However, depression is often treatable with pression from speech and on detecting dementia medication and therapy, whereas dementia usually from speech, but very little which combines the two. occurs as the result of an irreversible process of neu- We willfirst review the two tasks separately, and rodegeneration. It is therefore critical to be able to then discuss some of the complexity that arises when distinguish between these two conditions. depression and AD co-occur. However, distinguishing between depression and dementia can be extremely difficult because of over- 2.1 Detecting depression from speech lapping symptoms, including apathy, crying spells, Depression affects a number of cognitive and phys- changes in weight and sleeping patterns, and prob- ical systems related to the production of speech, in- lems with concentration and attention. cluding working memory, the phonological loop, ar- 1 Proceedings of the 3rd Workshop on Computational Linguistics and Clinical Psychology: From Linguistic Signal to Clinical Reality, pages 1–11, San Diego, California, June 16, 2016. c 2016 Association for Computational Linguistics ticulatory planning, and muscle tension and control 30 participants with depression and 30 healthy con- (Cummins et al., 2015). These changes can result trols. The speech was elicited through interview in word-finding difficulties, articulatory errors, de- questions about situations that had aroused signif- creased prosody, and lower verbal productivity. icant emotions. Higher accuracy was achieved on Over the past decade or so, there has been grow- detecting depression in women than in men. Energy, ing interest in measuring properties of the speech intensity, shimmer, and MFCC features were all in- signal that correlate with the changes observed in de- formative, and positive emotional speech was more pression, and using these measured variables to train discriminatory than negative emotional speech. machine learning classifiers to automatically detect Scherer et al. (2013) differentiated 18 depressed depression from speech. participants from 18 controls with 75% accuracy, Ozdas et al. (2004) found that mean jitter and the using interviews captured with a simulated virtual slope of the glottalflow spectrum could distinguish human. They found that glottal features such as between 10 non-depressed controls, 10 participants the normalized amplitude quotient (NAQ) and quasi- with clinical depression, and 10 high-risk suicidal open quotient (QOQ) differed significantly between participants. the groups. Moore et al. (2008) considered prosodic features Alghowinem et al. (2013) compared four clas- as well as vocal tract and glottal features. They per- sifiers and a number of different feature sets on formed sex-dependent classification and found that the task of detecting depression from spontaneous glottal features were more discriminative than vocal speech. They found loudness and intensity features tract features, but that the best results were achieved to be the most discriminatory, and suggested pitch using all three types of features. and formant features may be more useful for longi- Cohn et al. (2009) examined the utility of facial tudinal comparisons within individuals. movements and vocal prosody in discriminating par- While most of the literature concerning the detec- ticipants with moderate or severe depression from tion of depression from speech has focused solely those with no depression. They achieved 79% accu- on the speech signal, there is an associated body of racy using only two prosodic features: variation in work on detecting depression from writing that fo- fundamental frequency, and latency of response to cuses on linguistic cues. Rude et al. (2004) found interviewer questions. They used a within-subjects that college students with depression were signifi- design, in which they predicted which participants cantly more likely to use thefirst-person pronoun had responded to treatment in a clinical trial. I in personal essays than college students without Low et al. (2011) analyzed speech from ado- depression, and also used more words with neg- lescents engaged in normal conversation with their ative emotional valence. Other work has found parents (68 diagnosed with depression, 71 con- differences in the frequency of different parts-of- trols). They grouped their acoustic features into speech (POS) (De Choudhury et al., 2013) and in 5 groups: spectral, cepstral, prosodic, glottal, and the general topics chosen for discussion (Resnik et those based on the Teager energy operator (TEO, a al., 2015). Other work has accurately identified de- nonlinear energy operator). They achieved higher pression (and differentiated PTSD and depression) accuracies using sex-dependent models than sex- in Twitter social media texts with high accuracies independent models, and found that the best results usingn-gram language models (Coppersmith et al., were achieved using the TEO-based features (up to 2015). Similarly, Nguyen et al. (2014) showed 87% for males and 79% for females). that specialized lexical norms and Linguistic Inquiry Cummins et al. (2011) distinguished 23 depressed and Word Count1 features significantly differentiate participants from 24 controls with a best accuracy clinical and control groups in blog post texts. Howes of 80% in a speaker-dependent configuration and et al. (2014) showed that lexical features (in style 79% in a speaker-independent configuration. Spec- and dialogue) could also be used to predict the sever- tral features, particularly mel-frequency cepstral co- ity of depression and anxiety during Cognitive Be- efficients (MFCCs), were found to be useful. Alghowinem et al. (2012) analyzed speech from 1http://liwc.wpengine.com. 2 havioural Therapy treatment. It is not obvious that pothesis that AD patients would use more pronouns, these results generalize to the case where the topic verbs, and adjectives and fewer nouns than controls. and structure of the narrative is constrained to a pic- Rentoumi et al. (2014) considered a slightly dif- ture description. ferent problem: they used computational techniques to differentiate between picture descriptions from 2.2 Detecting Alzheimer’s disease from speech AD participants with and without additional vas- A growing number of researchers have tackled the cular pathology (n= 18 for each group). They problem of detecting dementia from speech and achieved an accuracy of 75% when they included language. Most of this work has focused on frequency unigrams and excluded binary unigrams, Alzheimer’s disease (AD), which is the most com- syntactic complexity features, measures of vocabu- mon cause of dementia. Although the primary diag- lary richness, and information theoretic features. nostic symptom of AD is memory impairment, this Orimaye et al. (2014) obtainedF-measure scores and other cognitive deficits often manifest in spon- up to 0.74 on transcripts from DementiaBank, com- taneous language through word-finding difficulties, bining participants with different etiologies rather a decrease in information content, and changes in than focusing on AD. In previous work, we also fluency,