药学学报

ML type	Definition	Common algorithm
Supervised learning	Utilizes labeled datasets (training sets) to predict outputs for unseen data through training	Linear regression; logistic regression; support vector machine (SVM); decision tree; random forest (RF)
Unsupervised learning	Discovers latent patterns or structures using unlabeled datasets	Clustering algorithms: K-means, hierarchical clustering; dimensionality reduction algorithms: principal component analysis (PCA), t-SNE
Reinforcement learning	Learns strategies through interactions with an environment to maximize cumulative rewards	Q-learning; deep Q-network (DQN)

ML type	Definition	Common algorithm
Supervised learning	Utilizes labeled datasets (training sets) to predict outputs for unseen data through training	Linear regression; logistic regression; support vector machine (SVM); decision tree; random forest (RF)
Unsupervised learning	Discovers latent patterns or structures using unlabeled datasets	Clustering algorithms: K-means, hierarchical clustering; dimensionality reduction algorithms: principal component analysis (PCA), t-SNE
Reinforcement learning	Learns strategies through interactions with an environment to maximize cumulative rewards	Q-learning; deep Q-network (DQN)

DL type	Definition	Application scenario
Convolutional neural network (CNN)	A specialized deep learning model designed for processing grid-structured data (e.g., images, audio, video). It achieves efficient feature learning and classification through convolutional operations, pooling operations, and hierarchical feature extraction. Inspired by the local perception mechanisms of biological visual systems, CNNs automatically extract spatially hierarchical features from data (e.g., edges → textures → shapes → semantics)	Image classification; object detection; semantic segmentation; video analysis; natural language processing; medical image analysis
Recurrent neural network (RNN)	A deep learning model specifically designed for sequential data processing. Its core features include recurrent structures and internal memory mechanisms. By introducing parameter sharing and state transfer across the temporal dimension, RNNs capture dynamic dependencies in sequences, making them suitable for tasks requiring contextual modeling (e.g., language, time series, speech)	Natural language processing; time series prediction; speech and audio processing; cross-modal tasks
Generative adversarial network (GAN)	A deep generative model based on a game-theoretic framework, comprising two neural networks: a generator (G) and a discriminator (D). Through adversarial training, GANs achieve data generation and distribution fitting, with the core concept summarized as the "game between a forger and an evaluator"	Image generation and enhancement; medical image analysis; drug discovery; cross-modal generation

DL type	Definition	Application scenario
Convolutional neural network (CNN)	A specialized deep learning model designed for processing grid-structured data (e.g., images, audio, video). It achieves efficient feature learning and classification through convolutional operations, pooling operations, and hierarchical feature extraction. Inspired by the local perception mechanisms of biological visual systems, CNNs automatically extract spatially hierarchical features from data (e.g., edges → textures → shapes → semantics)	Image classification; object detection; semantic segmentation; video analysis; natural language processing; medical image analysis
Recurrent neural network (RNN)	A deep learning model specifically designed for sequential data processing. Its core features include recurrent structures and internal memory mechanisms. By introducing parameter sharing and state transfer across the temporal dimension, RNNs capture dynamic dependencies in sequences, making them suitable for tasks requiring contextual modeling (e.g., language, time series, speech)	Natural language processing; time series prediction; speech and audio processing; cross-modal tasks
Generative adversarial network (GAN)	A deep generative model based on a game-theoretic framework, comprising two neural networks: a generator (G) and a discriminator (D). Through adversarial training, GANs achieve data generation and distribution fitting, with the core concept summarized as the "game between a forger and an evaluator"	Image generation and enhancement; medical image analysis; drug discovery; cross-modal generation

AI technology	Application scenarios in clinical pharmacy
ML, DL	Prediction of drug efficacy and toxicity Prediction of drug-drug interactions Prediction of pharmacokinetics Personalized medication and dose optimization MTM and antimicrobial stewardship
NLP	Pharmacovigilance and ADR monitoring Construction of knowledge graphs and medical record information extraction
LLMs	Prescription review and pharmaceutical care Clinical decision support Medication education Precision pharmacotherapy MTM Drug-drug interaction identification and prediction Pharmacovigilance

AI technology	Application scenarios in clinical pharmacy
ML, DL	Prediction of drug efficacy and toxicity Prediction of drug-drug interactions Prediction of pharmacokinetics Personalized medication and dose optimization MTM and antimicrobial stewardship
NLP	Pharmacovigilance and ADR monitoring Construction of knowledge graphs and medical record information extraction
LLMs	Prescription review and pharmaceutical care Clinical decision support Medication education Precision pharmacotherapy MTM Drug-drug interaction identification and prediction Pharmacovigilance

Strategies of AI in preventing medication error	Related AI technology
1. Analyzing massive amounts of information and identifying patterns to predict medication errors and proactively identify high-risk patients who may experience medication errors and implement preventative measures early, thereby preventing such errors from occurring	ML, DL
2. Identifying and parsing the content recorded in medical records or other healthcare documents, thereby identifying potential medication errors and providing early warnings	NLP
3. Providing real-time decision support to help prescribing physicians analyze medication prescriptions and medical orders, offering alerts regarding drug interactions, dosage, or other precautions, thereby reducing the occurrence of medication errors	ML, DL
4. Analyzing patient medication lists and individual data to identify medication usage patterns and predict potential medication errors. Simultaneously, this method enables the development of personalized medication plans based on patient-specific characteristics, reducing the incidence of medication errors	ML
5. Assisting pharmacists in medication verification and ensuring correct dispensing to the corresponding patients	Computer vision and DL
6. Using wearable devices to monitor patient medication usage and provide warnings for potential medication errors	DL
7. Integrating knowledge across multiple domains, interpreting complex textual information, and generating text, enabling earlier and more comprehensive identification of medication errors	LLMs

Strategies of AI in preventing medication error	Related AI technology
1. Analyzing massive amounts of information and identifying patterns to predict medication errors and proactively identify high-risk patients who may experience medication errors and implement preventative measures early, thereby preventing such errors from occurring	ML, DL
2. Identifying and parsing the content recorded in medical records or other healthcare documents, thereby identifying potential medication errors and providing early warnings	NLP
3. Providing real-time decision support to help prescribing physicians analyze medication prescriptions and medical orders, offering alerts regarding drug interactions, dosage, or other precautions, thereby reducing the occurrence of medication errors	ML, DL
4. Analyzing patient medication lists and individual data to identify medication usage patterns and predict potential medication errors. Simultaneously, this method enables the development of personalized medication plans based on patient-specific characteristics, reducing the incidence of medication errors	ML
5. Assisting pharmacists in medication verification and ensuring correct dispensing to the corresponding patients	Computer vision and DL
6. Using wearable devices to monitor patient medication usage and provide warnings for potential medication errors	DL
7. Integrating knowledge across multiple domains, interpreting complex textual information, and generating text, enabling earlier and more comprehensive identification of medication errors	LLMs

人工智能在临床药学中的应用进展

PDF下载

张微微 ¹^,² , 唐婧 ¹ , 冀召帅 ¹ , 胡永芳 ¹ , 聂广孟 ¹^,^* , 张翀 ²^,^*

药学学报 | 综述 2025,60(10): 3045-3059

收起

药学学报 |综述 2025 , 60 (10) : 3045 -3059

人工智能在临床药学中的应用进展

全屏

张微微¹^,², 唐婧¹, 冀召帅¹, 胡永芳¹, 聂广孟¹^,^*, 张翀²^,^*

作者信息

¹清华大学北京清华长庚医院药学部, 清华大学临床医学院, 北京 102218

²清华大学化学工程系, 北京 100084

通讯作者:

聂广孟, E-mail: ngma00216@btch.edu.cn

张翀, E-mail: chongzhang@tsinghua.edu.cn

作者简介:

作者贡献:

张微微完成了文献收集及论文初稿; 张翀及聂广孟是文章的构思者并进行了文章的修改及定稿; 唐婧、冀召帅、胡永芳参与了文章的修改。

Advances in the application of artificial intelligence in clinical pharmacy

Wei-wei ZHANG¹^,², Jing TANG¹, Zhao-shuai JI¹, Yong-fang HU¹, Guang-meng NIE¹^,^*, Chong ZHANG²^,^*

Affiliations

¹Department of Pharmacy, Beijing Tsinghua Changgung Hospital, School of Clinical Medicine, Tsinghua Medicine, Tsinghua University, Beijing 102218, China

²Department of Chemical Engineering, Tsinghua University, Beijing 100084, China

出版时间: 2025-10-12 doi: 10.16438/j.0513-4870.2025-0565

文章导航

摘要

收起

人工智能(artificial intelligence, AI) 已经逐渐融入各个领域中并发挥重要作用。随着我国老龄化时代的到来和人民群众对健康需求的增加, AI在临床药学领域的应用备受关注。临床药学通过优化药物治疗、避免用药错误对保证临床用药安全、有效非常重要。AI由于具有强大的数据处理及分析能力、基于模型的强大预测能力等优势, 在临床药学领域具有广阔应用前景。本文介绍了目前AI在临床药学领域的应用进展, 聚焦于其在提升药物有效性方面及安全性方面的应用, 包括辅助药物选择、给药剂量及给药时间优化、智能化药学服务、药物相互作用预测、药物警戒、避免用药错误, 在梳理目前应用进展的基础上同时针对其在应用中存在的挑战展开讨论, 并提出可能的解决策略, 以期为后续研究及相关从业者提供参考。

关键词

人工智能 / 临床药学 / 精准药物治疗 / 用药安全 / 应用进展

Abstract

收起

Artificial intelligence (AI) has progressively integrated into various fields and plays a significant role. With the advent of China's aging population era and the growing public demand for healthcare, the application of AI in the field of clinical pharmacy has garnered considerable attention. Clinical pharmacy is crucial for ensuring the safety and effectiveness of clinical medication use by optimizing drug therapy and preventing medication errors. Due to its advantages, such as powerful data processing and analysis capabilities, and robust model-based predictive abilities, AI holds broad application prospects in the field of clinical pharmacy. This article introduces the current progress in AI applications within clinical pharmacy, focusing on its use in enhancing drug efficacy and safety. These applications include: assisting drug selection, optimizing dosage and administration timing, providing intelligent pharmaceutical services, predicting drug-drug interactions, pharmacovigilance, and preventing medication errors. Building on a review of current advancements, the article also discusses the challenges faced in implementing these AI applications. It proposes potential solutions, aiming to provide a reference for subsequent research and relevant practitioners.

