Добавил:
kiopkiopkiop18@yandex.ru t.me/Prokururor I Вовсе не секретарь, но почту проверяю Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_5525_Библиотеки_им_академика_М_И_Перельмана.pdf
Скачиваний:
0
Добавлен:
31.08.2026
Размер:
29 Мб
Скачать
Artificial Intelligence in Adaptive Radiation Therapy
innovation, thereby contributing signicantly to societal advancements. Including more diverse data from multiple institutions to train AI models also reduces biases and enhances model generalizability. This collaborative approach not only enriches the research landscape but also promotes the overall welfare of society.
Ensuring data security and privacy [28] is of utmost importance when it comes to storing, transmitting, and sharing medical information. Particularly concerning patientssensitive health data, such as names, addresses, and medical records, stringent measures are in place to safeguard this information from any unauthorized access, use, or disclosure. Compliance with regulations such as health insurance portability and accountability act (HIPAA), national laws, and institutional policies is mandatory in this regard. To maintain the condentiality of patient data during transmission, secure methods such as encrypted connections are employed, guaran­teeing protection against interception or tampering. These measures serve to prevent unauthorized access and data breaches of health records. However, the process of data sharing is far from straightforward due to the intricate web of regulatory requirements and the imperative need to safeguard health data privacy. Addressing this challenge, a novel approach, federated learning [29], has gained traction in the eld of radiation oncology. This paradigm eliminates the necessity of data leaving their originating institution. Instead, only the model derived from the data is transferred. While this method partially resolves the issue, it does not always prove as effective as direct data sharing, which continues to be a complex and evolving endeavor in the realm of healthcare data management.
It is important to highlight that integrating certain standards into big data sharing and storing, particularly in the healthcare sector, can be highly advantageous. Standards such as DICOM, Health Level Seven (HL7), and the Fast Healthcare Interoperability Resource (FHIR) not only enhance interoperability but also drive cost efciency, ensure compliance with regulations, and bolster data security within the healthcare industry.
5.2.3 Data visualization
Data visualization [30] is the art of representing complex datasets using visual aids such as charts, graphs, maps, and interactive dashboards. This visual representation serves as a powerful tool, enabling users to explore, comprehend, and communicate intricate data patterns effectively. When dealing with extensive datasets, visual­ization becomes especially crucial as it allows for quick exploration and identica­tion of noteworthy patterns that might warrant further investigation. Additionally, it aids in error detection by highlighting outliers, ensuring data accuracy and reliability. Database tools integrated with visualization features offer users an interactive and exploratory approach to analyzing data. For instance, in the context of radiation oncology, visualizing OAR dose–volume histograms (DVHs) for an entire patient cohort can reveal how treatment plans align with clinical goals. Similarly, employing bar plots with dosimetric statistics grouped by outcome variables can unveil potential correlations. Another key advantage of data visual­ization is its ability to facilitate information sharing. By presenting data visually, it
5-7
Artificial Intelligence in Adaptive Radiation Therapy
becomes easier to convey insights to diverse audiences, fostering a better under­standing of the underlying information.
Several common visualization tools are widely used, including bar plots, histo­grams, line charts, and scatter plots. These tools provide different perspectives on the data, allowing users to gain insights into various aspects of their datasets. Moreover, advanced techniques such as tree maps, principal component analysis (PCA), clustering algorithms, and t-distributed stochastic neighbor embedding (t-SNE) [31] further enhance the visualization process. PCA, for example, reduces data dimensions, simplifying complex datasets and aiding in visualization. Clustering algorithms group similar data points, revealing inherent patterns, while t-SNE helps visualize high-dimensional data in low-dimensional space, preserving data relationships.
In summary, data visualization, coupled with sophisticated algorithms and tools, not only aids in exploring and understanding large datasets but also serves as a valuable tool for users to make informed decisions and communicate ndings effectively.
5.2.4 Knowledge creation and implementation
Knowledge creation involves acquiring insights through rigorous data analysis. This process entails building models, identifying patterns, and devising efcient strategies using data-driven approaches. Extensive validation and testing are essential com­ponents of this process. Once the tools are developed, they must be seamlessly integrated into real clinical settings, ensuring practical implementation and con­tinuous renement.
