Key points are not available for this paper at this time.
This scoping review aims to identify and assess the use of pre-trained, fine-tuned, and multimodal Generative Artificial Intelligence (GenAI) models in healthcare, focusing on their application areas, reported benchmarks, and environmental impact. A comprehensive search was conducted in PubMed, Scopus, and Cochrane Central Register of Controlled Trials databases, supplemented by manual searches. Studies were selected based on inclusion criteria that focused on the application of GenAI models in healthcare. The review followed PRISMA-SCR guidelines and synthesized data on the type of models, their tasks, accuracy, and carbon footprint. Out of 7,351 initial studies, 24 studies met the inclusion criteria. The models were utilized in tasks such as classification, medical Visual Question Answering (VQA), and conversational tasks, with reported accuracies ranging from 70.6% to 99.9%. Fine-tuned models showed superior accuracy with shorter training times compared to pre-trained models. Only one study reported the carbon footprint, revealing a significant gap in environmental reporting. Multimodal AI models, although less common, demonstrated promising results in handling complex healthcare data. GenAI models hold significant potential in healthcare. However, the review highlights the underutilization of multimodal models and the lack of reporting on carbon footprints. Future researches should focus on optimizing task-specific GenAI applications while addressing their environmental impact.
Nooruddin et al. (Mon,) studied this question.