Key words

artificial intelligence / clinical pharmacy / precision pharmacotherapy / medication safety / application advance

引用本文

张微微, 唐婧, 冀召帅, 胡永芳, 聂广孟, 张翀. 人工智能在临床药学中的应用进展. 药学学报, 2025 , 60 (10) : 3045 -3059 . DOI: 10.16438/j.0513-4870.2025-0565

Wei-wei ZHANG, Jing TANG, Zhao-shuai JI, Yong-fang HU, Guang-meng NIE, Chong ZHANG. Advances in the application of artificial intelligence in clinical pharmacy[J]. Acta Pharmaceutica Sinica, 2025 , 60 (10) : 3045 -3059 . DOI: 10.16438/j.0513-4870.2025-0565

正文

收起

人工智能(artificial intelligence, AI) 最早由McCarthy等^[1]提出, 是指一种编程机器, 可以学习和识别输入和输出之间的模式和关系, 并能有效地利用这些知识对新输入的数据进行决策。AI主要通过学习大量数据之间的关联关系, 使用样本数据开发预测模型从而实现其功能^[2]。AI技术主要包括机器学习(machine learning, ML)、深度学习(deep learning, DL)、自然语言处理(natural language processing, NLP) 和机器人自动化等^[3]。大型语言模型(large language models, LLMs) 是AI的一个特定技术方向, 特指具有大规模参数的DL模型, 实现AI在语言理解和生成方面的功能^[4]。

临床药学是通过实践和研究对药物使用进行优化, 以实现以人为本的公共卫生目标的相关活动与服务^[5], 对于保证药品安全、有效使用至关重要^[6]。随着数据、算法、算力和互联网技术的进步, AI在临床药学领域的应用呈现出广阔前景。AI能综合处理药物使用过程中的多种复杂信息包括患者基础信息、疾病信息、基因组信息和药理学等信息, 分析规律并且能够对疾病风险、药物治疗反应、相互作用、不良反应等进行预测, 从而拟定个体化的治疗方案或干预措施, 对于改善患者治疗结局、提升药物治疗效果具有重要意义^{[7, 8]}。此外, AI处理信息高效, 并且能够及时发现潜在用药错误或者药物相互作用并进行预警, 从而能够减少药物不良反应(adverse drug reaction, ADR) 及用药错误的发生, 降低治疗成本、死亡率及发病率^{[9, 10]}。同时, 基于LLMs的AI技术具备复杂文本解析、多步逻辑推理、多源知识融合等能力, 在处方审核、用药监护、用药重整、用药教育及药物治疗管理(medication therapy management, MTM) 等药学服务方面具有显著优势^[11-15]。因此, AI在临床药学领域展现出变革性的潜力。本文将系统梳理AI在临床药学中的应用, 特别聚焦于AI在促进精准药物治疗和提升药物安全性方面的应用进展, 从而为临床药学工作者及相关研究者提供参考。

1　临床药学领域相关的AI技术

收起

目前应用于临床药学领域相关的AI技术主要包括ML、DL、NLP及LLMs^{[3, 16, 17]}。ML通过改变算法从而更好地模仿人类思考和学习从而提升处理复杂信息的速度和准确性^[18]。ML过程包括数据收集、数据处理、算法选择、算法训练、测试集评价、参数调整和模型应用几个步骤。ML包括监督学习、无监督学习和强化学习, 其具体分类及相关算法如表1所示^[19]。DL是ML的一种新的形式, 主要基于深层神经网络模型和方法从大量数据中进行学习、分类、预测, 常用的DL模型及相关应用场景如表2所示^[20-22]。总的来说, ML是AI的一个子集, 侧重于使用样本数据开发预测模型; DL是机器学习的一个子集, 主要通过从样本数据中学习最佳特征集从而处理更复杂的任务, 两者在预测药物作用及毒性、药物剂量、药物相互作用以及ADR等方面有广泛应用^[23-34]。NLP是AI的另外一个子集, 专注于计算机和人类通过自然语言的交互, 包括理解、解释和生成人类语言, 涉及的技术包括文本挖掘、情感分析、语音识别和机器翻译等, 在药物警戒及ADR监测方面具有优势^[35-40]。NLP的工作流程包括准备源文件、构建NLP模型、自动化数据提取^[41]。LLMs能够解读复杂文本及语言并能生成相应的答案, 在审核处方、提供用药监护、用药教育、避免用药错误等多个方面具有应用前景^{[11-15, 42-45]}, 但由于其准确性及一致性仍存在不足, 目前主要处于研究及验证阶段。ML、DL、NLP与AI的关系如图1所示, 不同AI技术在临床药学中的潜在应用领域如表3所示。

2　AI在临床药学领域中的应用进展

收起

目前AI在临床药学多个领域得到了应用, 包括个体化药物选择、药物剂量及给药时间优化、药学服务、药物警戒、药物相互作用预测以及避免药物错误等, 具体详见图2。通过AI在这些领域的应用可以促进精准药物治疗, 加强用药安全, 从而实现临床用药更加有效、安全。

2.1　AI在精准药物治疗中的应用进展

2.1.1　AI在药物选择中的应用

人类生理的多样性及个体之间的差异性可能导致不同患者对药物的反应迥异, 进而可能导致药物产生的疗效及毒性有所差别。传统药物治疗通常遵循统一的治疗原则而可能忽略个体化特征的差异^[46]。随着基因组学、蛋白组学等领域的发展, 个体化、精准化的药物治疗将成为治疗复杂疾病及严重患者的重要手段。AI可以整合诸多疾病和治疗相关信息如基因组学、蛋白质组学、药物信息等, 从而可以帮助医生针对药物选择进行快速决策^{[47, 48]}。通过运用ML、DL及自动解码器的方法可以实现对药物品种、药物剂量的选择, 并能对药物产生的疗效及毒性进行预测。

肿瘤在基因层面存在异质性, 基因差异导致患者对抗肿瘤药物的疗效不同。因此, AI在肿瘤疾病的药物选择方面具有重要价值。Chang等^[31]利用DL开发了CDRscan模型, 该模型采用卷积神经网络架构, 在每组化合物约1 000个药物反应实验的数据集上进行训练, 与其他经典ML方法(如随机森林和SVM) 相比, 他们的模型性能显著提高, 未来可能应用于根据个体患者的基因组特征选择最有效的抗癌药物。Rampášek等^[49]开发了一种基于变分自动编码器的深度生成模型—药物反应变分自编码器(Dr.VAE)。该模型可同时模拟药物反应的两个方面: 细胞存活率和转录组的扰动, 通过利用细胞系中治疗前后的基因表达变化预测不同药物的反应。Dincer等^[50]开发DeepProfile方法, 该方法通过整合变分自编码器学习急性髓系白血病患者基因表达的8维表征, 并利用该表征训练Lasso线性模型以预测药物反应, 与未进行特征提取的方法相比, 其性能显著提升。Chiu等^[51]基于癌症基因组图谱(TCGA) 数据集中的突变数据和表达特征对自编码器进行预训练, 随后训练了一个深度药物反应预测模型。其方法的核心创新在于引入预训练策略, 从而能够利用外部未标记数据(如TCGA), 而非仅依赖药物反应实验中的有限标记基因表达数据, 这一策略大幅扩展了可用样本量, 相比仅使用标记数据的模型, 其预测性能显著提高。Wang等^[52]提出了一种新的DL模型, 该模型整合了作为基因表达、拷贝数变异、基因突变、反向相蛋白阵列表达和代谢物表达的多种组学数据, 以预测对已知抗癌药物的反应。该研究证实这种方法能够捕捉基因和蛋白质的相互作用, 并有效地整合多种组学特征, 从而实现抗肿瘤药物治疗反应的精准预测。这些预测模型可以帮助临床医生判断药物反应性及耐药性, 从而确定更加敏感、高效的治疗药物。

除了单一治疗药物选择, AI针对不同的药物联合方案的选择也有较好的预测价值。联合用药可以实现协同增效的治疗效果, 特别是针对耐药菌导致的感染性疾病和肿瘤性疾病^[53-57]。但是, 联合用药也可能导致药物毒性或不良反应增加。在这种情况下就可以借助AI预测多种药物联合使用对疾病的协同治疗作用或可能的毒性作用, 以确定最佳联合用药组合。目前, 基于AI的模型已经应用于传染性疾病和肿瘤药物治疗组合的预测。Mason等^[53]根据药物分子结构的实验组合数据通过ML的方法建立了用于预测抗疟疾药物协同作用的模型。该模型针对253种药物组合进行预测, 发现其中20种药物组合具有协同作用, 9个药物组合的协同作用得到了实验验证, 其中最高的一组协同作用概率为62%。AI针对药物组合的预测作用还应用于艾滋病的药物治疗中。Wang等^[54]应用ANN、RF和SVM三种ML方法建立了抗艾滋病药物组合对艾滋病毒的作用的预测模型, 该模型同时纳入了病毒变异数据和一些临床生物标志物, 从而得到了较好的预测准确度, R²值为0.707。在肿瘤药物治疗方面, 目前已经有多种基于基因组、药物结构和生物网络数据的计算方法用于预测抗癌药物联用的协同效应。例如, Li等^[57]通过研究不同药物组合的特征, 包括药物化学结构相似性、蛋白质–蛋白质网络中的目标距离和靶向途径相似性, 使用RF算法构建了一个预测模型, 该模型能够有效识别抗肿瘤药物协同药物组合。Gayvert等^[58]同样开发了一种RF算法仅使用单一药物疗效信息就能预测协同和有效的药物组合, 研究者将该方法应用于伴BRAF突变的黑色素瘤相关40种药物及相关的780种药物组合的高通量筛选, 并预测及验证了11种既往未测试的药物组合。除了感染疾病和肿瘤, 科学家们也在慢性病方面探索应用AI预测最佳药物组合。Silva等^[59]将慢性病患者症状及主诉信息与药物基因组学及药代动力学特征信息通过“临床语义网路(CSN)”进行整合, 通过分析约90个单核苷酸多态性变异、CYP2D6基因拷贝数及临床药代动力学特征, 对具有显著药物基因组学特征的药物进行药物-基因配对和药物相互作用监测。这种将多源信息整合进行分析及预测的方法在虚拟患者中得到了验证, 提示整合的信息学与生物分析平台可为初级医疗提供便捷的临床决策支持。其他用于可视化、评估和预测药物组合的AI工具还包括SynergyFinder、Combenefit、CImbinator、DIGREM、RACS、DeepSynergy等^[60]。

由此可见, AI可以有效预测适宜的药物及药物组合, 为优化药物治疗选择提供参考。目前已经有部分AI工具应用于临床药物选择, 例如基于NLP的IBM Watson for Oncology (WFO) 可以针对患者个体情况推荐对应的肿瘤治疗药物, 目前已在我国开始使用^[61]。虽然目前尚未实现全面临床验证及应用, 但这些研究将为未来应用AI辅助临床选择更佳的药物品种提供了广阔前景。

