Speech-Based Automated Cognitive Impairment Detection From Remotely-Collected Cognitive Test Audio

• Published in: IEEE access Vol. 6; pp. 40494 - 40505
• Main Authors: Yu, Bea, Williamson, James R., Mundt, James C., Quatieri, Thomas F.
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2018; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Animals; Audio data; Automation; Cognitive ability; Correlation; cross-validation; Dementia; Feature extraction; feature selection; formant frequencies; Impairment; Learning; Mild cognitive impairment; Monitoring; motor coordination; phoneme durations; remote health care monitoring; Remote monitoring; Speech; Speech recognition; Statistical models; Task analysis; Testing; vocal biomarkers
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2018.2856478

Remote-automated cognitive impairment (CI) monitoring has the potential to facilitate care for the elderly with mobility restrictions. In particular, CI detection based on speech features from audio data collected for remote cognitive testing holds significant promise to improve remote cognitive health monitoring. This requires no additional testing for speech analysis and, combined with cognitive test scores, can improve CI detection over using cognitive test scores alone. This paper builds on previous work with an expanded set of speech features extracted from a larger suite of remotely administered cognitive tests. The speech features tested include measures of phoneme characteristics, pitch, and articulation. The relative merits of using speech features, a common cognitive test score, and both combined for CI prediction are also explored. The best performing system uses a combination of speech features and the cognitive test score, obtaining a performance outcome of area under the ROC curve (AUC) = 0.77. This outcome is better at the 5% significance level than that obtained using the speech features alone (AUC = 0.74) or the cognitive test alone (AUC = 0.54). Additionally, the influence of validation methodology on performance estimation is addressed in detail. Learning statistical models for speech-based CI diagnosis is challenging due to limited availability of audio data from subjects with clinical CI diagnoses. Rigorous validation methods for model learning are important in this context. The stringent validation methodology developed in this paper produces more conservative, and likely, more generalizable performance estimates compared with methodologies used in prior art.