Search published articles


Showing 23 results for Learning
Diagnosing the severity of non-alcoholic fatty liver disease based on explainable artificial intelligence
Hamed Zamanian, Ahmad Shalbaf,
Volume 82, Issue 10 (1-2025)
Abstract
Background: Nonalcoholic fatty liver disease (NAFLD) represents a growing global health burden, strongly associated with rising rates of obesity, diabetes, and metabolic syndrome. This study introduces a machine learning framework to precisely diagnose NAFLD, classify disease severity, and stratify risk using routine clinical data. Our model improves early detection and risk prediction, supporting evidence-based clinical decisions. Leveraging predictive analytics, this scalable approach identifies high-risk patients and enables personalized interventions. The data-driven strategy optimizes NAFLD management by extracting maximal value from standard healthcare records, delivering both clinical and operational advantages.
Methods: This study examined 181 NAFLD patients across disease stages. The dataset was compiled from February 2010 to January 2019 at Eheim University Hospital, comprising general volunteers who were diagnosed with or without fatty liver based on histopathological evaluation of liver biopsy samples. Forward selection and mutual information identified predictive features, applied in classification models (e.g., random forest) to assess steatosis severity. Explainable AI (XAI) improved model interpretability. Combining robust feature selection, machine learning, and XAI ensured accurate, clinically actionable NAFLD severity evaluation.
Results: The XGBoost classifier with forward feature selection attained a classification accuracy of 69.23%±5.5% for steatosis severity. Interpretability analysis highlighted age, Body Mass Index (BMI), High-Density Lipoprotein (HDL), Low-Density Lipoprotein (LDL), A1c Hemoglobin (HbA1c), and glutamate pyruvate transaminase (GPT) as the most impactful variables across three severity classes. Furthermore, GPT, age, BMI, HDL, HbA1c, LDL, triglycerides, and cholesterol were critical to model performance, emphasizing their diagnostic significance in NAFLD progression. These findings suggest their utility in clinical assessments and risk stratification.
Conclusion: This study developed a machine learning model for accurate NAFLD diagnosis and severity stratification using routine clinical data. Accessible biomarkers reliably predicted disease progression, enabling gastroenterologists to facilitate early intervention. This cost-effective approach reduces healthcare costs while improving outcomes through precision medicine. Implementing such predictive tools in clinical practice could optimize resource allocation and enhance long-term NAFLD management. The framework supports timely diagnostics and targeted therapies, advancing patient-centered care.

The role of artificial intelligence in glaucoma: a review article
Zakieh Vahedian Ardakani , Mehran Zarei-Ghanavati , Hamid Riazi-Esfahani , Seyed Mehdi Tabatabaei , Mohammad Reza Mehrabi Bahar, Sadegh Ghafarian, Ahmad Masoomi,
Volume 83, Issue 1 (4-2025)
Abstract
Artificial intelligence (AI) has emerged as a transformative force in modern medicine, with ophthalmology standing at the forefront of its clinical integration. Among ophthalmic disorders, glaucoma—a leading cause of irreversible blindness worldwide—presents unique opportunities and challenges for AI-based solutions due to its chronic, progressive nature and reliance on multimodal data, including structural and functional assessments. This review article offers a comprehensive synthesis of the current and emerging roles of AI in the detection, monitoring, and management of glaucoma. AI algorithms, particularly deep learning and machine learning models, have demonstrated exceptional capabilities in interpreting fundus photographs, optical coherence tomography (OCT) images, and visual field data to identify glaucomatous damage. These systems often approach or even exceed the diagnostic performance of human experts. Moreover, AI has shown significant promise in facilitating large-scale population-based screening, improving early detection rates, and addressing disparities in access to subspecialty care, particularly in low-resource and remote settings. In the monitoring of disease progression, AI tools are being developed to detect subtle structural or functional changes over time, predict future visual outcomes, and support more precise and individualized treatment decisions. Despite these advancements, the widespread clinical adoption of AI in glaucoma care faces several critical barriers. Key limitations include poor generalizability of models across diverse populations, imaging devices, and clinical settings; scarcity of well-annotated, high-quality, and demographically representative datasets; and a lack of transparency and interpretability in algorithmic decision-making—commonly referred to as the “black box” problem. Ethical concerns, regulatory uncertainty, integration challenges within existing healthcare infrastructures, and medico-legal accountability also require thoughtful resolution before AI can be reliably deployed in clinical practice. This review critically evaluates the strengths, limitations, and real-world potential of AI technologies in glaucoma. It provides clinicians, researchers, and healthcare policymakers with a balanced and up-to-date perspective, highlighting promising avenues for future research, including explainable AI, federated learning, multi-modal data integration, and longitudinal validation studies. By fostering a deeper understanding of both the opportunities and challenges associated with AI, this article aims to guide the responsible, equitable, and evidence-based integration of AI into comprehensive glaucoma care.

Analysis of human electrocardiogram (ECG) signals using machine learning techniques
Hossein Akhavan, Fatemeh Rezaei,
Volume 83, Issue 3 (6-2025)
Abstract
Background: An Electrocardiogram is a non-invasive method for receiving heart signals. Despite advances in imaging methods, the electrocardiogram still plays an important role remains a vital tool in the diagnosis of heart diseases. Analysis of electrocardiogram signals plays an important role in the early detection of heart diseases such as arrhythmias and heart attacks. Today, with the advancement of science and technology, computer methods have received more and more attention from doctors. In this study, machine learning methods were used to classify normal and abnormal heartbeats.
Methods: The data under study were extracted from a dataset called Heartbeat published on the Kaggle website. This dataset includes samples of audio ECG signals that are divided into healthy and unhealthy categories. First, the data were preprocessed and normalized to prepare them for input into the model. Then, temporal and frequency features were extracted from the signals. Next, a hybrid model consisting of one-dimensional convolutional layers was designed and trained. Also, by using the early stopping method, overfitting was prevented and the stability of the model was improved.
Results: In this study, it was shown that by using deep learning, especially using CNN and 1D Conv, an accuracy of 0.99% and a loss of 0.0350 for test data in detecting normal and abnormal heartbeats can be achieved. This model has the ability to analyze complex structures and temporal dynamics of ECG signals and is able to detect patterns related to cardiac disorders.
Conclusion: Today, the electrocardiogram has received more attention than ever before. Appropriate selection of the model, data standardization, and a qualitative range of data are among the factors of high accuracy in this study. This study can be an effective step in the development of intelligent systems for diagnosing cardiac disorders and can be used in medical applications, especially in the field of continuous patient monitoring.

 
Page 2 from 2     

© 2026 , Tehran University of Medical Sciences, CC BY-NC 4.0

Designed & Developed by : Yektaweb