2.1.2　AI在优化药物剂量及给药时间中的应用

精准的药物治疗除了选择正确的药物, 还要保证药物剂量和给药时间的准确性。AI对给药时间和药物剂量的预测主要通过基于药物代谢动力学模型的预测、基于临床数据和统计模型的预测、基于药物基因组学的预测以及综合使用以上三种方式进行预测。科学家们希望通过应用AI技术使药物按照零级释放, 使人体中的药物浓度保持恒定, 从而避免药物不足导致的疗效不佳以及药物过量导致的不良反应。通过应用AI可以分析药物溶解度、渗透性、剂型特征等因素通过ML的方法预测药物吸收参数及理化性质, 从而能够进一步预测血药浓度^[62]。除此之外, AI还可以通过ML识别特定代谢转化相关的结构特征, 从而预测药物代谢位点、代谢酶或转运体, 以及酶促动力学^[63]。AI通过对以上药物代谢动力学方面的预测可以帮助临床确定适宜的给药时间及给药剂量, 从而优化药物治疗效果。例如基于ML通过结合合成数据与有限真实数据进行训练, 已实现CYP2D6单倍体功能的精准预测^[64]; 通过ML建立神经网络模型能够较传统基因检测方法更精确地预测乳腺癌患者体内CYP2D6介导的他莫昔芬代谢过程^[65]。除了考虑药代动力学特征, 科学家们还将药物代谢动力学特征和临床数据相结合并利用AI建立相关预测模型。Panetta等^[32]使用蒙特卡洛树启发的方法建立一种抗肿瘤药物替莫唑胺给药剂量和时间参数的模型。在模型中应用了替莫唑胺的药物代谢动力学和药物效应动力学数据描述药物特征, 应用中性粒细胞绝对值的最低值、替莫唑胺骨髓抑制作用和生理学参数描述药物毒性。基于该模型设计的给药剂量和给药时间预测可使肿瘤显著缩小(95% CI: 7.36~7.97), 但该模型及其方案并未在实际肿瘤患者上应用^[32]。此外, 尚有其他一些AI模型应用药物吸收、分布、代谢、排泄和毒性特征, 并通过应用遗传算法进行特征选择和共轭梯度法进行参数优化以获得更佳的预测性^[62]。Yang等^[66]通过应用ML模型建立了P-糖蛋白底物和非底物的鉴定模型、人类肠道吸收的预测模型、诱导尖端扭转性心动过速的化合物的鉴定模型以及血脑屏障穿透的预测模型。Lim等^[33]开发了一种预测疾病随时间变化的剂量反应的循环神经网络模型, 即“循环边际结构网络(RMSM)”, 并将该模型和其他方法比较, 发现RMSM方法具有更小的均方根误差, 并且其预测性能较传统方法高80.9%。Steiner等^[67]通过应用多元线性回归模型和三种非线性回归模型即贝叶斯加法回归树、多元自适应回归样条和支持向量回归模型在美国拉丁裔与拉丁美洲人(ULLA) 富集的大型队列中, 比较了线性和非线性ML模型预测稳定华法林剂量的准确性。各模型测试均采用国际华法林药物遗传学联合会(IWPC) 公布的相同变量集, 以及包含种族和用药指征的扩展变量集。结果发现, 除了国际华法林药物基因组学联盟推荐的变量集之外, 纳入种族和华法林适应症两个变量可以在预测最佳剂量方面带来小幅但显著的改善。总之, 目前研究显示AI在预测给药剂量和给药时间方面具有很大的潜力, 但是仍需大量临床研究验证其可靠性。另外, 针对不同药物及疾病预测变量的选择、训练样本、模型的选择等也会影响其预测能力及临床辅助决策能力。

2.1.3　AI在智能药学服务中的应用

LLMs由于具备演绎、溯因和逻辑推理能力, 在药学服务相关工作如处方审核、用药重整、处方精简、ADR识别、用药指导及MTM等方面具有优势和应用潜力。目前已经有多种LLMs应用于临床药学领域, 其中研究较多的是ChatGPT。我国开发的DeepSeek LLM也逐渐在该领域进行了应用探索。然而, 由于LLMs也会生成错误信息, 甚至伪造数据或者提供不存在的信息, 目前LLMs在临床药学中的应用尚需要进一步的验证和优化, 相关研究也主要集中LLMs的临床药学相关工作中应用效果评价及不同LLMs间的应用差异。

Pais等^[11]针对四个LLMs即ChatGPT 3.5、Gemini Pro、Claude-Instant及Llama-2-13b在用药剂量方案错误、药物相互作用、治疗药物监测及基因组学决策支持方面进行了评价。评价标准包括回答内容的准确性、相关性、风险管理、幻觉抑制、文献引用与遵循指南情况, 错误回答根据对患者安全产生的影响被分类为重大或轻微。结果显示, 四个LLMs在识别用药剂量错误、检测药物相互作用、提供基于治疗监测和基因组学的剂量调整建议方面回答一致性较高且无显著质量差异。其中, 剂量错误识别方面, ChatGPT除辛伐他汀案例外表现最为全面; 药物相互作用检测方面, 所有模型均识别出华法林相关风险, 但均遗漏了美托洛尔与维拉帕米相互作用; 治疗药物监测建议中, Claude-Instant在两个场景给出恰当建议, 另两个场景建议也较为适用; 基因组学决策支持方面, Claude-Instant在四个场景表现满意, Gemini在三个场景表现良好^[11]。

Albogami等^[13]比较了GPT系列模型(GPT-3、GPT-3.5、GPT-4) 与人类药师针对70例真实药物咨询案例的回答质量。咨询来源主要为医师(41%) 和药师(44%), 咨询问题包括剂量/给药方案(34.2%)、药物相互作用(12.8%) 及妊娠/哺乳期用药(15.7%), 涉及成人患者占83%、女性患者占54.3%。结果显示, GPT-4的完全准确应答率达64.3%, 与人类药师相当; 在信息完整性方面, GPT-4不仅达到人类药师水平, 而且还额外提供了相关细节; 安全性评估方面, GPT-4与人类药师均实现95%的安全应答率, 而且GPT-4在70%的案例中主动提供了风险防控相关信息, 显著高于人类药师(25.7%)。该研究验证了LLMs在用药咨询中的应用价值, 可以辅助临床药师更好地提供用药咨询服务。

Socrates等^[12]评价了LLMs在老年多重用药患者中处方精简中的作用。其研究发现LLMs较培训学生能够更加高效地识别并给出处方精简建议, 从而为急诊科处方精简提供决策支持, 但其对复杂性较高的决策任务表现较差, 无法在复杂临床场景及缺乏部分临床信息的情况下给出具体处理建议。

Roosan等^[15]评价了ChatGPT不同复杂程度MTM案例中的表现。39名测试案例按照用药复杂程度分为简单、复杂、高度复杂, 评价内容包括ChatGPT 4.0在识别药物相互作用方面的能力、推荐替代方案的精确性以及制定管理计划的合理性。结果发现, ChatGPT 4.0能针对所有测试案例成功识别药物相互作用、提供治疗建议并制定总体管理计划, 但未推荐具体药物剂量, 表明其可辅助药师开展MTM工作并提高整体效率。

Shin等^[68]针对两个具体多学科会诊场景下ChatGPT在药物信息检索(涉及利妥昔单抗、华法林、圣约翰草)、用药标签错误识别、处方分析、不确定性决策支持及多学科会诊中的作用进行了研究。结果显示, 针对两个多学科会诊场景, ChatGPT能够完成相关药物信息检索并引用相关的数据库及专著; 成功识别用药标签中的错误包括药物浓度、剂型、给药途径及单位换算; 同时能够针对两个病例提供与药师给出的处理策略相当的处理方案, 显示出其在药学会诊场景下的积极作用, 从而有望成为临床药师有效的辅助工具。

Williams等^[69]比较了四款公开可用LLMs (ChatGPT 4.0、Microsoft Copilot、Google Gemini 1.5及Meta AI) 对重症成人患者药物相关问题给出的用药建议的准确性, 准确性的判断标准依据美国重症医学会(SCCM) 发布的六项指南。研究发现, 不同AI程序回答类型存在显著差异(P = 0.007)。正确建议比例由高到低依次为Microsoft Copilot、Meta AI、ChatGPT 4.0、Google Gemini。所有AI均存在错误建议, 其中Gemini错误最多, ChatGPT次之。Copilot的“其他”类回答占比最高(11.63%, 5/43)。该研究显示LLMs在解决重症患者药物相关问题方面仍存在显著的不一致性, 准确性也不稳定, 目前还难以直接应用于临床决策。

van Nuland等^[70]评价了ChatGPT在医院药学监护中解决用药相关问题方面的能力。该研究从临床药学领域收集了30个常见药学监护问题, 并以标准化方式输入ChatGPT并提供病例具体信息, 询问ChatGPT如果作为临床药师如何处理。同时该研究应用了不同语言(英语及荷兰语) 以及不同提问方式进行了同样内容的提问以验证其重复性及不同因素对回答准确性的影响。结果显示, ChatGPT能针对77%的问题提供建议, 其中26%的回答正确且完整, 22%正确但不完整, 30%部分正确部分错误, 22%完全错误。虽然ChatGPT在回答一致性方面表现佳, 但在可重复性方面差, 仅10%的回答在重复提问时保持一致。该研究同样支持了ChatGPT目前暂不适用于解决医院药师日常工作中的用药监护相关问题。

我国学者Huang等^[43]也针对ChatGPT在临床药学核心实践领域的表现, 包括处方审核、患者用药教育、ADR识别、ADR因果关系判定及用药咨询中的作用进行了评价。结果发现ChatGPT在用药咨询表现优异(ChatGPT: 8.77 vs药师: 9.50, P = 0.079 1), 但在处方审核(5.23 vs 9.90, P = 0.008 9)、患者用药教育(6.20 vs 9.07, P = 0.003 2)、ADR识别(5.07 vs 9.70, P = 0.048 3) 及ADR因果关系判定(4.03 vs 9.73, P = 0.023) 中表现较弱。由此可见, ChatGPT虽然具备强大的检索与信息整合能力及对话生成优势, 但复杂逻辑推理方面仍存在局限, 且缺乏专门的医药数据库支持。

此外, 我国学者还研究了基于DeepSeek LLMs的处方合理性评估与实施效果评价^[71]。该研究应用DeepSeek-R1模型针对5 099条(抗菌药物处方数据约10.41%, 儿童处方约1.17%, 慢病处方约40.25%) 处方信息进行分析, 结果显示, DeepSeek-R1模型在处方合理性评估上的准确率较高(93.82%), 特异度也较高(97.25%), 但模型的精确率(25.64%) 和召回率(22.73%) 相对较低, 说明模型在识别不合理处方方面存在一定的局限性, 未能充分识别出所有不合理处方^[71]。因此, DeepSeek模型现阶段尚无法独立完成类似工作, 未来还需要不断优化模型结构和训练策略, 进一步提高模型的精确率和召回率, 从而能更好地辅助药师开展处方点评工作。

由此可见, LLMs在处方审核、用药咨询、处方精简、药物治疗管理、药学会诊、药学监护、ADR识别方面均显示出一定的应用潜力, 但由于不同模型一致性差异、对复杂场景难以分析、准确性不稳定、精确率及召回率差等局限性尚无法真正独立应用于相应的临床药学工作。

