Back

Social media language patterns reveal early signs of impending stroke: An observational study

Dahari, Y.; Ziv, T. A.; Yom-Tov, E.

2025-11-07 cardiovascular medicine
10.1101/2025.11.05.25339635 medRxiv
Show abstract

BackgroundStroke remains a major global health concern, contributing substantially to mortality and long-term disability. Current clinical tools lack effective mechanisms for early detection. Here, we investigate whether linguistic and behavioral patterns in online social media can serve as early indicators of impending stroke. MethodsWe analyzed posts from 1,683 Reddit users who reported experiencing a stroke and 2,438 users across three control groups al made between 2015 and 2024. Linguistic features were extracted from posts, and predictive models were trained to distinguish individuals who experienced a stroke from controls. ResultsOur results reveal changes in several linguistic markers, e.g., the rate of spelling errors and the use of word tokens, beginning approximately 20 weeks prior to the stroke. Using posts from the four months preceding the event, the predictive model achieved an area under the curve (AUC) of 0.87. DiscussionThese findings highlight the potential of social media-derived linguistic signals to predict strokes several months in advance, offering a promising avenue for early detection and preventive interventions in digital medicine.

Matching journals

The top 8 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.