Abstract
Brain stroke remains one of the primary causes of death and long-term
disability worldwide, making early detection and ongoing monitoring essential
for improving patient outcomes. Although current diagnostic systems
use machine learning for stroke detection, they frequently lack robustness
across diverse datasets and do not provide mechanisms for tracking stroke
evolution over time. This limitation reduces the impact of real-time clinical
decision support and disrupts continuity of care.
This study aims to develop and evaluate a robust, explainable diagnostic
framework that integrates Variational Autoencoders (VAEs) with image Re-
Identification (Re-ID) techniques for early detection and dynamic monitoring
of brain strokes using Computed Tomography (CT) imaging. The framework
addresses challenges of adaptability, catastrophic forgetting, and interpretability
by combining generative modelling, supervised classification,
and metric learning within a continual learning setting.
The methodology follows a Design Science Research approach, implementing
the DB-VAE-Re-ID architecture with a ResNet18 encoder, dynamic
transposed convolution decoder, and classification head. Synthetic patient
profiles were generated to simulate longitudinal scans, and experiments were
conducted on a publicly available stroke Computed Tomography dataset.
Performance was assessed using reconstruction metrics, classification metrics
(accuracy, precision, recall, F1-score), and re-identification metrics (mAP,
Top 1 accuracy), alongside latent space figures for interpretability.
The results demonstrate an overall stroke detection accuracy of 85%, with
robust performance for ischemic stroke detection (Recall: 0.96, F1: 0.90) and
robust latent space clustering enabling patient level re-identification. However,
bleeding cases show a lower Recall (0.60), indicating a need for improved
feature discrimination. The framework successfully maintained identity
consistency across synthetic patient profiles and supported risk scoring
for longitudinal monitoring.
The implications of this work are substantial. By incorporating image reidentification
into stroke assessment, the proposed system shifts traditional
static classification toward a longitudinal, patient focused decision support
framework, improving diagnostic robustness and enabling earlier, targeted
interventions. This strategy establishes a basis for scalable and transparent
Artificial Intelligence solutions in healthcare, with promising extensions to
multi-modal imaging and real-time integration into clinical workflows.