基于多模态大数据的机器学习模型的研究和应用，用于急性髓系白血病的精确输血管理

摘要(中文译文)

急性髓系白血病（AML）是一种高度异质性的血液系统恶性肿瘤，其中输血支持是患者综合诊疗的重要组成部分。本综述旨在对基于多模态大数据的机器学习模型在AML精准输血管理中的开发与临床应用最新进展进行更新性总结，以应对传统经验性输血实践长期存在的局限性。我们系统回顾了多模态数据整合（包括电子健康记录、基因组学、蛋白质组学及其他高维数据集）在AML输血管理背景下的相关文献。分析了决策树、随机森林和神经网络等机器学习算法在输血需求预测和输血反应风险评估中的应用，并参考了具有代表性的临床病例研究以展示其实际应用价值。多模态大数据在优化AML患者输血策略方面展现出重要价值。机器学习模型在预测输血需求和评估输血反应风险方面表现出良好性能，临床病例研究亦支持其实际应用价值。然而，仍存在重大挑战，包括数据隐私保护、跨平台数据标准化以及模型可解释性对临床采纳的影响。多模态大数据与先进机器学习方法的整合为在AML中实现精准、个体化输血管理带来了巨大前景。涉及联邦学习和可解释人工智能的未来方向有望解决当前局限，最终有助于改善AML患者的输血安全性和临床结局。

Abstract

Acute myeloid leukaemia (AML) is a highly heterogeneous haematologic malignancy in which transfusion support represents an essential component of comprehensive patient care. This review aims to provide an updated synthesis of recent progress in the development and clinical application of machine learning models based on multimodal big data for precision transfusion management in AML, addressing the persistent limitations of conventional, empirically guided transfusion practices. We systematically reviewed the literature on multimodal data integration-including electronic health records, genomic, proteomic and other high-dimensional datasets-in the context of AML transfusion management. The applications of machine learning algorithms such as decision trees, random forests and neural networks were analysed in the contexts of transfusion demand prediction and transfusion reaction risk assessment, with reference to representative clinical case studies demonstrating their practical utility. Multimodal big data demonstrates substantial value in optimising transfusion strategies for AML patients. Machine learning models have shown promising performance in predicting transfusion demand and assessing transfusion reaction risks, with clinical case studies supporting their practical utility. However, major challenges persist, including data privacy protection, data standardisation across platforms and model interpretability for clinical adoption. The integration of multimodal big data with advanced machine learning methodologies holds substantial promise for enabling precision, individualised transfusion management in AML. Future directions involving federated learning and explainable artificial intelligence are anticipated to address current limitations, ultimately contributing to improved transfusion safety and clinical outcomes in AML patients.

如何引用