2.2　AI在药物安全中的应用进展

2.2.1　AI在药物警戒中的应用

药物警戒是发现、评估、理解和预防药物不良作用或任何其他与药物相关问题的科学与活动, 其贯穿从药物研发到使用直至退出市场的全生命周期, 核心思想是防控用药风险, 保障患者与公众安全^[72]。医疗工作者, 特别是药师在药物警戒中扮演者重要角色, 其通过识别、评价、报告、记录、解决、监测、预防ADR等多个环节开展药物警戒工作。传统的基于人工的ADR上报工作以自发性报告为主, 存在严重的报告不足问题。研究显示仅有约5%~10%的ADR被记录及报告^{[73, 74]}。其次, 传统报告方法需要数月甚至更长的时间才能识别到安全信号, 例如罗非昔布(Vioxx) 所致的心血管风险在其上市5年后才被确认, 从而导致了近三万名患者受害^[75]。此外, 传统ADR监测工作中大部分药物安全信息来源于非结构文本(例如临床记录) 需要人工进行信息提取及重新记录和上报工作, 可能导致重要信息缺失或错误。不同上报人员对ADR因果判定也存在主观性, 缺乏统一的客观标准^[76]。由此可见, 传统基于人工的ADR监测工作不仅耗费人力及时间, 还存在信息漏报、时间延误及准确性不佳等局限性。

AI的诞生成为药物警戒的有力工具, 在很多方面可以弥补传统方式的不足。在数据处理速度及准确度方面, 一方面AI可以通过NLP自动提取非结构文本中的关键信息(例如ADR术语、药物名称等), 显著缩短单个病例安全报告(individual case safety report, ICSR) 处理时间^[36-38]; 另外一方面, AI可以整合电子病历、基因组信息、社交媒体等多模态信息, 从而构建全景式药物安全图谱^{[77, 78]}。由此, AI可以实现自动化报告生成及多源数据整合, 从而提升数据处理速度及准确度。在ADR预警及安全信号监测方面, AI通过ML模型(如RF、DL) 可以实现数据流的实时分析, 能够提前数周至数月识别潜在安全信号^[37], 例如AI在COVID-19疫苗监测中快速捕捉心肌炎的风险^[73]。同时, AI能够获取除了电子病历、医疗文书之外的信息源, 例如社交媒体, 从而能够从更广泛的信息中实现信息挖掘, 补充传统报告方式的盲区。在风险评估方面, AI模型通过分析时间关联性、去混杂因素(如患者基础疾病) 和文献证据链, 从而提供客观的因果关系概率评分, 避免了人为主观因素的干扰^[79]。此外, AI能够基于患者基因组、代谢特征和用药史建立预测模型从而预测个体化ADR风险^[80]。

已有学者应用ML的方法从社交媒体如Twitter收集潜在的ADR或药物不良事件(adverse drug event, ADE), 从而发现医疗人员无法获取的ADR及ADE, 并对信息进行快速处理及分析^[81]。但不可否认的是, 这种收集方式往往需要排除“噪音”的影响, 此外还需要对非正式文本或不规则文本进行进一步的处理。康涅狄格儿童医学中心利用ML的技术将基于规则查询和半监督ML结合简化了ADE报告效率^[82]。针对药物上市前的研究也可以用来训练ML模型, Ménard等^[83]使用104个已完成的临床研究中的数据包括人口统计数据、生命体征和疾病数据训练ML模型, 从而可以用于ADE发生数量的预测。针对某些特殊疾病, 例如糖尿病, 通过一种NLP系统—HypoDetect, 可以及时捕捉到患者体内血糖值, 及时发现患者低血糖的情况, 从而有助于及时进行治疗^[84]。我国学者以国家食品药品监督管理总局、美国食品药品监督管理总局及某大型三甲医院药品库中药品说明书为数据源, 构建了基于NLP的临床合理用药知识图谱, 能够在标引少量药品说明书的前提下挖掘出药品说明书中所有用药相关信息从而有助于促进合理用药, 减少药物不良反应^[40]。另有国内学者也通过NLP技术对本院住院患者的病历数据进行挖掘和利用, 建立ADR主动监测方案, 从而可以提高ADR监测的敏感性, 提高ADR的识别能力, 促进患者用药安全^[85]。

AI可以针对患者报告的药物不良反应或事件中的文本进行分类处理, 并且已有学者利用这一技术评价药物不良反应或事件后果的严重程度, 发现通过NLP可以识别发生严重不良反应或者死亡的病例^[86]。然而, 目前这种技术并不能完全替代人工审查, 因为医学术语等专业程度高的语言还很难处理。Yang等^[87]通过应用DL模型可以识别医院药物安全性报告中关于过敏反应的文本描述, 从而可以针对过敏患者进行及时干预及防护, 避免医疗过失, 提升安全性。

此外, AI还可以通过ML的方法识别药物不良反应高危人群。Chandak和Tatonetti^[88]发明了一项针对女性药物毒性风险分析的ML算法—AwareDX, 他们通过应用倾向性评分匹配的方法从而精准预测由性别差异导致的药物不良反应风险。此外, 还有学者利用ML模型识别二氢吡啶脱氢酶(DPD) 缺乏人群, 这类人群由于DPD缺乏而容易出现氟嘧啶毒性^[89]。

总之, AI在药物警戒中的应用包括提升药物不良反应上报效率及早期预警, 整合多模态信息并及时捕获风险信号, 预测不良反应后果及识别高危人群等, 其在药物警戒中的应用逐渐得到证实, 但是其未来广泛应用还需要依靠稳健有效的训练数集、兼具高度敏感性及特异性的算法以及完备且标准的数据来源。

2.2.2　AI在预测药物相互作用中的应用

多重疾病患者常需要同时服用多种药物治疗疾病, 在使用多种药物时特别需要关注的一点就是药物相互作用。药物相互作用不仅会影响治疗效果还可能产生用药风险。美国每年有超过74 000人因药物相互作用导致的不良后果而在急诊治疗, 超过195 000人因不良药物相互作用导致住院治疗^[19]。在中国, 药物相互作用的发生率约为5%, 每年约有250万人因药物滥用或不良反应而住院^[90]。药物在研发过程中虽然会在临床试验中针对某些药物的相互作用进行研究, 但是由于真实世界中使用药物种类庞大, 很难逐一证实所有药物之间是否存在相互作用。此外, 通过临床试验验证所有药物相互作用通常需要花费大量金钱和时间, 因此较难完成大量药物相互作用的验证。AI的诞生为解决多种药物复杂相关作用提供了解决方案, AI可以通过分析和学习大量药物相关数据库从而可以识别不同数据之间的关联, 实现对药物相关作用的预测。目前已有部分研究显示AI在预测药物相互作用方面具有一定价值。

Ryu等^[91]在2018年通过调取DrugBank数据库中药物及发生相互作用事件数据, 包括1 704种药物和191 400个药物相互作用关联, 应用DL的方法建立药物-药物相互作用和药物-食物相互作用预测模型, 该模型能够准确预测其中86种药物相互作用类型。Hou等^[92]在2019年也通过从DrugBank中提取药物数据并构建了一种深度神经网络架构模型, 该模型从4 432个药物特征中发现了80种不同的药物相互作用事件。随后, Deng等^[93]在2020年也通过分析DrugBank中65种药物-药物相互作用事件数据, 并使用多个数据源和药物作用中涉及的化学亚基、靶点、酶和通路以及计算相似性作为特征, 构建了一个多通道深度神经网络模型从而实现对药物相互作用的预测功能。近年来, 更多研究药物相互作用的AI模型逐渐涌现, 虽然使用的数据库和方法不同, 但均展现了较好的预测效果^[94-98]。另外, 随着大量LLMs的涌现, 学者们开始尝试应用不同的LLMs进行药物相互作用识别及严重程度分级, 从而避免药物相关性损害^[42]。

由此可见, 随着各种AI预测模型的涌现, AI将在预测药物相关作用方面扮演越来越重要的角色, 但是由于这些模型参考的数据库及数据来源存在差异, 尚无法确定哪种方法预测价值更高, 未来仍需相关实验进行研究。

2.2.3　AI在避免用药错误中的应用

用药错误是指药品在临床使用及管理全过程中出现的、任何可以防范的用药疏失, 这些疏失可导致患者发生潜在的或直接的损害^[99]。用药错误直接关系到患者的用药安全, 目前已成为全球医疗领域的主要致害因素, 每年造成约420亿美元的经济损失, 在美国每天至少有1例患者因用药错误导致死亡, 每年因用药错误导致受伤的人群高达130万^[100]。目前用于避免及减少用药错误的手段包括使用电子化系统例如电子处方^[101]、条码技术^[102]、自动化药房例如单剂量分包机^[102]、标准化流程例如应用高警示药品标识^{[99, 102]}、双重核对制度^[102]; 加强人员管理以及开展相关教育与培训^[99]等。

近年来AI逐渐应用于减少和避免用药错误, 其在避免用药错误方面的手段及相关技术如表4所示^{[11, 103-109]}。

AI在防止用药错误方面的作用在一些研究中已经得到证实。2023年发表的一篇系统综述显示, 71%的研究显示与传统临床实践相比, AI能够减少用药错误, 凸显了AI在初级医疗环境中合理应用的可行性^[110]。另外一项2024年发表的针对医院环境中用药警报的系统综述显示, 与当前广泛应用的临床决策支持系统生成的药物警报相比, 基于AI优化的药物警报可提高不当或非典型处方的识别率, 减少用药错误的出现, 并支持使用者进行响应及预测, 但基于AI的警报系统尚需更多的外部验证及临床实践验证^[105]。Zhao等^[111]探索并研究了利用AI进行居家药物自我管理, 其研究显示应用内置于传感器的AI系统可以解析环境无线电反射信号, 追踪患者使用吸入器或胰岛素笔的特定动作, 并检测患者是否遵循规范步骤(如胰岛素注射前是否排气、吸入器使用前是否摇晃) 进行给药, 当发现操作错误时, 系统自动生成警报, 从而实现了非接触式自我用药的被动监测, 避免用药错误的发生。Pais等^[11]将药学领域专业知识与LLMs相结合建立了一个名为“MEDIC”的系统, 该系统能够模拟药剂师的临床推理逻辑对处方进行审查, 通过测试发现, 该系统能够及时发现并减少33%的用药错误。

由此可见, AI在避免用药错误方面具有显著优势并且已经逐渐应用于临床实践, 虽然未来还需更多的外部验证及临床实践验证, 其在用药监测、错误预警及自我用药管理方面的应用将极大地提高临床医生及药师的工作效率, 提高患者用药的安全性。

3　AI在临床药学中应用的挑战及对策

收起

3.1　伦理方面的挑战及对策

AI在临床药学领域虽然有相当大的应用优势, 但是不可否认的是, AI通过ML和DL实现的预测结果大多时候仅能提供预测结果但无法给出内部学习及计算机制。例如, 应用ML及DL用于药物反应、药物剂量及给药时间等预测时, 可以获取相关的预测结果, 但是无法了解具体计算过程及预测机制。这种“过程不透明性”导致人类对AI决策结果往往只能选择信任, 但是无法解释其决策原因^[112]。针对这一挑战目前有学者提出可以通过“可解释性人工智能(XAI)”进行解决, XAI可以克服基于人AI应用程序理解不足的问题, 并提高这些应用程序和内部机制在关键决策领域的可解释性^{[113, 114]}。

