Toward Dementia Diagnosis via Artificial Intelligence
Alzheimer's disease (AD) is the most common type of dementia and affects 44 million people worldwide; this is projected to triple by 2050.1 Assessing for AD remains an expensive and laborious process that is completely unsustainable given the rapidly aging populations of many nations. Even with the savings associated with early assessment, the amortized cost for each diagnosis is currently at least $6,000,2 which does not take into account indirect costs such as lost time by patients (and their caregivers) in travel, wait times, and hours spent in assessment, which is a process so laborious and stressful that it is often repeated only every year, or even less frequently. This is doubly unfortunate because, given the high variability of symptoms in AD, it often cannot be ascertained accurately from a single assessment,3 especially in its early stages. Repeatable, remote, and cost-effective assessment is therefore essential.
Modern artificial intelligence (AI) research is producing tools that will help to solve this challenge. In particular, machine learning, a subarea of AI, is developing algorithms that can accurately learn models of real-world phenomena, or categorical differences among those phenomena, from representative data. Progress in this area has recently accelerated dramatically, due to the discovery of new algorithms in artificial neural networks, the availability of powerful hardware capable of vast parallel processing, and the wide proliferation of data across a variety of domains, including health care. In particular, recent advances are showing that AI can be used to assess and monitor AD and other types of dementia, using no more than simple measures of a speaker's voice.
Language in AD
AD affects not only the words and sounds of speech but also higher-level aspects. For example, discourse-level aspects such as understanding metaphor, sarcasm,11 and global coherence12 are negatively impacted in AD. Grammar is also diminished, including reduced syntactic complexity13 and fewer subordinate clauses.14
Crucially, whether an individual becomes more impaired in either his or her grammar or word finding can be indicative of the location of the etiology. For example, damage to Broca's area (Brodmann area 44/45) in the left frontal lobe can result in agrammatical speech, whereas damage to Wernicke's area (Brodmann area 22), in the posterior superior temporal gyrus can reduce the ability to understand or produce meaningful language. Our own work in the linguistic analysis of AD shows that syntax, semantics, and even acoustic properties of the voice can be differentially impacted, and are indicative of this disease.15
Computational Approaches to Assessing Dementia
While still in its infancy, assessing clinically relevant aspects of cognition through automated analysis of language is growing in popularity. Bucks et al could identify individuals with AD, from a cohort of eight, with 87.5% accuracy by applying eight linguistic features within a linear discriminant analysis.16 Along these lines, Guinn and Habash built software that identified AD, with up to 79.5% accuracy within 80 conversations and found that measuring the syntactic categories of words (eg, proper nouns, gerund verbs) was less useful than more acoustically identifiable measures such as filled pauses, repetition, and incomplete words.17 Jarrold et al also used acoustic features but combined these with syntactic word categories and psychologically motivated word lists, obtaining 88% accuracy.18
Members of the computational linguistics group at the University of Toronto have recently shown that the early warning signs of dementia were detectable algorithmically through diachronic changes in literary writing19 and were sensitive even to individual writing style.20 At the Toronto Rehabilitation Institute and the Rotman Research Institute, the author of this article and colleagues have expanded on that work, combining a vast collection of linguistic measurements of speech with modern machine learning to develop computational methods that accurately identify primary progressive aphasia and its subtypes,21 Parkinson's disease,22 indications of confusion in spoken dialogue in AD,23 and AD itself.15 The same core algorithms are widely applicable across pathologies and are exceedingly accurate, as long as appropriate linguistic measures are selected by machine learning, given sufficient amounts of data. We are currently in the process of commercializing these technologies in a company called Winterlight Labs.
The Future of AI in Clinical Practice
— Frank Rudzicz, PhD, is a scientist at the Toronto Rehabilitation Institute in the University Health Network, an assistant professor of computer science at the University of Toronto, and cofounder of Winterlight Labs Inc. He is the president of the international joint ACL/ISCA Special Interest Group on Speech and Language Processing for Assistive Technologies, and Young Investigator of the Alzheimer's Society. His work involves machine learning, human-computer interaction, speech-language pathology, rehabilitation engineering, signal processing, and linguistics.