The implementation of data-driven tools in healthcare presents numerous challenges [32, 33]. First, these tools ideally should be interpretable, offering insights into their decision-making processes to establish trust among healthcare professio­nals. Second, they need to be user-friendly and seamlessly integrated into existing workows or systems. Rigorous validation through clinical trials is also essential to ensure their effectiveness and safety before they can be widely adopted. Furthermore, continuous QA programs are essential to monitor their performance over time. Ethical concerns, such as biases in algorithms and accountability in cases of failure, must be thoroughly investigated. Additionally, the regulatory approval process is required to ensure both efcacy and safety.
5.2.5 Data archiving and deletion
Effectively managing the continuous inux of data is essential, with archiving and deletion playing pivotal roles in this process [34]. Archiving involves preserving data for the long term, typically governed by a well-dened data archiving policy that outlines the frequency and timing of these activities. This strategic approach not only facilitates efcient data management but also ensures alignment with organiza­tional data retention policies and regulatory standards. Archiving data, often transitioned from active databases to cost-effective cold storage, becomes an efcient tool for achieving compliance while mitigating storage costs. Retrieval
5-8
Artificial Intelligence in Adaptive Radiation Therapy
mechanisms are in place, allowing organizations to access archived data when necessary. In sectors such as healthcare, archived data serve as a valuable resource for clinical trials and medical research.
Data deletion is a crucial step in optimizing storage resources and streamlining the management of expansive datasets. By eliminating outdated or inaccurate information, organizations can maintain high data quality, ensuring the accuracy and reliability of analytical processes. Additionally, the removal of unused or outdated data plays a pivotal role in mitigating security risks, reducing the potential surface area for breaches, and minimizing the impact in case of a security incident.

5.3 Big data analytics with AI

5.3.1 Data processing and integration
Data processing is a comprehensive procedure designed to transform raw data into a format suitable for analysis. This intricate process encompasses various essential stages. Data cleaning involves identifying and rectifying outliers, errors, inconsis­tencies, and inaccuracies. Data ltering allows the selection of specic data subsets based on predened inclusion or exclusion criteria. The transformative phase involves several methods. Normalization, or standardization, scales numerical values to a standard range, eliminating biases. Discretization or binning converts continuous data into discrete variables, facilitating categorical analysis. Encoding transforms categorical data into numeric values, ensuring compatibility with analytical models. Data smoothing techniques, such as moving averages, reduce noise, aiding in the identication of underlying trends. Additionally, data imputa­tion could be used to replace missing data with values computed using statistical methods or advanced machine learning techniques, ranging from basic mean or median imputation to more sophisticated regression-based approaches.
In the advanced stages of processing, an analytical component becomes integral. It aims to delve into patterns and relationships among data samples. Clustering, for instance, unveils inherent patterns by grouping similar samples, while regression forecasts relationships between variables, offering valuable insights for data trans­formation. Moreover, feature extraction methods such as PCA reveal essential features contributing to dataset variance. This wealth of information provides insights into processing and integrating the data for downstream modeling tasks.
In specic domains such as radiation oncology, additional processing steps are required for handling various data sources. For text data, NLP techniques such as tokenization, stemming, and sentiment analysis are applied to clinical notes, patient histories, radiology reports, pathology reports, and EHRs. In imaging data, processes such as image enhancement, registration, augmentation, and feature extraction are implemented for the meticulous processing of medical images, such as CT, MRI, and PET. These advanced processing steps not only ensure data integrity and quality but also unlock deeper insights, particularly in complex elds such as RO. Leveraging sophisticated methods becomes pivotal in extracting meaningful information from diverse datasets as technology continues to evolve.
5-9
Artificial Intelligence in Adaptive Radiation Therapy
Data integration serves as a pivotal process, bringing together information from diverse sources to enhance analysis, reporting, and decision-making. In specialized elds such as radiation oncology, there is a growing emphasis on integrating patient data from systems such as EHRs, ROIS, TPS, etc. By amalgamating patients medical histories, lab results, imaging data, clinical information, and treatment plans, a comprehensive patient-specic view emerges, signicantly supporting clinical decision-making, particularly in the context of precision medicine.
Various methods can be employed for effective data integration. Matching and linkage algorithms, alongside merge and join algorithms, play a crucial role in identifying and linking data related to the same entity. Transformation and clustering algorithms, as previously discussed, prove efcient when dealing with data from diverse sources, contributing to a more unied and coherent dataset. Additionally, the integration landscape benets from the application of machine learning algorithms such as decision trees, random forest (RF) [35], k-nearest neighbors (kNN) [36], support vector machine (SVM) [37], and deep learning methods.
