Generative AI Empowering Clinical Decision-Making: A Review of Research from Medical Record Analysis to Treatment Optimization

Abstract

This review explores generative AI’s role in empowering clinical decision-making, covering its applications,challenges, and future directions. In medical record analysis, generative AI—via NLP—extracts key data from unstructured text (e.g., TNM stages from radiology reports, SMI symptoms from discharge summaries) and generates synthetic, de-identified notes for privacy-preserving research, while also synthesizing patient timelines and clinician-friendly summaries. For diagnosis and prognosis, it creates synthetic medical images (e.g.,CMR via GANs) to augment limited datasets, predicts disease progression (e.g., CKD’s need for RRT) and severity, enables early detection, and generates differential diagnoses (e.g., GPT-4’s 98.21 F1-score for anemia subtypes). In personalized care and drug discovery, it predicts treatment responses (e.g., 73.52% accuracy for pulmonary fibrosis corticotherapy), designs tailored plans, accelerates drug development (e.g., de novo molecule design), and forecasts drug-drug interactions (e.g., via MKGFENN). Key challenges include data privacy (addressed via encryption/synthetic data), bias mitigation (to avoid care disparities), ensuring AI reliability (aided by 32-item evaluation checklists), and ethical concerns (preventing clinician over-reliance). Future directions involve integrating generative AI with RL for adaptive care, developing explainable models, and expanding to mental health (e.g., schizophrenia prognosis) and public health. Generative AI holds great promise for more efficient, equitable healthcare.