Page 51 - AIH-2-3

Artificial Intelligence in Health AI in medical diagnostics: A multi-disease approach

errors and delays are leading causes of adverse patient By addressing the technical, ethical, and regulatory
outcomes, underscoring the urgent need for diagnostic challenges associated with AI in healthcare, this research
solutions that are more accurate, timely, and scalable. aims to deliver a comprehensive perspective on how AI
AI offers notable advancements in this area, enhancing can enhance diagnostic processes and improve patient
diagnostic precision, reducing human error, and enabling outcomes across multiple disease domains.
the early detection of diseases by analyzing vast amounts
of complex data. 1-6 2. Methods and experimental analysis

This research aims to investigate AI’s role in medical To evaluate the integration of AI in multi-disease
diagnostics across multiple diseases, focusing on cancer, diagnostics, this study compiled a diverse dataset drawn
cardiovascular diseases, neurological disorders, and from publicly available medical imaging repositories and
infectious diseases. These conditions were selected due to electronic health records (EHRs) across several healthcare
their significant global health burden, diverse diagnostic institutions. The dataset was constructed to represent four
requirements, and considerable impact on medical health primary disease areas: cancer, cardiovascular diseases,
systems. Each of these diseases brings unique diagnostic neurological disorders, and infectious diseases. Specifically,
challenges, such as managing various data complexities, cancer data were sourced from the Cancer Imaging Archive,
requiring rapid diagnosis, and accurately interpreting cardiovascular data from the UK Biobank, neurological
medical images, clinical data, and patient histories. AI data from the Alzheimer’s Disease Neuroimaging Initiative
7-9
algorithms, particularly those based on deep learning (DL) and the Human Connectome Project, and infectious
and machine learning (ML), have shown notable potential disease data from the COVIDx dataset. Comprehensive
in addressing these challenges. Recurrent neural networks preprocessing steps were implemented to standardize
(RNNs) with long short-term memory (LSTM) units, for image sizes, resolutions, and formats, ensuring data
instance, excel in analyzing time-series data, which is uniformity across the disease types. In addition, patient
crucial for cardiovascular monitoring, while support vector demographics, medical histories, and clinical parameters
machines (SVMs) effectively handle smaller, structured were anonymized to comply with data privacy standards
datasets often encountered in disease diagnostics. Despite such as the Health Insurance Portability and Accountability
these advancements, incorporating AI into clinical Act (HIPAA) and General Data Protection Regulation
practice poses significant challenges. Issues surrounding (GDPR), preserving patient confidentiality.
data quality, model bias, and the availability of extensive The analysis employed a combination of supervised and
annotated datasets present hurdles to AI deployment in unsupervised AI models tailored to the specific diagnostic
healthcare. Further, regulatory and ethical frameworks requirements of each disease category. For image-based
are essential to govern AI’s role in diagnostics, ensuring diagnostics, convolutional neural networks (CNNs)
patient safety, data privacy, and the ethical use of AI tools. were utilized due to their robust capabilities in medical
The need for international guidelines, such as TRIPOD-AI image classification and segmentation. Sequential time-
and CONSORT-AI, highlights the importance of series data, particularly for cardiovascular applications,
developing standardized reporting for AI in healthcare to were processed with RNNs incorporating LSTM units,
promote transparency and accountability. Bias within AI which are adept at capturing temporal patterns in
models, if left unaddressed, can contribute to disparities physiological signals. SVMs were applied to smaller,
in diagnostic outcomes, particularly for minority and tabular datasets, particularly where linear separability
low-resource settings, raising ethical concerns around was beneficial for diagnostic accuracy. In addition,
fairness and equity. 10-12 This study provides an in-depth unsupervised learning methods, such as autoencoders
examination of AI applications in diagnostics, presenting and variational autoencoders (VAEs), were incorporated
case studies of AI implementations and assessing the to detect anomalies in unlabeled data, aiming to identify
benefits and limitations of various models across disease rare conditions that may otherwise elude traditional
areas. In addition, emerging security considerations, such diagnostics. Separate models were trained for each disease
as fault detection and cryptography, will be discussed in category using a cross-validation approach to enhance
the context of secure AI-driven healthcare applications. generalizability and mitigate overfitting. Hyperparameter
The discussion will explore future directions for AI in tuning was conducted through grid search to optimize
diagnostics, including real-time patient monitoring, model performance, focusing on parameters such as
personalized medicine, and multispectral imaging, learning rate, batch size, and layer architecture. The
emphasizing the need for continued research to enhance models were evaluated on several key metrics, including
the robustness and applicability of AI in medical classification accuracy, sensitivity (recall), specificity,
diagnostics. precision, F1-score, and the area under the curve (AUC)

Volume 2 Issue 3 (2025) 45 doi: 10.36922/aih.5173