另外一个伦理方面的挑战是数据来源的不均衡。AI应用于临床药学时需要基于大量数据, 而数据来源可能无法代表所有人类群体, 因此其模型应用时可能会导致偏差^[115]。例如, 非洲裔儿童患者平均接受的阿片类镇痛剂量低于白人儿童患者, AI系统若基于此类医疗记录学习, 可能延续系统性偏见, 而非基于生物学差异, 错误建议降低非洲裔儿童患者的镇痛剂量^[116]。我国幅员辽阔, 有多个少数民族, 不同民族间可能也存在群体差异, 因此如果仅应用单一的数据来源也可能影响AI的准确性。因此, 在构建临床药学相关AI模型时可能需要纳入多样化的人群数据, 尽可能代表所有人类群体。

此外, 针对数据来源的隐私保护和数据安全性也值得关注。患者的健康信息具有高度敏感性, 因此保护患者健康信息至关重要^{[117, 118]}。既往已有在未经患者知情同意的情况下应用其健康数据进行AI训练而引起争议的先例^{[119, 120]}。尽管目前AI能够实现数据保护并防止网络攻击, 但是仍有AI系统被黑客入侵或者植入虚假数据的可能性^[112]。针对于此, 将AI利用到临床药学领域时应建立相关法律制度及伦理规范进行约束。欧盟已于2018年制定了通用数据保护条例(GDPR) 用于保护数据安全规范AI使用^[121]。此后欧盟委员会又制定了人工智能法案(AIA) 以解决人工智能相关社会风险的问题^[122]。我国在2024年也发布了《中华人民共和国数据安全法》、《中华人民共和国个人信息保护法》, 对数据安全提供了法律支持^[123], 但目前尚未有专门针对AI在临床药学领域应用的具体法律。未来可能需要更多更有针对性的法律对AI应用于临床药学工作进行规范和引导。此外, 要想在临床药学中更好地应用AI, 必须在获得患者知情同意的基础上通过一定的技术手段充分保护患者的个人健康信息。

其他伦理方面的挑战还包括基于AI的临床用药指导及决策如果出现错误后其责任归属以及AI交互能力不足等。AI在临床药学应用时如果出现错误决策可能导致威胁健康的严重后果, 如果出现此类情况如何确定责任主体仍是一大难题^[124], 在一些医疗场景中, 参与患者药物治疗的不仅包括医生、患者、护士, 而且还包括临床药师, 多种角色参与以及不同角色AI的使用使其责任归属模糊。如果归咎于临床使用者但其并未参与算法开发或控制, 如果归咎于开发者但其并未参与临床直接应用^[119]。除此之外, 利用AI开展临床药学工作可能无法实现人文化的关怀和照护, 例如无法综合考虑患者的经济状况、家庭支持等情况而为其提供更适合的、更个性化的治疗药物和方案, 也无法实现与患者共情以及提供情感支持。针对以上情况, 未来可能需要制定相关政策及制度限定不同AI使用者在临床用药决策中的权责及具体使用情景。此外, 未来可能需要综合AI语音识别、情感分析等技术, 例如通过摄像头捕捉患者在用药咨询时的面部表情以辅助判读, 从而从一定程度上提升AI的交互性。

3.2　技术方面的挑战

AI在技术方面的一大挑战是数据整合^[125]。如前所述, AI在药物选择、药物剂量及给药时间优化、药物相互作用及不良反应预测时通常基于ML、DL预测模型, 而模型的建立则需要纳入相关数据变量, 如果变量参数与结果关联过多时可能导致算法依赖不恰当的特征进行预测^[126], 而如果变量与结果之间找不到关联时则会出现过度拟合。因此, 为了更好预测模型, 就需要纳入所有相关的变量, 并对相关的多源数据包括文本、数值、影像及视频数据通过同一算法进行整合。然而, 医疗数据的多样性及庞大使这些数据的整合并不容易^[127]。此外, 不同地区及医疗机构所使用的药物名称、疾病名称、不良反应名称等可能并不相同, 针对同一种疾病的描述可能也存在偏差, 其他数据来源例如社交媒体中使用的语言或者描述可能并非专业术语及药品通用名称, 而且可能存在方言、错误拼写、缩写和信息遗漏, 由此可能导致AI无法精确识别和捕捉具体的含义^[79], 阻碍其更好地应用。

另外一个技术挑战是数据质量和数量问题。高质量、标准化的数据是AI算法的基础, 但目前临床药学领域缺乏足够且统一的数据。医院采集的数据常存在质量低、准确性差及数据不完整等问题, 因此可能造成AI应用障碍^[119]。如果选择的算法与数据不匹配或者数据可靠性不佳无法满足神经网络、决策树、贝叶斯网络等分类算法, 也会造成ML决策错误^[127]。此外, 数据获取通常需要高额成本, 另外隐私法规限制了数据共享, 临床用药数据缺失、质量不高等都会影响AI模型的可靠性和准确性^[128]。

这些问题会严重影响AI在临床药学中的应用。例如, AI处方审核系统多依赖本地历史数据训练, 因此存在地区差异和时间差异, 难以覆盖所有地区的药品及处方问题, 也没有办法针对新出现的药品进行审核, 因此使其应用受到局限。在治疗药物选择及用法用量决策支持方面, 如果无法良好地整合药物信息、基因组学和代谢组学等信息, 那么AI就难以推荐最佳的候选药品和方案。在药物警戒和不良反应识别方面, 如果没有充足、完整的数据来源就无法保证AI模型的准确性, 进而无法识别药物相关风险信号。

针对以上这些问题, 可以尝试通过促进数据共享、建立统一的数据标准、对数据进行清洗及验证等方式应对。例如通过建立跨区域、多中心的医疗数据库, 整合不同医疗机构的用药信息, 从而提升AI模型的泛化能力。通过推动国际通用的数据格式(例如FHIR标准), 建立药物警戒联盟共享脱敏数据, 从而更好地推动AI在药物警戒中的利用。通过不断增强AI学习能力, 实现高质量数据筛选及融合, 从而为药物治疗决策提供更可靠的支持和参考。此外, 还可以通过应用开源工具和库降低开发门槛与加速研究进程、应用云计算平台实现弹性资源与安全合规、应用迁移学习破解数据稀缺难题、使用协作注释工具提升标注效率与质量、促进医疗机构与科技公司协作更好地开发适合特定临床用药场景的AI模型等方法进行解决^[129]。

3.3　应用方面的挑战

虽然很多研究证实了AI在临床药学领域具有重大应用价值, 但目前临床药学领域AI研究多为受控条件下的回顾性分析, 缺乏前瞻性临床试验证实其干预的有效性, 从而可能导致研究结果无法外推到真实的临床场景, 限制了其在临床中的应用^[130]。因此, 在当前情况下, AI只能作为临床药学工作中的辅助工具, 最终决策仍需要药师结合患者的具体情况综合判断。后续可能需要更多前瞻性临床试验及外推试验验证AI在临床药学领域的应用价值。

此外, AI在临床药学中的应用需要具备一定的人力和基础设施, 例如开发人员、维护人员、数据存储设施及备份设施等, 这些需要一定资金投入和人力投入^[131]。此外, 目前药师针对AI应用仍存在一定的技术壁垒, 例如无法确定选择何种AI技术, 如何解读AI预测结果及回答内容等, 这对于AI有效使用造成了障碍。例如, 药师在利用AI进行药物相互作用分析时, 需要理解AI的决策逻辑, 这样才能在最终决策时正确参考AI的建议。而对于开发者来说, 则需要充分考虑各种临床药学场景, 例如药物治疗管理和不良反应分析等需求差异。针对于此, 未来在临床药学领域应用AI可能需要增加相关资金及人力投入, 并针对药师开展相关AI课程培训, 从而促进AI技术更好地应用于临床用药工作。对于AI技术开发者来说, 则应了解临床药学相关工作内容, 并根据不同应用场景的具体需求优化AI, 使使用者能够更好地了解相关功能, 充分地应用AI提升工作效率。此外, 需要将AI整合到药师工作流中, 例如在进行处方审核时, 既显示处方中存在的问题, 也能显示问题风险及相关参考文献, 从而增加AI应用的可解释性, 建立流畅的人机协作流程。

4　总结及未来展望

收起

AI在临床药学领域已经在多个方面展现出广阔前景, 通过其在药物选择、剂量及给药时间优化、智能药学服务等方面可以促进药物治疗的有效性; 通过其在药物警戒、药物相互作用及用药错误的识别及预警中的应用可以促进药物治疗的安全性。但是后续如何更好地将AI应用于临床药学还需要克服伦理、技术及应用方面的挑战。未来可能需要聚焦于如何将目前开发的AI技术与临床药学实践进行整合, 真正能运用AI提升药师决策能力, 为患者提供更加个体化的药物治疗。这需要在大量相关前瞻性研究、伦理及隐私框架、患者的积极参与等多方面支持下确保其能在真实世界中的有效验证及安全实施^[17]。相对于其他国家, 我国AI在临床药学中的应用还需要大量探索和验证。AI的快速发展将进一步推动临床药学工作的智能化转型, 为患者提供更高效、更精准、更安全的临床药学服务。

参考文献

收起

文献

收起

参考文献引证文献

排序方式：

McCarthy J, Minsky ML, Rochester N, et al. A proposal for the dartmouth summer research project on artificial intelligence [J]. AI Mag, 2006, 27: 12-14.

Muthukrishnan N, Maleki F, Ovens K, et al. Brief history of artificial intelligence [J]. Neuroimaging Clin N Am, 2020, 30: 393-399.

Lu M, Yin J, Zhu Q, et al. Artificial intelligence in pharmaceutical sciences [J]. Engineering, 2023, 27: 37-69.

Shah NH, Entwistle D, Pfeffer MA. Creation and adoption of large language models in medicine [J]. JAMA, 2023, 330: 866-869.

Dreischulte T, van den Bemt B, Steurbaut S. European Society of Clinical Pharmacy definition of the term clinical pharmacy and its relationship to pharmaceutical care: a position paper [J]. Int J Clin Pharm, 2022, 44: 837-842.

Claire M, Claire A, Matthew B. The role of clinical pharmacists in general practice in England: impact, perspectives, barriers and facilitators [J]. Res Social Adm Pharm, 2022, 18: 3432-3437.

Bajwa J, Munir U, Nori A, et al. Artificial intelligence in healthcare: transforming the practice of medicine [J]. Future Healthc J, 2021, 8: e188-e194.

Javaid M, Haleem A, Pratap Singh R, et al. Significance of machine learning in healthcare: features, pillars and applications [J]. Int J Intel Netw, 2022, 3: 58-73.

Davenport T, Kalakota R. The potential for artificial intelligence in healthcare [J]. Future Healthc J, 2019, 6: 94-98.

Bohr A, Memarzadeh K. Chapter 2-The rise of artificial intelligence in healthcare applications [M]// Bohr A, Memarzadeh K. Artificial Intelligence in Healthcare. Copenhagen: Academic Press, 2020: 25-60.