These AI techniques not only assist in integrating data but also bring additional advantages. They are adept at reducing dimensionality, extracting latent features, and uncovering intricate patterns within a diverse set of input data. This dimension­ality reduction and feature extraction contribute signicantly to the seamless integration of data from disparate sources, enhancing the overall efcacy of the data integration process. As technology advances, AI techniques further rene the ability to extract meaningful insights from complex datasets, making data integra­tion an ever-evolving and powerful facet of modern data management.
5.3.2 AI modeling
AI modeling stands as a pivotal phase for unraveling inherent patterns in datasets and crafting tools essential for informed clinical decision-making. The standard approach involves constructing AI models and validating their performance on independent datasets. During model development, a subset of data is often ear­marked as a validation dataset. This subset becomes instrumental in rening the training hyperparameters, such as tweaking the strength of penalization terms or optimizing the choice of activation functions based on the models validation performance.
The process of partitioning datasets into distinct sets involves various strategies. Stratied sampling ensures an equitable distribution of target classes across subsets, while time-based splitting arranges data chronologically. k-fold cross-validation (CV) divides data into k folds, iteratively training the model on k – 1 folds and testing on the remaining one. There are variations such as leave-one-out CV and stratied CV. The nested cross-validation method employs an inner loop of CV to pinpoint optimal model parameters and an outer loop to rigorously test the rened model. In essence, the choice of dataset partitioning method should align with the datasets characteristics and the nature (e.g. size) of the AI task at hand. These techniques collectively contribute to the development of models that not only
5-10
Artificial Intelligence in Adaptive Radiation Therapy
Figure 5.4. Overview of four machine learning paradigms and the example algorithms. . Abbreviations: cML
= classical machine learning; DL = deep learning; SVM = support vector machine; RF = random forest; MLP = multi-layer perceptron; CNN = convolutional neural network; FCN = fully convolutional network; LSTM = long short-term memory; GRU = gated recurrent units; PCA = principal component analysis; t-SNE = t-distributed stochastic neighbor embedding; AE= autoencoder; SAE = sparse autoencoders; DAE = denoising autoencoders; VAE = variational autoencoders; GAN = generative adversarial networks. models could be trained using all three learning paradigms (supervised, unsupervised, and semi-supervised).
*
These
unravel intricate associations within data but also generalize effectively to new and unseen instances.
AI models encompass a wide array of categories, each tailored to the specic nature of the datasets they handle. Figure 5.4 shows different paradigms in machine learning along with the corresponding algorithms.
In supervised learning, which deals with labeled datasets, models are employed to address classi cation or regression challenges. Classical machine learning methods such as SVM, RF, k NN, and naive Bayes [38] are commonly utilized for these tasks. Furthermore, advanced deep learning architectures such as multi-layer perceptron (MLP), convolutional neural networks (CNNs) and their variants such as U-Net [39], V-Net [40], and the fully convolutional network (FCN) [41] demonstrate exceptional performance, particularly with tasks involving image inputs. Recurrent neural networks (RNNs) and their variants, including gated recurrent units (GRU) [42] and long short-term memory (LSTM) networks [43], play a pivotal role in analyzing sequential data such as time series data, natural language, DNA sequences, etc. However, RNNs struggle to capture long-term dependencies in sequences due to the vanishing gradient problem. It is also challenging to parallelize them due to the sequential processing manner. In contrast, transformer models [44] can process entire sequences simultaneously thanks to parallel computing and the self-attention mechanism.
Unsupervised learning deals with unlabeled datasets, focusing on tasks such as clustering and dimensionality reduction. Techniques such as PCA, k-means cluster­ing, and t-SNE are commonly employed for these purposes. Additionally,
5-11
Artificial Intelligence in Adaptive Radiation Therapy
sophisticated deep learning methods such as generative adversarial networks (GANs) [45] and autoencoders, including variants such as denoising autoencoders (DAEs) [46], sparse autoencoders (SAEs) [47], and variational autoencoders (VAEs) [48], offer powerful tools for unsupervised learning tasks.
Semi-supervised learning emerges as a crucial eld, particularly in scenarios where labeled data are scarce or expensive to obtain. By leveraging both labeled and unlabeled data, semi-supervised learning approaches effectively learn from labeled examples while also capitalizing on the additional insights provided by unlabeled data. This approach proves particularly valuable when resource-intensive labeling processes pose constraints, enabling a more nuanced understanding of data patterns. Various training strategies, including self-training, low-density separation [49], and graph-based methods [50], contribute to the efcacy of semi-supervised learning algorithms. Notably, certain architectures such as CNNs, RNNs, and transformers exhibit versatility, allowing them to be applied across various learning paradigms, including supervised, semi-supervised, and unsupervised learning.
