You are here

Article Text

Article menu

Download PDFPDF

SNIS 20th annual meeting late-breaking oral abstracts
LB-009 Smartphone-enabled machine learning algorithms for autonomous stroke detection
Free
  1. R Raychev1,
  2. J Saver1,
  3. D Liebeskind1,
  4. M Nour1,
  5. S Petkov2,
  6. D Angelov2,
  7. D Georgiev3,
  8. A Koralova3,
  9. F Aleksiev4,
  10. T Sakelarova4,
  11. D Kalpachka4,
  12. R Kalpachki4,
  13. E Kostadinova5,
  14. T Manolova6,
  15. I Milanov7
  1. 1Neurology, UCLA, Los Angeles, CA
  2. 2Neuronics Medical, LOS ANGELES, CA
  3. 3Neurology, Haskovo General Hospital, Haskovo, BULGARIA
  4. 4Neurology, St. Ana University Hospital, Sofia, BULGARIA
  5. 5Neurology, Pulmed University Hospital, Plovdiv, BULGARIA
  6. 6Neurology, University Hospital Prof. Dr. Stoyan Kirkovich, Stara Zagora, BULGARIA
  7. 7Neurology, St. Naum University Hospital, Sofia, BULGARIA

Abstract

Background Using the well-established FAST paradigm, we developed an automated smart phone application for detection of acute stroke signs using machine learning (ML) algorithms for recognition of facial asymmetry, arm weakness, and speech changes (figure 1).

Methods We analyzed collected data from patients admitted to 4 major metropolitan stroke centers. Speech and facial data were captured via smartphone video recording and arm data was captured via device sensors.

A. Face. This module extracts standard 68 facial landmark points that are passed through a machine learning pipeline consisting of a dimensionality reduction step and an asymmetry classifier (figure 2).

B. Arm. Using data extracted from the 3D accelerometer, gyroscope, and magnetometer, while the smartphone is being held and moved, we designed a grasp agnostic classifier to process motion trajectories and detect arm weakness. (figure 3) C. Speech. We developed a model based on frequency analysis and Mel Frequency Cepstral Coefficients (MFCC) to detect abnormal/slurred speech. (figure 4) All tests were conducted within 72 hours of symptoms onset. Each of the three ML outputs was correlated with neurologists’ clinical impression and brain imaging results. Characteristics of the studied population are included in table 1. Results Results of final analyses of each individual modality for detection of abnormal speech, arm weakness, and facial asymmetry are summarized in table 2. Confirmed diagnosis of stroke is based on merged data from all 3 modalities.

View this table:
Abstract LB-009 Table 1

Characteristics of the studied population

View this table:
Abstract LB-009 Table 2

Results from correlation of individual and *merged ML outputs with neurologists’ clinical impression

Abstract LB-009 Figure 1–4
Abstract LB-009 Figure 1–4

Conclusions Our results confirm that smartphone enabled ML-algorithms can reliably identify acute stroke features with high accuracy comparable to neurologists’ clinical impression.

Disclosures R. Raychev: 1; C; Boehringer Ingelheim. 2; C; Rapid Medical, Perflow Medical. 4; C; Neuronics Medical. J. Saver: 2; C; Neuronics Medical. D. Liebeskind: 2; C; Neuronics Medical. M. Nour: 2; C; Neuronics Medical, Ischemia View. S. Petkov: 4; C; Neuronics Medical. D. Angelov: 4; C; Neuronics Medical. D. Georgiev: None. A. Koralova: None. F. Aleksiev: None. T. Sakelarova: None. D. Kalpachka: None. R. Kalpachki: None. E. Kostadinova: None. T. Manolova: None. I. Milanov: None.

https://doi.org/10.1136/jnis-2023-SNIS.383

Statistics from Altmetric.com

    Request Permissions

    If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.