Pais C, Liu J, Voigt R, et al. Large language models for preventing medication direction errors in online pharmacies [J]. Nat Med, 2024, 30: 1574-1582.

Socrates V, Wright DS, Huang T, et al. Identifying deprescribing opportunities with large language models in older adults: retrospective cohort study [J]. JMIR Aging, 2025, 8: e69504.

Albogami Y, Alfakhri A, Alaqil A, et al. Safety and quality of AI chatbots for drug-related inquiries: a real-world comparison with licensed pharmacists [J]. Digit Health, 2024, 10: 20552076241253523.

Sridharan K, Sivaramakrishnan G. Unlocking the potential of advanced large language models in medication review and reconciliation: a proof-of-concept investigation [J]. Explor Res Clin Soc Pharm, 2024, 15: 100492.

Roosan D, Padua P, Khan R, et al. Effectiveness of ChatGPT in clinical pharmacy and the role of artificial intelligence in medication therapy management [J]. J Am Pharm Assoc (2003), 2024, 64: 422-428. e8.

Mottaghi-Dastjerdi N, Soltany-Rezaee-Rad M. Advancements and applications of artificial intelligence in pharmaceutical sciences: a comprehensive review [J]. Iran J Pharm Res, 2024, 23: e150510.

Chalasani SH, Syed J, Ramesh M, et al. Artificial intelligence in the field of pharmacy practice: a literature review [J]. Explor Res Clin Soc Pharm, 2023, 12: 100346.

Al Kuwaiti A, Nazer K, Al-Reedy A, et al. A review of the role of artificial intelligence in healthcare [J]. J Pers Med, 2023, 13: 951.

Chen S, Li T, Yang L, et al. Artificial intelligence-driven prediction of multiple drug interactions [J]. Brief Bioinform, 2022, 23: bbac427.

Wang R, Wang S, Shi TL. An introduction to the application of deep learning in the biomedical field [J]. Sci Sin Vitae (中国科学: 生命科学), 2024, 54: 1268-1287.

Zhang HY, Liu YL. Research based on deep learning technology [J]. Cybersecur Technol Appl (网络安全技术与应用), 2024, (08): 41-43.

Wenzel M. Machine learning for brain disorders [M] // Wenzel M. Generative Adversarial Networks and Other Generative Models. New York: Humana, 2023: 139-192.

Amorim AMB, Piochi LF, Gaspar AT, et al. Advancing drug safety in drug development: bridging computational predictions for enhanced toxicity prediction [J]. Chem Res Toxicol, 2024, 37: 827-849.

Rafique R, Islam SMR, Kazi JU. Machine learning in the prediction of cancer therapy [J]. Comput Struct Biotechnol J, 2021, 19: 4003-4017.

Khan MZI, Ren JN, Cao C, et al. Comprehensive hepatotoxicity prediction: ensemble model integrating machine learning and deep learning [J]. Front Pharmacol, 2024, 15: 1441587.

Hajim WI, Zainudin S, Mohd Daud K, et al. Optimized models and deep learning methods for drug response prediction in cancer treatments: a review [J]. PeerJ Comput Sci, 2024, 10: e1903.

Kong W, Midena G, Chen Y, et al. Systematic review of computational methods for drug combination prediction [J]. Comput Struct Biotechnol J, 2022, 20: 2807-2814.

Wu L, Wen Y, Leng D, et al. Machine learning methods, databases and tools for drug combination prediction [J]. Brief Bioinform, 2022, 23: bbab355.

Tran TTV, Tayara H, Chong KT. Recent studies of artificial intelligence on in silico drug distribution prediction [J]. Int J Mol Sci, 2023, 24: 1815.

Tran TTV, Tayara H, Chong KT. Artificial intelligence in drug metabolism and excretion prediction: recent advances, challenges, and future perspectives [J]. Pharmaceutics, 2023, 15: 1260.

Chang Y, Park H, Yang HJ, et al. Cancer drug response profile scan (CDRscan): a deep learning model that predicts drug effectiveness from cancer genomic signature [J]. Sci Rep, 2018, 8: 8857.

Panetta JC, Kirstein MN, Gajjar AJ, et al. A mechanistic mathematical model of temozolomide myelosuppression in children with high-grade gliomas [J]. Math Biosci, 2003, 186: 29-41.

Lim B, Alaa A, Schaar MVD. Forecasting treatment responses over time using recurrent marginal structural networks [C] // Proceedings of the 32nd International Conference on Neural Information Processing Systems, 2018. Montréal: Curran Associates Inc., 2018: 7494-7504.

Simpson MD, Qasim HS. Clinical and operational applications of artificial intelligence and machine learning in pharmacy: a narrative review of real-world applications [J]. Pharmacy (Basel), 2025, 13: 41.

Del Rio-Bermudez C, Medrano IH, Yebes L, et al. Towards a symbiotic relationship between big data, artificial intelligence, and hospital pharmacy [J]. J Pharm Policy Pract, 2020, 13: 75.

Tao W, Mingjie Z, Hongliang L, et al. Current applications of artificial intelligence in pharmacovigilance in the USA and implications [J]. Chin Drug Monit (中国药物警戒), 2023, 20: 1129-1133.

Kompa B, Hakim JB, Palepu A, et al. Artificial intelligence based on machine learning in pharmacovigilance: a scoping review [J]. Drug Saf, 2022, 45: 477-491.

Golder S, Xu D, O'Connor K, et al. Leveraging natural language processing and machine learning methods for adverse drug event detection in electronic health/medical records: a scoping review [J]. Drug Saf, 2025, 48: 321-337.

Zheng C, Lee MS, Bansal N, et al. Identification of recurrent atrial fibrillation using natural language processing applied to electronic health records [J]. Eur Heart J Qual Care Clin Outcomes, 2024, 10: 77-88.

Zhang XL, Wang ZM, Wang YQ, et al. Construction of knowledge graphs for rational use of drugs in clinical practice based on natural language processing [J]. Chin J Med Libr Inf Sci (中华医学图书情报杂志), 2019, 28: 1-5.

Hsu JC, Wu M, Kim C, et al. Applications of advanced natural language processing for clinical pharmacology [J]. Clin Pharmacol Ther, 2024, 115: 786-794.

Ong JCL, Chen MH, Ng N, et al. A scoping review on generative AI and large language models in mitigating medication related harm [J]. NPJ Digit Med, 2025, 8: 182.

Huang X, Estau D, Liu X, et al. Evaluating the performance of ChatGPT in clinical pharmacy: a comparative study of ChatGPT and clinical pharmacists [J]. Br J Clin Pharmacol, 2024, 90: 232-238.

Sridharan K, Sivaramakrishnan G. Investigating the capabilities of advanced large language models in generating patient instructions and patient educational material [J]. Eur J Hosp Pharm, 2024. DOI: 10.1136/ejhpharm-2024-004245.

Li D, Wu L, Lin YC, et al. Enhancing pharmacogenomic data accessibility and drug safety with large language models: a case study with Llama3.1 [J]. Exp Biol Med (Maywood), 2024, 249: 10393.

Hripcsak G, Ryan PB, Duke JD, et al. Characterizing treatment pathways at scale using the OHDSI network [J]. Proc Natl Acad Sci U S A, 2016, 113: 7329-7336.

Álvarez-Machancoses Ó, DeAndrés Galiana EJ, Cernea A, et al. On the role of artificial intelligence in genomics to enhance precision medicine [J]. Pharmgenomics Pers Med, 2020, 13: 105-119.

Pandi MT, Koromina M, Tsafaridis I, et al. A novel machine learning-based approach for the computational functional assessment of pharmacogenomic variants [J]. Hum Genomics, 2021, 15: 51.

Rampášek L, Hidru D, Smirnov P, et al. Dr. VAE: improving drug response prediction via modeling of drug perturbation effects [J]. Bioinformatics, 2019, 35: 3743-3751.

Dincer A, Celik S, Hiranuma N, et al. DeepProfile: deep learning of patient molecular profiles for precision medicine in acute myeloid leukemia [J]. BioRxiv, 2018. https://doi.org/10.1101/278739.

Chiu YC, Chen HH, Zhang T, et al. Correction to: predicting drug response of tumors from integrated genomic profiles by deep neural networks [J]. BMC Med Genomics, 2019, 12: 119.

Wang C, Lye X, Kaalia R, et al. Deep learning and multi-omics approach to predict drug responses in cancer [J]. BMC Bioinformatics, 2022, 22: 632.

Mason DJ, Eastman RT, Lewis RPI, et al. Using machine learning to predict synergistic antimalarial compound combinations with novel structures [J]. Front Pharmacol, 2018, 2: 1096.

Wang D, Larder B, Revell A, et al. A comparison of three computational modelling methods for the prediction of virological response to combination HIV therapy [J]. Artif Intell Med, 2009, 47: 63-74.

Prosperi MC, Altmann A, Rosen-Zvi M, et al. Investigation of expert rule bases, logistic regression, and non-linear machine learning techniques for predicting response to antiretroviral treatment [J]. Antivir Ther, 2009, 14: 433-442.

Xia F, Shukla M, Brettin T, et al. Predicting tumor cell line response to drug pairs with deep learning [J]. BMC Bioinformatics, 2018, 19: 486.

Li X, Xu Y, Cui H, et al. Prediction of synergistic anti-cancer drug combinations based on drug target network and drug induced gene expression profiles [J]. Artif Intell Med, 2017, 83: 35-43.

Gayvert KM, Aly O, Platt J, et al. A computational approach for identifying synergistic drug combinations [J]. PLoS Comput Biol, 2017, 13: e1005308.

Silva P, Jacobs D, Kriak J, et al. Implementation of pharmacogenomics and artificial intelligence tools for chronic disease management in primary care setting [J]. J Pers Med, 2021, 11: 443.

Adam G, Rampášek L, Safikhani Z, et al. Machine learning approaches to drug response prediction: challenges and recent progress [J]. NPJ Precis Oncol, 2020, 4: 19.

Zhou N, Zhang CT, Lv HY, et al. Concordance study between IBM Watson For Oncology and clinical practice for patients with cancer in China [J]. Oncologist, 2019, 24: 812-819.

Li LJ, Ji PY, Zheng AL, et al. Research progress of artificial intelligence combined with physiologically based pharmacokinetic models [J]. Acta Pharm Sin (药学学报), 2024, 59: 2491-2498.

Yang X, Wang Y, Byrne R, et al. Concepts of artificial intelligence for computer-assisted drug discovery [J]. Chem Rev, 2019, 119: 10520-10594.

McInnes G, Dalton R, Sangkuhl K, et al. Transfer learning enables prediction of CYP2D6 haplotype function [J]. PLoS Comput Biol, 2020, 16: e1008399.

van der Lee M, Allard WG, Vossen R, et al. Toward predicting CYP2D6-mediated variable drug response from CYP2D6 gene sequencing data [J]. Sci Transl Med, 2021, 13: eabf3637.

Yang SY, Huang Q, Li LL, et al. An integrated scheme for feature selection and parameter setting in the support vector machine modeling and its application to the prediction of pharmacokinetic properties of drugs [J]. Artif Intell Med, 2009, 46: 155-163.

