DeepAphasia: Transformer-based aphasia screening for acute stroke patients using contrastive segment-level labels

Peiqi Sui; Zhihao Wan; Kelvin Wong; Xiaohui Yu; Rachel Leicht; John J. Volpi; Stephen T. Wong

DeepAphasia: Transformer-based aphasia screening for acute stroke patients using contrastive segment-level labels

Peiqi Sui, Zhihao Wan, Kelvin Wong, Xiaohui Yu, Rachel Leicht, John J. Volpi, Stephen T. Wong

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Abstract

We develop DeepAphasia, the first audio-based pipeline for rapid and reliable aphasia screening
in stroke patients. Early detection of aphasia is crucial for improving stroke outcomes, while current AI-based
methods are time-consuming and require additional assessments. Our approach better fit the limitations of
the acute setting as it could detect aphasia with only the brief (60s) voice recordings of patients describing the
cookie theft picture, a standard component of the NIHSS. Motivated by the clinical insight that not all parts of
aphasia patient speech could be uniformly considered as aphasiac, we introduced contrastive labeling on
segments (10s) of patient voice recordings to better capture this difference. Our results showed that training
models on the segment-level significantly increases performance of aphasia detection.

Original language	English (US)
Title of host publication	11th European Stroke Organization Conference proceeding
State	Published - 2025
Event	European Stroke Organization Conference 2025 - Helsinki, Finland Duration: May 21 2025 → May 23 2025 Conference number: 11th https://eso-stroke.org/esoc2025/

Conference

Conference	European Stroke Organization Conference 2025
Abbreviated title	ESOC 2025
Country/Territory	Finland
City	Helsinki
Period	5/21/25 → 5/23/25
Internet address	https://eso-stroke.org/esoc2025/

Divisions

Medical Oncology

Cite this

@inproceedings{6a8a9f31bc3c44f68d5ab438c388ca9d,

title = "DeepAphasia: Transformer-based aphasia screening for acute stroke patients using contrastive segment-level labels",

abstract = "We develop DeepAphasia, the first audio-based pipeline for rapid and reliable aphasia screening in stroke patients. Early detection of aphasia is crucial for improving stroke outcomes, while current AI-based methods are time-consuming and require additional assessments. Our approach better fit the limitations of the acute setting as it could detect aphasia with only the brief (60s) voice recordings of patients describing the cookie theft picture, a standard component of the NIHSS. Motivated by the clinical insight that not all parts of aphasia patient speech could be uniformly considered as aphasiac, we introduced contrastive labeling on segments (10s) of patient voice recordings to better capture this difference. Our results showed that training models on the segment-level significantly increases performance of aphasia detection.",

author = "Peiqi Sui and Zhihao Wan and Kelvin Wong and Xiaohui Yu and Rachel Leicht and Volpi, \{John J.\} and Wong, \{Stephen T.\}",

year = "2025",

language = "English (US)",

booktitle = "11th European Stroke Organization Conference proceeding",

note = "European Stroke Organization Conference 2025, ESOC 2025 ; Conference date: 21-05-2025 Through 23-05-2025",

url = "https://eso-stroke.org/esoc2025/",

}

TY - GEN

T1 - DeepAphasia: Transformer-based aphasia screening for acute stroke patients using contrastive segment-level labels

AU - Sui, Peiqi

AU - Wan, Zhihao

AU - Wong, Kelvin

AU - Yu, Xiaohui

AU - Leicht, Rachel

AU - Volpi, John J.

AU - Wong, Stephen T.

N1 - Conference code: 11th

PY - 2025

Y1 - 2025

N2 - We develop DeepAphasia, the first audio-based pipeline for rapid and reliable aphasia screening in stroke patients. Early detection of aphasia is crucial for improving stroke outcomes, while current AI-based methods are time-consuming and require additional assessments. Our approach better fit the limitations of the acute setting as it could detect aphasia with only the brief (60s) voice recordings of patients describing the cookie theft picture, a standard component of the NIHSS. Motivated by the clinical insight that not all parts of aphasia patient speech could be uniformly considered as aphasiac, we introduced contrastive labeling on segments (10s) of patient voice recordings to better capture this difference. Our results showed that training models on the segment-level significantly increases performance of aphasia detection.

AB - We develop DeepAphasia, the first audio-based pipeline for rapid and reliable aphasia screening in stroke patients. Early detection of aphasia is crucial for improving stroke outcomes, while current AI-based methods are time-consuming and require additional assessments. Our approach better fit the limitations of the acute setting as it could detect aphasia with only the brief (60s) voice recordings of patients describing the cookie theft picture, a standard component of the NIHSS. Motivated by the clinical insight that not all parts of aphasia patient speech could be uniformly considered as aphasiac, we introduced contrastive labeling on segments (10s) of patient voice recordings to better capture this difference. Our results showed that training models on the segment-level significantly increases performance of aphasia detection.

M3 - Conference contribution

BT - 11th European Stroke Organization Conference proceeding

T2 - European Stroke Organization Conference 2025

Y2 - 21 May 2025 through 23 May 2025

ER -

DeepAphasia: Transformer-based aphasia screening for acute stroke patients using contrastive segment-level labels

Abstract

Conference

Divisions

Fingerprint

Cite this