Predicting the length of stay (LoS) is important for hospital administration, as it helps allocate proper resources, such as bed management and hospital staffing. Patients' Electronic Health Records (EHRs) contain highly relevant data for LoS prediction; however, their integration and effective use in predictive modeling for accurately estimating LoS remain challenging. To address this, we propose a homogeneous Graph Neural Network (GNN)-based framework for predicting LoS. This method employs a comprehensive data fusion strategy based on the hospital Visit-based Similarity Graph (VSG), which integrates diverse multi-modal clinical features into a coherent, homogeneous graph representation. Next, this VSG is fed into the GNN layers, followed by Multi-Layer Perceptron (MLP) layers, which further transform the aggregated node representation from message passing to enhance the performance of LoS classification. We also systematically evaluate multiple machine learning, GNN, and graph transformer models with various feature configurations. Results obtained by applying our models to the MIMIC-III database demonstrate (i) a GraphSAGE encoder followed by an artificial neural network classifier (SAGE_ANN) and (ii) a two-layer GraphSAGE model (SAGE_2L) consistently yield superior performance in global hospital LoS prediction. This study highlights the impact of multi-modal fusion and graph learning in advancing predictive modeling for complex clinical systems.

PURPOSE:Despite the frequency with which patients with cancer receive radiotherapy, integrating radiation oncology data with other aspects of the clinical record remains challenging because of siloed and variable software systems, high data complexity, and inconsistent data encoding. Recognizing these challenges, the Veterans Affairs (VA) National Radiation Oncology Program (NROP) is developing Granular Radiotherapy Information Database (GRID), a platform and pipeline to combine radiotherapy data across the VA with the goal of both better understanding treatment patterns and outcomes and enhancing research and data analysis capabilities. METHODS:This study represents a proof-of-principle retrospective cohort analysis and review of select radiation treatment data from the VA Radiation Oncology Quality Surveillance Program (VAROQS) initiative. Key radiation oncology data elements were extracted from Digital Imaging and Communications in Medicine Radiotherapy extension (DICOM-RT) files and combined into a single database using custom scripts. These data were transferred to the VA's Corporate Data Warehouse (CDW) for integration and comparison with the VA Cancer Registry System and tumor sequencing data. RESULTS:The final cohort includes 1,568 patients, 766 of whom have corresponding DICOM-RT data. All cases were successfully linked to the CDW; 18.8% of VAROQS cases were not reported in the existing VA cancer registry. The VAROQS data contributed accurate radiation treatment details that were often erroneous or missing from the cancer registry record. Tumor sequencing data were available for approximately 5% of VAROQS cases. Finally, we describe a clinical dosimetric analysis leveraging GRID. CONCLUSION:NROP's GRID initiative aims to integrate VA radiotherapy data with other clinical data sets. It is anticipated to generate the single largest collection of radiation oncology-centric data merged with detailed clinical and genomic data, primed for large-scale quality assurance, research reuse, and discovery science.

Accurate prediction of hospital length of stay (LoS) is a vital component in optimizing clinical workflows, resource allocation, and patient care. This study presents a comprehensive evaluation of machine learning models for both binary and multi-class LoS classification tasks using structured clinical variables, physiological measurements, and unstructured clinical notes. Seven data configurations were constructed from combinations of structured features (Z), including diagnoses, procedures, medications, laboratory tests, and microbiology results; MeSH-based symptoms (S); physiological signals (F); and textual representations (E): Z, F, E, ZS, ZSF, ZSE, and ZSEF. Five predictive models-Artificial Neural Networks (ANN), XGBoost, Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM)-were applied, with and without feature selection, where categorical features and Bag-of-Words representations were reduced to varied dimensions. Results indicate that the base structured feature set (Z) alone yields strong predictive performance across tasks. Moreover, the integration of additional data types-S, F, and E-either individually or in combination, consistently enhanced performance, with the ZSEF configuration achieving the highest F1-scores and AUC values in most cases. While the application of SMOTE did not yield substantial improvements in the global setting encompassing all hospital admissions, it demonstrated enhanced performance in disease-specific cohorts, particularly for patients admitted with lung cancer. Among the evaluated models, XGBoost and ANN demonstrated superior generalizability. These findings underscore the effectiveness of multimodal data integration and feature reduction techniques in advancing predictive modeling for hospital length of stay across diverse patient populations.