Steiner HE, Giles JB, Patterson HK, et al. Machine learning for prediction of stable warfarin dose in US latinos and latin Americans [J]. Front Pharmacol, 2021, 12: 749786.

Shin E, Hartman M, Ramanathan M. Performance of the ChatGPT large language model for decision support in community pharmacy [J]. Br J Clin Pharmacol, 2024, 90: 3320-3333.

Williams B, Erstad BL. Analysis of responses from artificial intelligence programs to medication-related questions derived from critical care guidelines [J]. Am J Health Syst Pharm, 2025, 22: zxaf075.

van Nuland M, Lobbezoo AH, van de Garde EMW, et al. Assessing accuracy of ChatGPT in response to questions from day to day pharmaceutical care in hospitals [J]. Explor Res Clin Soc Pharm, 2024, 15: 100464.

Ma HY, Zhou X, Zhang BY, et al. Practical exploration on the rationality of prescriptions based on the DeepSeek large language model [J]. Her Med (医药导报), 2025. https://link.cnki.net/urlid/42.1293.R20250528.1116.002.

Center for Drug Evaluation of NMPA, Chinese Pharmacists Association, Hospital Pharmacy Professional Committee of the Chinese Pharmaceutical Association. Expert consensus on the construction of drug monitoring system in medical institutions [J]. Chin J Drug Appl Monit (中国药物应用与监测), 2022, 19: 135-144.

Crisafulli S, Bate A, Brown JS, et al. Interplay of spontaneous reporting and longitudinal healthcare databases for signal management: position statement from the real-world evidence and big data special interest group of the international society of pharmacovigilance [J]. Drug Saf, 2025. DOI: 10.1007/s40264-025-01548-3.

Shamim MA, Shamim MA, Arora P, et al. Artificial intelligence and big data for pharmacovigilance and patient safety [J]. J Med Surg Public Health, 2024, 3: 100139.

Krumholz HM, Ross JS, Presler AH, et al. What have we learnt from Vioxx? [J]. BMJ, 2007, 334: 120-123.

Alatawi YM, Hansen RA. Empirical estimation of under-reporting in the U. S. Food and Drug Administration Adverse Event Reporting System (FAERS) [J]. Expert Opin Drug Saf, 2017, 16: 761-767.

Dsouza VS, Leyens L, Kurian JR, et al. Artificial intelligence (AI) in pharmacovigilance: a systematic review on predicting adverse drug reactions (ADR) in hospitalized patients [J]. Res Social Adm Pharm, 2025, 21: 453-462.

Dimitsaki S, Natsiavas P, Jaulent MC. Applying AI to structured real-world data for pharmacovigilance purposes: scoping review [J]. J Med Internet Res, 2024, 26: e57824.

Desai MK. Artificial intelligence in pharmacovigilance-opportunities and challenges [J]. Perspect Clin Res, 2024, 15: 116-121.

Ward IR, Wang L, Lu J, et al. Explainable artificial intelligence for pharmacovigilance: what features are important when predicting adverse outcomes? [J]. Comput Methods Programs Biomed, 2021, 212: 106415.

Hestenes M, Stiefel E. Methods of conjugate gradients for solving linear systems [J]. J Res Natl Bur Standards, 1952, 49: 409-435.

Benin AL, Fodeh SJ, Lee K, et al. Electronic approaches to making sense of the text in the adverse event reporting system [J]. J Healthc Risk Manag, 2016, 36: 10-20.

Ménard T, Barmaz Y, Koneswarakantha B, et al. Enabling data-driven clinical quality assurance: predicting adverse event reporting in clinical trials using machine learning [J]. Drug Saf, 2019, 42: 1045-1053.

Chen J, Lalor J, Liu W, et al. Detecting hypoglycemia incidents reported in patients' secure messages: using cost-sensitive learning and oversampling to reduce data imbalance [J]. J Med Internet Res, 2019, 21: e11990.

Xie JK, Ji ZS, Sun K, et al. Practice of active monitoring and reporting management of adverse drug reactions [J]. Eval Anal Drug-Use Hosp China (中国医院用药评价与分析), 2025, 25: 240-243.

Evans HP, Anastasiou A, Edwards A, et al. Automated classification of primary care patient safety incident report content and severity using supervised machine learning (ML) approaches [J]. Health Inform J, 2020, 26: 3123-3139.

Yang J, Wang L, Phadke NA, et al. Development and validation of a deep learning model for detection of allergic reactions using safety event reports across hospitals [J]. JAMA Netw Open, 2020, 3: e2022836.

Chandak P, Tatonetti NP. Using machine learning to identify adverse drug effects posing increased risk to women [J]. Patterns (N Y), 2020, 1: 100108.

Correia Pinheiro L, Durand J, Dogné JM. An application of machine learning in pharmacovigilance: estimating likely patient genotype from phenotypical manifestations of fluoropyrimidine toxicity [J]. Clin Pharmacol Ther, 2020, 107: 944-947.

Zhang L, Yan J, Liu X, et al. Pharmacovigilance practice and risk control of traditional Chinese medicine drugs in China: current status and future perspective [J]. J Ethnopharmacol, 2012, 140: 519-525.

Ryu JY, Kim HU, Lee SY. Deep learning improves prediction of drug–drug and drug–food interactions [J]. Proc Natl Acad Sci U S A, 2018, 115: E4304-E4311.

Hou X, You J, Hu P. Predicting drug-drug interactions using deep neural network[C]//Proceedings of the 2019 11th International Conference on Machine Learning and Computing, 2019. Zhuhai: Association for Computing Machinery, 2019: 168-172.

Deng Y, Xu X, Qiu Y, et al. A multimodal deep learning framework for predicting drug–drug interaction events [J]. Bioinformatics, 2020, 36: 4316-4322.

Huang K, Xiao C, Hoang T, et al. CASTER: predicting drug interactions with chemical substructure representation[C] // Proceedings of the AAAI Conference on Artificial Intelligence, 2019. Hawaii: Association for the Advancement of Artificial Intelligence, 2019: 702-709.

Lin S, Wang Y, Zhang L, et al. MDF-SA-DDI: predicting drug-drug interaction events based on multi-source drug fusion, multi-source feature fusion and transformer self-attention mechanism [J]. Brief Bioinform, 2021, 23: bbab421.

Lyu T, Gao J, Tian L, et al. MDNN: a multimodal deep neural network for predicting drug-drug interaction events [C] // Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, 2021. Montreal: International Joint Conference on Artificial Intelligence, 2022: 3536-3542.

Yue X, Wang Z, Huang J, et al. Graph embedding on biomedical networks: methods, applications and evaluations [J]. Bioinformatics, 2019, 36: 1241-1251.

Feng YH, Zhang SW, Zhang QQ, et al. DeepMDDI: a deep graph convolutional network framework for multi-label prediction of drug-drug interactions [J]. Anal Biochem, 2022, 646: 114631.

International Network for Rational Use of Drugs, Chinese Center Group for Clinical Safe Drug Use, Chinese Pharmacological Society's Committee on Drug-Induced Diseases, et al. Expert consensus on medication error management in China [J]. J Advers Drug React (药物不良反应杂志), 2014, 16: 321-326.

100

Härkänen M, Haatainen K, Vehviläinen-Julkunen K, et al. Artificial intelligence for identifying the prevention of medication incidents causing serious or moderate harm: an analysis using incident reporters' views [J]. Int J Environ Res Public Health, 2021, 18: 9206.

101

Bates DW, Leape LL, Cullen DJ, et al. Effect of computerized physician order entry and a team intervention on prevention of serious medication errors [J]. JAMA, 1998, 280: 1311-1316.

102

Vincent C. Patient safety, 2nd edition [J]. Am J Health Syst Pharm, 2010, 102: 62.

103

Topol EJ. High-performance medicine: the convergence of human and artificial intelligence [J]. Nat Med, 2019, 25: 44-56.

104

Young IJB, Luz S, Lone N. A systematic review of natural language processing for classification tasks in the field of incident reporting and adverse event analysis [J]. Int J Med Inform, 2019, 132: 103971.

105

Graafsma J, Murphy RM, van de Garde EMW, et al. The use of artificial intelligence to optimize medication alerts generated by clinical decision support systems: a scoping review [J]. J Am Med Inform Assoc, 2024, 31: 1411-1422.

106

Corny J, Rajkumar A, Martin O, et al. A machine learning-based clinical decision support system to identify prescriptions with a high risk of medication error [J]. J Am Med Inform Assoc, 2020, 27: 1688-1694.

107

Larios Delgado N, Usuyama N, Hall AK, et al. Fast and accurate medication identification [J]. NPJ Digit Med, 2019, 2: 10.

108

Merry AF, Webster CS, Hannam J, et al. Multimodal system designed to reduce errors in recording and administration of drugs in anaesthesia: prospective randomised clinical evaluation [J]. BMJ, 2011, 343: d5543.

109

Chan J, Nsumba S, Wortsman M, et al. Detecting clinical medication errors with AI enabled wearable cameras [J]. NPJ Digit Med, 2024, 7: 287.

110

Damiani G, Altamura G, Zedda M, et al. Potentiality of algorithms and artificial intelligence adoption to improve medication management in primary care: a systematic review [J]. BMJ Open, 2023, 13: e065301.

111

Zhao M, Hoti K, Wang H, et al. Assessment of medication self-administration using artificial intelligence [J]. Nat Med, 2021, 27: 727-735.

112

Artificial intelligence (AI) in healthcare and research [EB/OL]. London: Nuffield Council on Bioethics, 2018-05-15 [2025-05-17]. https://www.nuffieldbioethics.org/wp-content/uploads/Artificial-Intelligence-AI-in-healthcare-and-research-1.pdf.

113

Clement T, Kemmerzell N, Abdelaal M, et al. XAIR: a systematic metareview of explainable AI (XAI) aligned to the software development process [J]. Mach Learn Knowl Extr, 2023, 5: 78-108.

114

Tiwari R. Explainable AI (XAI) and its applications in building trust and understanding in AI decision making [J]. Int J Sci Res Eng Manag, 2023. DOI: 10.55041/IJSREM17592.

115

AI in the UK: ready, willing and able?[EB/OL]. London: Authority of the House of Lords, 2018 [2025-06-05]. https://publications.parliament.uk/pa/ld201719/ldselect/ldai/100/100.pdf.

116

Goyal MK, Kuppermann N, Cleary SD, et al. Racial disparities in pain management of children with appendicitis in emergency departments [J]. JAMA Pediatrics, 2015, 169: 996-1002.

117

Murdoch B. Privacy and artificial intelligence: challenges for protecting health information in a new era [J]. BMC Med Ethics, 2021, 22: 122.

118

Grzybowski A, Jin K, Wu H. Challenges of artificial intelligence in medicine and dermatology [J]. Clin Dermatol, 2024, 42: 210-215.

119

Aung YYM, Wong DCS, Ting DSW. The promise of artificial intelligence: a review of the opportunities and challenges of artificial intelligence in healthcare [J]. Br Med Bull, 2021, 139: 4-15.