Reinforcement learning (RL) represents a distinct paradigm in machine learning, where intelligent agents learn to optimize their strategies by interacting with environments to maximize rewards. RL methods, such as dynamic programming, Monte Carlo tree search [51], Q-learning [52], and temporal difference learning, exemplify this approach, offering powerful techniques for solving complex decision­making problems in dynamic environments.
Recently, there have been remarkable advancements in foundation models and generative AI, signaling a pivotal breakthrough in the eld of AI. A foundation model [53] is a large-scale AI model pre-trained on vast unlabeled data across various modalities such as text, images, audio, or video. It learns useful representa­tions and patterns of data and can be adapted to perform a wide range of downstream tasks. Foundation models exhibit exceptional exibility, adapting to supervised, semi-supervised, or unsupervised learning, depending on their training methodology. They can be classied based on the type of input data modalities they are designed to process. For example, large language models (LLMs) are text-based foundation models primarily used for NLP tasks. Examples of LLMs include generative pre-trained transformers (GPTs) [54] and bidirectional encoder repre­sentations from transformers (BERTs) [55]. Some LLMs are specically tailored for medical applications, such as pubMedBERT [56], BioMedLM [57], and clinicalBERT [58]. Image-based models, consisting of self-distillation with no labels (DINO) [59] and masked autoencoders (MAEs) [60], are trained on image datasets and can be used for image recognition, object detection, and segmentation. Multi­modal models are capable of handling multiple data types and are often trained to align different modalities for tasks that benet from cross-modal understanding. Example models include contrastive language-image pretraining (CLIP) [61] and the large language and vision assistant (LLaVA) [62]. Generative AI refers to AI systems capable of generating new content akin to the data they were trained on. By modeling the underlying data distribution, these systems produce novel samples resembling the training data. Generative AI often builds upon foundation models. For instance, an LLM could be initially trained to understand and represent text
5-12
Artificial Intelligence in Adaptive Radiation Therapy
data and be further ne-tuned or extended to perform generative tasks such as human language generation.

5.4 The application of big data in radiation oncology

5.4.1 Medical image segmentation
The applications of big data in medical image segmentation have shown great advancements in recent years [63]. In radiation oncology, contours of targets and OARs are required to generate treatment plans. However, manual segmentation is not only time-consuming but also prone to inter-observer variations. The slow segmentation process also decreases the efciency of the online adaptive workow and potentially affects the efcacy of adaptive plans. Therefore, it is critical to develop fast automatic segmentation methods that can help relieve clinical burden, standardize segmentation performance, and accelerate adaptive workow.
Numerous pairs of images and contours in radiation oncology provide a rich source for training AI-based segmentation models, which have achieved state-of-the­art performance in medical image segmentation. Notably, the U-Net [39], intro­duced by Ronneberger et al in 2015, demonstrated the potential to generate accurate segmentations from input image slices. U-Net consists of an encoder and a decoder connected with skip connections. This encoder–decoder architecture has been incorporated as the backbone of many advanced AI-based medical image segmen­tation models such as AnatomyNet [64] and nn-U-Net [65].
AI-based models have been trained to generate OAR segmentation based on single­or multi-modal images.As a CT scanis routinelyacquiredfor treatment planningin the conventional image-guided radiotherapy workow, numerous models have been trained with pairs of CT images and OAR contours. For example, a WBNet proposed by Chen et al was trained using 505 CT scans of the head and neck, thorax, abdomen, and pelvis, and could accurately delineate 50 OARs [66]. Additionally, commercial AI­based segmentation tools such as Limbus Contour (Limbus AI, Canada) and Contour ProtégéAI (MIM Software Inc, USA) have been integrated into many hospitals worldwide for automatic OAR segmentation on CT scans [67, 68]. Many studies also focused on OAR segmentation based on MRI [67, 69] or ultrasound images [70]dueto growing interest in using other image modalities for treatment guidance. The HaN-Seg dataset [71] contains both CT, MRI, and OAR contours of 56 head and neck cancer patientsand could be trained to train a multi-modalsegmentationmodel. However,CT andMRI scansof the same patientsarenot always available, which makesit challenging to acquire sufcient training data. Valindria et al demonstrated the benets of multi­model learning for multi-organ segmentation based on unpaired CT and MRI scans acquired from different subjects [72].
