Large language models (LLMs) are increasingly being incorporated into healthcare systems to support clinical communication, medical reasoning, information synthesis, patient interaction, documentation, and emerging forms of predictive analytics. Their adoption introduces a significant shift from conventional task-specific artificial intelligence toward general-purpose language-based systems capable of processing heterogeneous clinical information. However, the usefulness of these systems depends not only on linguistic performance but also on factual reliability, clinical fidelity, transparency, confidentiality, evaluation methodology, and regulatory suitability. This paper presents a structured review and analytical framework for understanding intelligent healthcare systems powered by LLMs, based exclusively on the provided literature. The analysis examines major application domains, including clinical decision support, patient-facing communication, clinical summarization, diagnostic reasoning, health trajectory forecasting, and digital-twin-oriented systems. Particular attention is given to the discrepancy between apparent model capability and clinically meaningful performance. The literature indicates that LLMs can provide substantial value in information-intensive healthcare workflows, but hallucinations, inconsistent reasoning, evaluation limitations, privacy concerns, and regulatory uncertainty constrain their safe deployment. The paper proposes a lifecycle-oriented perspective in which LLM-based healthcare systems should be evaluated through task-specific clinical benchmarks, human oversight, safety monitoring, confidentiality controls, and regulatory alignment. The findings position LLMs as augmentation technologies rather than autonomous replacements for healthcare professionals and identify research priorities involving robust evaluation, clinically grounded reasoning, patient safety, and trustworthy system architectures.