Abstract
This study presents the development and validation of a novel hybrid Artificial Intelligence
(AI) framework for the automated detection of COVID-19 from a limited dataset of 510
chest X-ray images sourced from a Zimbabwean radiology center, specifically addressing
the challenges of data scarcity and resource constraints in low-resource healthcare settings.
By combining Convolutional Neural Network (CNN) architectures (Visual Geometry Group
16 and EfficientNetB0) for deep feature extraction with a Random Forest (RF) classifier,
our proposed hybrid models demonstrated a significant leap in diagnostic performance over
traditional deep learning approaches. After addressing a severe class imbalance (90 positive
vs. 420 negative cases) using the Synthetic Minority Oversampling Technique, our hybrid
models achieved outstanding accuracy, with the EfficientNetB0 and Boruta derived model
reaching 96.0% accuracy, 93.8% precision, and 96.2% recall, a substantial improvement
from the 88.2% accuracy of a baseline CNN model. The integration of dimensionality reduction
techniques like the Boruta algorithm and Principal Component Analysis proved critical
in refining the feature space for the RF classifier, with Area Under Curve scores improving
from 0.68 for the baseline CNN model to a minimum of 0.98 for the hybrid approaches.
Further enhancing the framework, topological data analysis using the Mapper algorithm provided
useful qualitative insights into the data’s structure, revealing distinct clusters, loops,
and flares that correspond to underlying patterns of COVID-19 pathology, which are often
missed by conventional statistical methods. These topological maps demonstrated superior
separability between COVID-19 positive and negative cases, with the algorithm successfully
identifying subpopulations and continuous profile variations that may represent different disease
severity or presentations. The topological features extracted through Mapper graphs
captured the essential geometric and structural properties of the data set, achieving higher
classification accuracy than the hybrid models developed and evaluated in this study. The
findings demonstrate that hybrid AI models, enhanced by topological analysis, can achieve
diagnostic accuracy comparable to typically expensive laboratory tests while providing interpretable
insights crucial for clinical decision-making, offering a promising solution for
improving COVID-19 detection capabilities in settings where traditional diagnostic infrastructure
is limited or unavailable.