The combination of big data and AI makes it possible to automatically generate accurate tumor and target contours. Many studies have shown that AI models, trained with pairs of clinical target volume (CTV) contours and CT scans, achieved promising performance in delineating CTV for patients with head and neck cancer [73], breast cancer [74], and cervical cancer [75]. Additionally, tumor segmentation sometimes involves multi-modal images. MRI provides superior soft tissue contrast
5-13
Artificial Intelligence in Adaptive Radiation Therapy
compared with CT and has been increasingly used to delineate tumors in the brain, prostate, and abdomen. For instance, multi-modal MR images and manually drawn tumor contours from 180 glioblastoma patients were used to train a 3D multipath DenseNet that achieved accurate glioblastoma tumor segmentation [76]. Additionally, a commercial AI tool uses both CT and MRI scans to accurately detect and delineate metastatic brain tumors [77, 78]. On the other hand, PET images provide a functional assessment of a tumor and have become the modality of choice for delineating targets in patients with lung cancer, head and neck cancer, and esophageal cancer [77]. Groendahl et al showed that CNNs trained with multi­modal PET-CT images achieved better performance in delineating the gross tumor volume of head and neck images compared to the model trained with PET or CT images [79]. The modality-speci c segmentation network (MoSNet) based on PET­CT images outperforms the state-of-the-art lung tumor segmentation models [80].
MedSAM, a foundation model trained with more than one million pairs of medical images and contours, demonstrates signicant potential to enable universal medical image segmentation [81]. Public datasets for OAR segmentation include multi-organ abdominal CT reference standard segmentations [82], segmentation of thoracic organs-at-risk (SegTHOR) [83], and medical segmentation decathlon [84]. Public datasets for tumor segmentation include brain tumor segmentation (BraTS) [85], liver tumor segmentation (LiTS) [86], and DeepLesion [87]. More public datasets on image–contour pairs could be found in The Cancer Imaging Archive (TCIA) [88] and The Cancer Genome Atlas (TCGA) [89].
5.4.2 Automatic treatment planning
Intensity modulated radiation therapy (IMRT) and volumetric modulated arc therapy (VMAT) enable precise radiation delivery to tumors while sparing OARs. However, the manual treatment planning process normally involves multiple rounds of inverse optimization by trial and error to generate clinically acceptable plans. This manual workow is very time-consuming, requires signicant human expertise, and may not always yield optimal treatment plans. Online adaptive radiation therapy, which allows better inter-fraction organ motion management compared with non-adaptive treatment, has drawn signicant clinical interest. However, this requires decreasing planning time to a few minutes.
Automatictreatmentplanningcould be achievedby trainingAI modelsto predict3D dose distribution,predict uence maps,or automatethe hyperparameter tuningprocess during inverse optimization. U-Net and its variants could be trained to predict clinical dose distributions based on the input of CT and/or contours of targets and OARs. Promising results have been achieved for cancer sites including the head and neck [90], breast [91], and prostate [92]. The predicted dose distributions or the corresponding DVHs could be fed into the dose optimization engine to generate deliverable plans. However, as historical plans used for model training may not always be optimal, plans generated using this method may be suboptimal. Instead of clinical dose distribution, these models can also be trained to predict Pareto optimal dose distributions based on
5-14
Artificial Intelligence in Adaptive Radiation Therapy
the additional input of user-specified contourweights [93]orDVHs[94].This allowsthe user to search for the optimal trade-off dose in real-time.
Fluence map prediction models could help skip the optimization step and potentially speed up the automatic treatment planning process. These models can be trained to take the AI-predicted dose distributions as their input and predict uence maps [95], predict uence maps directly based on contours [96], or simultaneously predict dose distribution and uence maps based on CT and contours [97]. On the other hand, RL agents could be trained to observe intermediate plan DVHs and take action to adjust planning parameters for generating the plan with the maximized quality score [98].
5.4.3 Treatment response prediction
Precision oncology represents a paradigm shift in cancer treatment, leveraging the power of big data to tailor therapy to individual patients. This approach is grounded in the understanding that cancer is not a singular disease but rather a highly heterogeneous disease with unique genomic and phenotypic characteristics that vary among individual patients [99]. AI models, trained on extensive datasets such as patient demographics, genetic test results, medical images, and treatment plan dose distribution, can predict how individual patients will respond to specic treatment regimens. This could potentially allow us to improve patient outcomes and reduce the toxicity and burden of unnecessary treatments.
