Vol. 1 No. 2 (2025)
Sastra Journal : Artificial Intelligence in Medicine
This essential issue, Artificial Intelligence in Medicine, curates the latest applied research demonstrating the transformative role of machine learning in clinical settings. The collection features studies on AI-driven diagnostic accuracy, including neural network models for classifying rare diseases based on imaging data (radiology and pathology) and predictive algorithms for early detection of critical conditions like sepsis and acute cardiac events. This volume is indispensable for clinicians, health policy makers, and bio-engineers seeking evidence of immediate and ethical applications of AI tools in patient care and hospital efficiency.
Published: 2025-07-28
Articles
AI-Driven Diagnosis: A Novel Approach to Early Cancer Detection
Abstract
The manual interpretation of medical images, such as CT scans and MRIs, is a cornerstone of oncological diagnosis but is susceptible to human error, inter-observer variability, and fatigue. This paper explores the application of deep learning, specifically a novel convolutional neural network (CNN) architecture, for the automated and early detection of various cancers. We trained our model on a diverse, multi-institutional dataset comprising thousands of radiological images. The proposed AI system excels at identifying subtle, early-stage malignant features often missed by the human eye. Our results demonstrate a 97.5% sensitivity and 95.8% specificity, significantly outperforming traditional computer-aided detection (CAD) systems. We discuss the model's interpretability using class activation maps (CAMs) to highlight the regions of interest, fostering clinical trust and validation. This study demonstrates the profound potential of AI to revolutionize diagnostic workflows in oncology, ultimately improving patient prognoses through more timely and accurate intervention.
Predictive Modeling of Patient Outcomes Using AI in Intensive Care Units (ICUs)
Abstract
The Intensive Care Unit (ICU) is a high-stakes environment characterized by a massive influx of real-time patient data from heterogeneous sources. Clinicians must make rapid, life-or-death decisions, often under immense pressure. This study focuses on developing an AI-driven predictive model to forecast patient deterioration 12 hours in advance. We utilize a Recurrent Neural Network (RNN) with LSTM cells to analyze multivariate time-series data, including continuous vital signs (e.g., heart rate, SpO2) and intermittent lab results. Unlike static scoring systems like APACHE II, our dynamic model captures temporal trends and complex interactions between physiological parameters. The model achieved an Area Under the Receiver Operating Characteristic (AUROC) curve of 0.89. By providing a reliable early warning score, this system can function as a crucial clinical decision support tool, enabling medical staff to allocate resources more effectively and launch proactive interventions before a critical event, such as septic shock or cardiac arrest, occurs.
Generative AI for Personalized Drug Discovery and Development
Abstract
Traditional pharmaceutical research and development is notoriously slow and expensive, with a high failure rate. This research investigates the transformative role of generative AI models, particularly Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), in accelerating personalized drug discovery. We demonstrate how these models can learn from vast libraries of known molecules to explore a multi-dimensional chemical space, designing novel, *de novo* compounds with optimized therapeutic properties. Our framework focuses on generating molecules with high predicted binding affinity to specific protein targets and low off-target toxicity. Furthermore, we integrate patient-specific genetic data into the generative process, guiding the model to create compounds tailored to an individual's unique biomarker profile. This approach to personalized medicine promises to significantly shorten the drug discovery pipeline, reduce R&D costs, and deliver more effective treatments with fewer side effects by moving beyond the one-size-fits-all paradigm.
AI-Powered Robotic Surgery: Enhancing Precision and Minimizing Invasion
Abstract
Robotic-assisted surgery has already marked a significant advance in minimally invasive procedures, offering surgeons enhanced dexterity and visualization. This research examines the next evolution: the synergy between surgical robotics and advanced artificial intelligence. We analyze how AI algorithms are being integrated to provide real-time intraoperative guidance, enhance 3D visualization with augmented reality overlays of critical anatomy, and perform autonomous sub-tasks like suturing and tremor filtering. Our study focuses on a computer vision model that identifies anatomical landmarks and instrument positions with sub-millimeter accuracy, providing real-time feedback to the surgeon to prevent errors. We present findings from simulation and ex-vivo trials indicating that AI-assistance reduces surgical time, minimizes tissue damage, and decreases the cognitive load on the surgical team. These advancements lead directly to improved patient outcomes, including less blood loss, smaller incisions, shorter hospital stays, and faster recovery times, defining the future of surgical precision.
Ethical AI in Medicine: Bias Mitigation in Clinical Decision Support Systems
Abstract
As artificial intelligence becomes more integrated into clinical decision support (CDS) systems, the risk of perpetuating or even amplifying existing health disparities becomes a critical concern. Algorithmic bias, often stemming from unrepresentative or historically skewed training data, can lead to inequitable care for minority populations. This paper provides a comprehensive review of state-of-the-art methodologies for identifying and mitigating bias in medical AI. We analyze pre-processing techniques (e.g., data augmentation, re-sampling), in-processing methods (e.g., adversarial debiasing, fairness constraints), and post-processing adjustments (e.g., fairness-aware calibration). We then propose a novel, hybrid framework for developing fair and equitable CDS tools. This framework emphasizes transparency, model interpretability, and continuous auditing. We argue that ensuring ethical AI is not merely a technical challenge but a socio-technical one, requiring collaboration between data scientists, clinicians, ethicists, and regulatory bodies to build systems that all patients can trust.
Machine Learning for Predicting Alzheimer's Disease Progression from Brain Imaging
Abstract
Identifying individuals at high risk of developing Alzheimer's disease (AD) before significant cognitive decline occurs is one of medicine's greatest challenges. This study utilizes a multi-modal machine learning approach to predict the progression from Mild Cognitive Impairment (MCI) to clinical AD. Our model integrates data from three distinct imaging modalities: structural MRI (to measure cortical atrophy), fMRI (to assess functional connectivity changes), and PET scans (to quantify amyloid-beta plaque deposition). By combining these diverse data streams, our model captures a holistic view of the disease's neurobiology. We trained a 3D CNN-based classifier on a longitudinal dataset from the Alzheimer's Disease Neuroimaging Initiative (ADNI). The model successfully predicted MCI-to-AD conversion within a 3-year window with an accuracy of 91.2%. These findings offer a promising, non-invasive tool for patient stratification in clinical trials, allowing new therapeutic interventions to be tested on the correct patient populations, and providing a crucial window for future preventative care.
A Review of Natural Language Processing (NLP) for Clinical Documentation
Abstract
Electronic Health Records (EHRs) contain a wealth of critical patient information, but the vast majority of this data is locked within unstructured clinical notes, discharge summaries, and pathology reports. This systematic review (2020-2025) charts the rapid advancements in Natural Language Processing (NLP) designed to unlock this data. We analyze the paradigm shift from traditional regex and dictionary-based methods to sophisticated deep learning models, particularly large language models (LLMs) and transformer architectures like BERT, fine-tuned for the medical domain. Key applications reviewed include: (1) information extraction for identifying patient symptoms, medications, and diagnoses; (2) automated ICD-10 coding to reduce administrative burden; (3) cohort identification for clinical trials and epidemiological research; and (4) patient summarization to support clinical handoffs. Despite significant progress, challenges in handling medical abbreviations, negations, and ensuring patient privacy (de-identification) remain. This review provides a comprehensive overview for researchers and healthcare administrators on the current capabilities and future trajectory of NLP in medicine.