Science1 publisher2 min readPublished
How people retell a story in a memory test predicted lower cognitive scores seven years later
Boston University researchers used language software on story-recall recordings to flag mild cognitive impairment and forecast lower scores seven years later. The test is one clinics already give, so the open question for a cheap screen is how much the speech adds to the ordinary score.
The Scientist · Science desk
What happened
- Conventional story-recall scoring gives a point for each detail a patient recalls, verbatim or synonymous, and adds the points into one composite score.
- People with mild cognitive impairment lost core story details, made more off-task remarks and admissions of forgetting, and used simpler, more repetitive sentences.
- The study is published in the Journal of the International Neuropsychological Society.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- capability The input is a recording of a test clinics already run, so a validated profile would give each test a second reading, delivered by phone app or in a primary care visit.
- constraint The reported outcome is lower test performance, so a flag from this profile cannot yet be treated as a forecast of dementia itself.
- decision Thousands of candidate variables drawn from one cohort mean anyone considering deployment has good reason to wait for a frozen model tested on a second population.
The appeal of the design is that it asks nothing new of the patient. The recordings came from a standard paragraph-recall assessment [3], the kind already scored by checklist in clinics [7]. A remark such as "I know there was another person, but I forgot her name" earns nothing on that checklist, according to the release [11]. The study treats it as a measurement.
"Our method allows us to go beyond right or wrong scoring and capture qualities of the spoken test responses which shows us how a person thinks and remembers information and whether they notice their own mistakes," said lead author Seho Park, a postdoctoral associate at the school [8][13].
The release does not give the number of participants, an accuracy figure for either the impairment classification or the seven-year forecast, or any test of the profile in a second cohort.
The open question is how much the speech profile adds to the score clinicians already compute. The marker groups [5] split into two kinds. Loss of salient detail sits close to what the checklist counts [7], so a model leaning on it may partly be rescoring the same test. Off-target commentary, admissions of forgetting and flattened syntax are invisible to the checklist. Any genuinely new signal has to come from those.
The outcome needs the same care. According to the release, the combined profile forecast lower cognitive performance seven years later [6]. A lower test score is not a dementia diagnosis, and a screen sold for dementia risk would eventually need to predict diagnoses.
Then there is the search space. The team parsed thousands of acoustic, semantic and syntactic variables [4] from a single study population, the Long Life Family Study [3]. With that many candidates, some features will separate groups by chance. I'd want the profile fixed in advance and applied unchanged to recordings from a different cohort before putting weight on the seven-year result.
The screening case follows directly from the input: the analysis is automated and runs on spoken answers to an existing test, and the release suggests it could be deployed through smartphone apps or routine primary care visits before conventional scores drop [10]. Standard cognitive evaluations are time-consuming, the release notes [12]. Park also points to a diagnostic use beyond early flagging. "These behavioral differences have great potential to help us detect cognitive impairment earlier and pinpoint different types of cognitive impairment," Park said [9].
What to watch
- Whether the full paper reports how much the speech profile improves the seven-year prediction over the standard recall score alone.
- A replication applying the same fixed profile to recordings from a cohort other than the Long Life Family Study.
- Whether the profile predicts clinical diagnoses of mild cognitive impairment or dementia, and not only lower test scores.