120

Powles J, Hodson H. Google DeepMind and healthcare in an age of algorithms [J]. Health Technol (Berl), 2017, 7: 351-367.

121

Forcier MB, Gallois H, Mullan S, et al. Integrating artificial intelligence into health care through data access: can the GDPR act as a beacon for policymakers? [J]. J Law Biosci, 2019, 6: 317-335.

122

The European commission's artificial intelligence act [EB/OL]. Standford, CA: Standford Univeristy Human-Centered Artificial Intelligence, 2022-03-10. https://hai.stanford.edu/policy/issue-brief-european-commissions-artificial-intelligence-act.

123

The State Council of the People's Republic of China. The regulation on network data security management [EB/OL]. Beijing: The State Council of China, 2024 [2025-05-01]. https://www.gov.cn/zhengce/zhengceku/202409/content_6977767.htm.

124

Rodrigues R. Legal and human rights issues of AI: gaps, challenges and vulnerabilities [J]. J Responsible Technol, 2020, 4: 100005.

125

Murray M, Macedo M, Glynn C. Delivering health intelligence for healthcare services [C]// Proceedings of the 2019 First International Conference on Digital Data Processing (DDP), London, 2019. London: IEEE, 2019: 88-91.

126

Wiens J, Shenoy ES. Machine learning for healthcare: on the verge of a major shift in healthcare epidemiology [J]. Clin Infect Dis, 2018, 66: 149-153.

127

Ali O, Abdelbaki W, Shrestha A, et al. A systematic literature review of artificial intelligence in the healthcare sector: benefits, challenges, methodologies, and functionalities [J]. J Innov Knowl, 2023, 8: 100333.

128

Alowais SA, Alghamdi SS, Alsuhebany N, et al. Revolutionizing healthcare: the role of artificial intelligence in clinical practice [J]. BMC Med Educ, 2023, 23: 689.

129

Carini C, Seyhan AA. Tribulations and future opportunities for artificial intelligence in precision medicine [J]. J Transl Med, 2024, 22: 411.

130

Kelly CM, Osorio-Marin J, Kothari N, et al. Genetic improvement in cotton fiber elongation can impact yarn quality [J]. Ind Crops Prod, 2019, 129: 1-9.

131

Tachkov K, Zemplenyi A, Kamusheva M, et al. Barriers to use artificial intelligence methodologies in health technology assessment in central and east European countries [J]. Front Public Health, 2022, 10: 921226.

2025年第60卷第10期

PDF下载

321

155

引用本文

BibTeX

文章信息

doi: 10.16438/j.0513-4870.2025-0565

接收时间：2025-05-17
首发时间：2026-09-20
出版时间：2025-10-12

补充材料

相关文章

文章信息

作者

出版历史

收稿日期：2025-05-17
修回日期：2025-06-20

基金

作者信息

¹清华大学北京清华长庚医院药学部, 清华大学临床医学院, 北京 102218

²清华大学化学工程系, 北京 100084

通讯作者:

聂广孟, E-mail: ngma00216@btch.edu.cn

张翀, E-mail: chongzhang@tsinghua.edu.cn

参考文献

分享链接

https://castjournals.cast.org.cn/joweb/yxxb/CN/10.16438/j.0513-4870.2025-0565

分享至

全文二维码

扫描看全文

引用本文

BibTeX

本文的引用情况

2种不同金属材料的力学参数

科 Family	属数 Number of genus	种数 Number of species	占总种数比例 Percentage of total species (%)	属 Genus	种数 Number of species	占总种数比例 Percentage of total species (%)
鹅膏菌科Amanitaceae	2	11	5.26	鹅膏菌属 Amanita	10	4.78
小菇科 Mycenaceae	2	12	5.74	丝盖伞属 Inocybe	5	2.39
多孔菌科 Polyporaceae	8	14	6.70	蜡蘑属 Laccaria	5	2.39
红菇科 Russulaceae	3	23	11.00	小皮伞属 Marasmius	6	2.87
				小菇属 Mycena	11	5.26
				光柄菇属 Pluteus	5	2.39
				红菇属 Russula	17	8.13
				栓菌属 Trametes	5	2.39

关闭全屏

BibTeX
EndNote
RefWorks
TxT

ML type	Definition	Common algorithm
Supervised learning	Utilizes labeled datasets (training sets) to predict outputs for unseen data through training	Linear regression; logistic regression; support vector machine (SVM); decision tree; random forest (RF)
Unsupervised learning	Discovers latent patterns or structures using unlabeled datasets	Clustering algorithms: K-means, hierarchical clustering; dimensionality reduction algorithms: principal component analysis (PCA), t-SNE
Reinforcement learning	Learns strategies through interactions with an environment to maximize cumulative rewards	Q-learning; deep Q-network (DQN)

ML type

Definition

Common algorithm

Supervised learning

Utilizes labeled datasets (training sets) to predict outputs for unseen data through training

Linear regression; logistic regression; support vector machine (SVM); decision tree; random forest (RF)

Unsupervised learning

Discovers latent patterns or structures using unlabeled datasets

Clustering algorithms: K-means, hierarchical clustering; dimensionality reduction algorithms: principal component analysis (PCA), t-SNE

Reinforcement learning

Learns strategies through interactions with an environment to maximize cumulative rewards

Q-learning; deep Q-network (DQN)

DL type	Definition	Application scenario
Convolutional neural network (CNN)	A specialized deep learning model designed for processing grid-structured data (e.g., images, audio, video). It achieves efficient feature learning and classification through convolutional operations, pooling operations, and hierarchical feature extraction. Inspired by the local perception mechanisms of biological visual systems, CNNs automatically extract spatially hierarchical features from data (e.g., edges → textures → shapes → semantics)	Image classification; object detection; semantic segmentation; video analysis; natural language processing; medical image analysis
Recurrent neural network (RNN)	A deep learning model specifically designed for sequential data processing. Its core features include recurrent structures and internal memory mechanisms. By introducing parameter sharing and state transfer across the temporal dimension, RNNs capture dynamic dependencies in sequences, making them suitable for tasks requiring contextual modeling (e.g., language, time series, speech)	Natural language processing; time series prediction; speech and audio processing; cross-modal tasks
Generative adversarial network (GAN)	A deep generative model based on a game-theoretic framework, comprising two neural networks: a generator (G) and a discriminator (D). Through adversarial training, GANs achieve data generation and distribution fitting, with the core concept summarized as the "game between a forger and an evaluator"	Image generation and enhancement; medical image analysis; drug discovery; cross-modal generation

DL type

Definition

Application scenario

Convolutional neural network (CNN)

A specialized deep learning model designed for processing grid-structured data (e.g., images, audio, video). It achieves efficient feature learning and classification through convolutional operations, pooling operations, and hierarchical feature extraction. Inspired by the local perception mechanisms of biological visual systems, CNNs automatically extract spatially hierarchical features from data (e.g., edges → textures → shapes → semantics)

Image classification; object detection; semantic segmentation; video analysis; natural language processing; medical image analysis

Recurrent neural network (RNN)

A deep learning model specifically designed for sequential data processing. Its core features include recurrent structures and internal memory mechanisms. By introducing parameter sharing and state transfer across the temporal dimension, RNNs capture dynamic dependencies in sequences, making them suitable for tasks requiring contextual modeling (e.g., language, time series, speech)

Natural language processing; time series prediction; speech and audio processing; cross-modal tasks

Generative adversarial network (GAN)

A deep generative model based on a game-theoretic framework, comprising two neural networks: a generator (G) and a discriminator (D). Through adversarial training, GANs achieve data generation and distribution fitting, with the core concept summarized as the "game between a forger and an evaluator"

Image generation and enhancement; medical image analysis; drug discovery; cross-modal generation

AI technology	Application scenarios in clinical pharmacy
ML, DL	Prediction of drug efficacy and toxicity Prediction of drug-drug interactions Prediction of pharmacokinetics Personalized medication and dose optimization MTM and antimicrobial stewardship
NLP	Pharmacovigilance and ADR monitoring Construction of knowledge graphs and medical record information extraction
LLMs	Prescription review and pharmaceutical care Clinical decision support Medication education Precision pharmacotherapy MTM Drug-drug interaction identification and prediction Pharmacovigilance

AI technology

Application scenarios in clinical pharmacy

ML, DL

Prediction of drug efficacy and toxicity
Prediction of drug-drug interactions
Prediction of pharmacokinetics
Personalized medication and dose optimization
MTM and antimicrobial stewardship

NLP

Pharmacovigilance and ADR monitoring
Construction of knowledge graphs and medical
record information extraction

LLMs

Prescription review and pharmaceutical care
Clinical decision support
Medication education
Precision pharmacotherapy
MTM
Drug-drug interaction identification and prediction
Pharmacovigilance

Strategies of AI in preventing medication error	Related AI technology
1. Analyzing massive amounts of information and identifying patterns to predict medication errors and proactively identify high-risk patients who may experience medication errors and implement preventative measures early, thereby preventing such errors from occurring	ML, DL
2. Identifying and parsing the content recorded in medical records or other healthcare documents, thereby identifying potential medication errors and providing early warnings	NLP
3. Providing real-time decision support to help prescribing physicians analyze medication prescriptions and medical orders, offering alerts regarding drug interactions, dosage, or other precautions, thereby reducing the occurrence of medication errors	ML, DL
4. Analyzing patient medication lists and individual data to identify medication usage patterns and predict potential medication errors. Simultaneously, this method enables the development of personalized medication plans based on patient-specific characteristics, reducing the incidence of medication errors	ML
5. Assisting pharmacists in medication verification and ensuring correct dispensing to the corresponding patients	Computer vision and DL
6. Using wearable devices to monitor patient medication usage and provide warnings for potential medication errors	DL
7. Integrating knowledge across multiple domains, interpreting complex textual information, and generating text, enabling earlier and more comprehensive identification of medication errors	LLMs

Strategies of AI in preventing medication error

Related AI technology

1. Analyzing massive amounts of information and identifying patterns to predict medication errors and proactively identify high-risk patients who may experience medication errors and implement preventative measures early, thereby preventing such errors from occurring

ML, DL

2. Identifying and parsing the content recorded in medical records or other healthcare documents, thereby identifying potential medication errors and providing early warnings

NLP

3. Providing real-time decision support to help prescribing physicians analyze medication prescriptions and medical orders, offering alerts regarding drug interactions, dosage, or other precautions, thereby reducing the occurrence of medication errors

ML, DL

4. Analyzing patient medication lists and individual data to identify medication usage patterns and predict potential medication errors. Simultaneously, this method enables the development of personalized medication plans based on patient-specific characteristics, reducing the incidence of medication errors

5. Assisting pharmacists in medication verification and ensuring correct dispensing to the corresponding patients

Computer
vision and DL

6. Using wearable devices to monitor patient medication usage and provide warnings for potential medication errors

7. Integrating knowledge across multiple domains, interpreting complex textual information, and generating text, enabling earlier and more comprehensive identification of medication errors

LLMs