Instrumentation & Measurement Magazine 24-6 - 15

cognitive testing instruments mentioned above, the MMSE
has specific known issues, including non-linearity as well as
floor and ceiling effects-a highly-educated individual, for example,
may report a decline in their own cognition while still
scoring within the 'normal' range on the MMSE. Therefore,
engineering a speech-based instrument which faithfully reproduces
the MMSE would represent progress, but could not
claim to measure the 'true' value of cognition, which we must
take to be " not only unknown but also unknowable " [11].
Repeatability
Ideally, if we use the same instrument to measure the same
quantity under the same conditions, we will receive the
same measurement. However, the vast majority of research
on speech-based dementia detection considers only a single
speech sample per person, making it impossible to assess the
stability of the measurement. When uncertainty is discussed,
it is typically only in the context of the machine learning algorithm:
if some parameter is changed (e.g., the training data in
the case of cross-validation), then what effect does it have on
the accuracy of the prediction with respect to the reference?
This source of variance in the outcome is sometimes reported
in terms of confidence intervals or statistical significance. However,
to our knowledge, the following basic experiment has not
been conducted: If we took a speech sample from a person today
and another sample tomorrow, extracted the same features
from each and fed them to the same machine learning model,
how closely would the outputs match? This concept of repeatability
(or in the medical community, test-retest reliability) is not
assessed in single-sample, cross-sectional study designs.
Resolution
As discussed, many studies characterize cognitive status as a
binary classification: participants are labelled as either having
dementia, or not. To monitor change over time, it will be necessary
to recognize cognition as a continuous measure. Some
work has started to move in this direction by reformulating the
classification task as a regression task, with the aim of predicting
an MMSE score rather than a diagnostic category. While
this would certainly represent an improvement in resolution,
even MMSE is a relatively coarse scale from the metrology perspective,
with only 30 possible values to cover the entire range
of human cognitive ability.
Sensitivity
Most studies on speech-based measures of cognition involve
participants who have already been diagnosed with dementia,
due to the difficulties in identifying and recruiting study
participants who will develop cognitive impairment over the
course of a time-limited study period. However, this limits our
understanding of how early in the disease trajectory we can
actually pick up on changes. Computational studies of speech
from public figures, such as politicians and professional athletes
[12], suggest that the changes may be detectable from
speech prior to formal diagnosis, but this work has yet to be extended
to a large sample of ordinary citizens.
September 2021
Continuous speech sensing in a smart home environment
can mitigate some of these issues. While reference to an established
clinical instrument such as the MMSE is still necessary,
continuous longitudinal sensing makes it possible to use an individual's
own normal functioning as a reference and detect
decline with respect to that baseline-that is, to calibrate the
measurement relative to each individual. Furthermore, continuous
(or high-frequency) speech recording should make
it possible to measure the variance in the signal, develop
confidence intervals around the measurement, and detect deviations
from that normal range. Finally, if sensors are installed
in the homes of cognitively healthy seniors and allowed to operate
unobtrusively over time, then in cases where the user
does develop dementia, we will have a better understanding
of the sensitivity of speech as a measurement of cognitive
decline.
Smart Home Speech Sensing
There have been a number of proposals for how smart home
speech sensing might be implemented (Fig. 2). These proposals
fall broadly into two categories: active sensing, in which
users must perform some action to initiate the recording, and
passive sensing, in which the recording occurs unobtrusively,
without the user's intervention. In either case, the encrypted
audio data is then sent to an external server for processing and
analysis. If a concerning trend is detected, the user's healthcare
provider is notified for further assessment and possible
intervention.
Active Sensing
There has been growing interest in providing mobile cognitive
assessments in the form of apps that can be deployed on a
tablet or smart phone. Many of these assessments are simply
digitized versions of traditional pen-and-paper tests and are
designed to be administered by a health professional, rather
than independently in the home. Here, we focus instead on a
few examples of remote, repeated, home-based assessments
where speech was a primary outcome measure.
Jeffrey Kaye at the Oregon Health and Science University
in Portland, Oregon, USA proposed that cognitive tests
displayed on a home computer screen, recorded over the microphone,
and scored using ASR would greatly improve
outcome assessment in clinical trials for AD medications, by
allowing assessments to occur frequently, over a longer period
of time, and in a manner convenient to participants [13].
A test system was deployed, along with wireless passive infrared
motion sensors, in the homes of 265 older participants for
an average of 33 months [14]. As part of the protocol, participants
took part in daily conversations with research assistants
over a computer webcam. Subsequent analyses of the conversations
were able to detect MCI based on vocabulary richness
measures with AUC = 0.71 [15] and with an accuracy of 84%
based on the psycholinguistic properties of the words used
[16]. In a similar vein, the CAL: Cognitive Assisted Living smart
home architecture [17] and the Lilly Exploratory Digital Assessment
Study of smart device sensors [18] both incorporated
IEEE Instrumentation & Measurement Magazine
15
Instrumentation & Measurement Magazine 24-6

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-6
