Back to search

University of North Texas

Deep learning-based Assessment of Gait and Fall Risk in Motor-Impaired Individuals Using Electronic Medical Records

Abstract

dc:description

Motor impairment conditions such as lower limb amputation and cerebral palsy (CP) significantly affect mobility, functional independence, and quality of life. Clinical assessment of these populations is often multidimensional, heterogeneous, and reliant on subjective interpretation. These make the evaluation challenging. At the same time, the rapid adoption of electronic medical records (EMRs) has generated large volumes of clinical data that remain underutilized for decision support. To utilize and make assessments easier, we present progressive, data-driven methodologies to improve assessments in motor-impaired populations. 1) We propose an unsupervised autoencoder-based approach to evaluate prosthetic performance in K2-level lower limb amputees. By compressing multiple correlated clinical and self-reported outcome measures into a single latent representation, the model generates a composite metric that captures overall functional improvement while reducing measurement noise. This summary metric enables clearer interpretation of health changes and supports holistic evaluation in small clinical cohorts. 2) To extend the assessment process, we utilize unstructured clinical narratives by adopting BERT-based transformer models to predict fall risk in children with CP using unstructured EMRs from 1,604 subjects. This study demonstrates that language models can extract clinically meaningful signals from free-text documentation, enabling objective and scalable fall risk stratification from real-world hospital data. 3) Finally, to address limited labeled data in clinical settings, we develop an instruction-tuned large language model (LLM) for fall risk assessment. Using a small set of 135 instructional examples, we fine-tune the GPT-2 model and iteratively improve performance through synthetic data generation. This data-efficient strategy enhances robustness, reduces annotation burden, and demonstrates practical deployment potential in low-resource environments. Collectively, these studies illustrate an evolution from unsupervised representation learning on structured outcomes to transformer-based and instruction-tuned language models applied to unstructured EMRs. The proposed methodologies provide scalable and clinically practical tools by establishing a pathway toward more objective and personalized decision support for populations with motor impairments.

Degree

thesis:*
Grantor dc:publisher
University of North Texas
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Tabashum, Thasina
Contributors dc:contributor
  • Albert, Mark V.
  • Xiao, Ting
  • Bozdag, Serdar
  • Jayaraman, Arun

Subjects

dc:subject × 3

Rights

dc:rights
Statement dc:rights
  • Public
  • Tabashum, Thasina
  • Copyright
  • Copyright is held by the author, unless otherwise noted. All rights Reserved.
Language dc:language
English

Identifiers

dc:identifier.*
Identifier
local-cont-no: submission_4848
ark: ark:/67531/metadc2652423

Chain of custody

source
Harvested from
University of North Texas
Base URL
digital.library.unt.edu/oai/
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Tabashum, Thasina. Deep learning-based Assessment of Gait and Fall Risk in Motor-Impaired Individuals Using Electronic Medical Records. University of North Texas, 2026. https://digital.library.unt.edu/ark:/67531/metadc2652423/