2. Getsios D, Blume S, Ishak KJ, Maclaine G, Hernández L. An economic evaluation of early assessment for Alzheimer's disease in the United Kingdom. Alzheimers Dement. 2012;8(1):22-30.
3. Rockwood K, Fay S, Hamilton L, Ross E, Moorhouse P. Good days and bad days in dementia: a qualitative analysis of variability in symptom expression. Int Psychogeriatr. 2014;26(8):1239-1246.
4. Forbes KE, Venneri A, Shanks MF. Distinct patterns of spontaneous speech deterioration: an early predictor of Alzheimer's disease. Brain Cogn. 2002;48(2-3):356-361.
5. Faber-Langendoen K, Morris JC, Knesevich JW, LaBarge E, Miller JP, Berg L. Aphasia in senile dementia of the Alzheimer type. Ann Neurol. 1988;23(4):365-370.
6. Ahmed S, Haigh M, de Jager CA, Garrard P. Connected speech as a marker of disease progression in autopsy-proven Alzheimers disease. Brain. 2013;136(Pt 12):3727-3737.
7. Weiner MF, Neubecker KE, Bret ME, Hynan LS. Language in Alzheimer's disease. J Clin Psychiatry. 2008;69(8):1223-1227.
8. Kirshner HS, Webb WG, Kelly MP. The naming disorder of dementia. Neuropsychologia. 1984;22(1):23-30.
9. Oppenheim G. The earliest signs of Alzheimer's disease. J Geriatr Psychiatry Neurol. 1994;7(2):116-120.
10. Robinson KM, Grossman M, White-Devine T, D'Esposito M. Category-specific difficulty naming with verbs in Alzheimer's disease. Neurology. 1996;47(1):178-182.
11. Rapp AM, Wild B. Nonliteral language in Alzheimer dementia: a review. J Int Neuropsychol Soc. 2011;17(2):207-218.
12. Glosser G, Deser T. Patterns of discourse production among neurological patients with fluent language disorders. Brain Lang. 1991;40(1):67-88.
13. Sajjadi SA, Patterson K, Tomek M, Nestor PJ. Abnormalities of connected speech in semantic dementia vs Alzheimer's disease. Aphasiology. 2012;26(6):847-866.
14. Croisile B, Ska B, Brabant MJ, et al. Comparative study of oral and written picture description in patients with Alzheimer's disease. Brain Lang. 1996;53(1):1-19.
15. Fraser KC, Meltzer JA, Rudzicz F. Linguistic features identify Alzheimer's disease in narrative speech. J Alzheimers Dis. 2015;49(2):407-522.
16. Bucks RS, Singh S, Cuerden JM, Wilcock GK. Analysis of spontaneous, conversational speech in dementia of Alzheimer type: evaluation of an objective technique for analysing lexical performance. Aphasiology. 2000;14:71-91.
17. Guinn CI, Habash A. Language analysis of speakers with dementia of the Alzheimer's type. AAAI Fall Symposium Series. 2012;8-13.
18. Jarrold W, Peintner B, Wilkins D, at al. Aided diagnosis of dementia type through computer-based analysis of spontaneous speech. ACL Workshop on Computational Linguistics and Clinical Psychology. 2014;27-36.
19. Le X, Lancashire I, Hirst G, Jokel R. Longitudinal detection of dementia through lexical and syntactic changes in writing: a case study of three British novelists. Lit Linguist Comp. 2011;26(4):435-461.
20. Hirst G, Feng VW. Changes in style in authors with Alzheimer's disease. English Stud. 2012;93(3):357-370.
21. Fraser KC, Rudzicz F, Rochon E. Using text and acoustic features to diagnose progressive aphasia and its subtypes. Proc Interspeech. 2013;2177-2181.
22. Zhao S, Rudzicz F, Carvalho LG, Márquez-Chin C, Livingstone S. Automatic detection of expressed emotion in Parkinson's disease. IEEE International Conference on Acoustics, Speech and Signal Processing. 2014:4846-4850.
23. Rudzicz F, Wang R, Begum M, Mihailidis A. Speech interaction with personal assistive robots supporting aging at home for individuals with Alzheimer's disease. ACM Trans Accessible Comput. 2015;7(2).