Many efforts have been made to utilize large-scale omics data to predict radiation therapy response. For example, genomic classiers have been developed to stratify local recurrence risk after radiation therapy in breast cancer patients and identify those who would benet from radiation therapy [100102]. Radiomic features extracted from medical images have been found to correlate with genomic biomarkers and achieve promising performance in predicting treatment responses for patients with lung cancer, rectal cancer, etc [103]. Moreover, the integration of multi-omics data holds the potential to further enhance model prediction accuracy [104106].
5.4.4 Quality assurance and patient safety
Ensuring QA and patient safety is a top priority in radiation oncology. The data relevant to this domain come from a variety of sources, including QA procedures and incident learning systems (ILSs) [107, 108]. Data generated from QA procedures include machine-specic data and patient-specic data. Machine-specic data can be from various categories, including dosimetric, mechanical, imaging, and respiratory gating. These data are normally obtained by measurements or checks during machine and TPS acceptance/commissioning/upgrading and routine daily/monthly/annual QA. Patient-specic data may include patient-specic QA conducted using either an array detector or electronic portal imaging device [109], treatment plan review [110], and machine delivery log les [119]. The ILS within radiation oncology [111] serves as a pivotal platform for healthcare professionals to both report and glean insights from incidents and near-misses tied to cancer treatment. Through meticulous data collection and analysis, ILS fosters a continuous drive toward enhancing quality
5-15
Artificial Intelligence in Adaptive Radiation Therapy
within radiation oncology practices. It empowers healthcare providers to discern patterns, pinpoint root causes, and highlight areas ripe for improvement in patient safety and treatment delivery. Radiation Oncology ILS (RO-ILS) [112], jointly sponsored by ASTRO and AAPM, spearheads the establishment of a national database tasked with housing incidents, near-misses, and unsafe conditions reported across various institutions. Advocating for a collaborative approach, RO-ILS encourages sharing and learning from events submitted nationwide. This initiative promotes a culture of safety and collaboration within the radiation oncology community, ultimately culminating in elevated patient care outcomes.
With a wealth of data stemming from both QA and ILS, AI emerges as a powerful ally in several key areas. It can adeptly discern trends in QA task performance, identifying outliers that warrant closer examination. Furthermore, AI streamlines QA workows by automating tasks such as plan review and imaging QA, thereby enhancing both efciency and precision. By leveraging predictive analytics, AI forecasts which aspects are more prone to failure, allowing for proactive intervention and resource allocation [113].
5.4.5 Clinical decision support
Big data and AI tools hold signicant promise in aiding healthcare professionals in decision-making processes by furnishing them with relevant, timely, and evidence­based information [114]. Clinical decision support systems, underpinned by big data and AI, can amalgamate patient-specic data and medical knowledge to offer recommendations for diagnosis, treatment, and patient care. In the realm of medicine, there is a burgeoning interest in leveraging foundation models across various data sources, including EHRs, clinical guidelines, medical literature, medical imaging, biological sequences, and molecular proles [115, 116].
EHRs encompass both structured datasuch as billing codes, demographics, and medicationsand unstructured data, including medical notes, radiology reports, and lab reports. Challenges include irregularities such as ambiguous jargon and non­standard phrasal structure, necessitating domain expertise. Some example datasets include the deidentied clinical acronym sense inventory (CASI) [117] which contains snippets of clinical notes across specialties in four University of Minnesota-affiliated hospitals. The MIMIC-III criticalcare databasecontainsapproximately 2 million notes written between 2001 and 2012 in the ICU of Beth Israel Deaconess Medical Center [118, 119]. Foundation models have demonstrated capabilities such as extracting drug names from medical reports [120], responding to patient queries [121], summarizing clinical text [122], and predicting clinical needs based on clinical notes [123].
The wealth of information found in the scientic literature, clinical guidelines, and knowledge bases such as NCT guidelines and PubMed identiers is invaluable for clinical decision-making. However, manually extracting predictive information from these sources can be time-consuming and requires extensive expertise. LLMs offer signicant promise in this regard. For instance, a few-shot prediction model can leverage LLM representations, which encapsulate prior knowledge in scientic literature, to predict drug pair synergy in rare tissues with limited data [124].
5-16