In the era of big data, it is necessary to provide novel and efficient platforms for training machine learning models over large volumes of data. The MapReduce approach and its Apache Spark implementation are among the most popular methods that provide high-performance computing for classification algorithms. However, they require dedicated implementations that will take advantage of such architectures. Additionally, many real-world big data problems are plagued by class imbalance, posing challenges to the classifier training step. Existing solutions for alleviating skewed distributions do not work well in the MapReduce environment. In this paper, we propose a novel KD-tree based classifier, together with a variation of the SMOTE algorithm dedicated to the Spark platform. Our algorithms offer excellent predictive power and can work simultaneously with binary and multi-class imbalanced data. Exhaustive experiments conducted using the Amazon Web Service platform showcase the high efficiency and flexibility of our proposed algorithms.

Purpose Radiation Oncology Learning Health System (RO-LHS) is a promising approach to improve the quality of care by integrating clinical, dosimetry, treatment delivery, research data in real-time. This paper describes a novel set of tools to support the development of a RO-LHS and the current challenges they can address. Methods We present a knowledge graph-based approach to map radiotherapy data from clinical databases to an ontology-based data repository using FAIR concepts. This strategy ensures that the data is easily discoverable, accessible, and can be used by other clinical decision support systems. It allows for visualization, presentation, and data analyses of valuable information to identify trends and patterns in patient outcomes. We designed a search engine that utilizes ontology-based keyword searching, synonym-based term matching that leverages the hierarchical nature of ontologies to retrieve patient records based on parent and children classes, connects to the Bioportal database for relevant clinical attributes retrieval. To identify similar patients, a method involving text corpus creation and vector embedding models (Word2Vec, Doc2Vec, GloVe, and FastText) are employed, using cosine similarity and distance metrics. Results The data pipeline and tool were tested with 1660 patient clinical and dosimetry records resulting in 504,180 RDF tuples and visualized data relationships using graph-based representations. Patient similarity analysis using embedding models showed that the Word2Vec model had the highest mean cosine similarity, while the GloVe model exhibited more compact embeddings with lower Euclidean and Manhattan distances. Conclusions The framework and tools described support the development of a RO-LHS. By integrating diverse data sources and facilitating data discovery and analysis, they contribute to continuous learning and improvement in patient care. The tools enhance the quality of care by enabling the identification of cohorts, clinical decision support, and the development of clinical studies and machine learning programs in radiation oncology.

The popularity of learning from data with machine learning and neural networks has lead to the creation of many new datasets for almost every problem domain. However, even within a single domain, these datasets are often collected with disparate features, sampled from different sub-populations, and recorded at different time points. Even with the plethora of individual datasets, large data science projects can be difficult as it is often not trivial to merge these smaller datasets. Inherent challenges in some domains such as medicine also makes it very difficult to create large single source datasets or multi-source datasets with identical features. Instead of trying to merge these non-matching datasets directly, we propose a neural network architecture that can provide data augmentation using features common between these datasets. Our results show that this style of data augmentation can work for both image and tabular data.

In this study, we investigated 3D convolutional neural networks (CNNs) with input from radiographic and dosimetric datasets of primary lung tumors and surrounding lung volumes to predict the likelihood of radiation pneumonitis (RP). Pre-treatment, 3- and 6-month follow-up computed tomography (CT) and 3D dose datasets from one hundred and ninety-three NSCLC patients treated with stereotactic body radiotherapy (SBRT) were retrospectively collected and analyzed for this study. DenseNet-121 and ResNet-50 models were selected for this study as they are deep neural networks and have been proven to have high accuracy for complex image classification tasks. Both were modified with 3D convolution and max pooling layers to accept 3D datasets. We used a minority class oversampling approach and data augmentation to address the challenges of data imbalance and data scarcity. We built two sets of models for classification of three (No RP, Grade 1 RP, Grade 2 RP) and two (No RP, Yes RP) classes as outputs. The 3D DenseNet-121 models performed better (F1 score [0.81], AUC [0.91] [three class]; F1 score [0.77], AUC [0.84] [two class]) than the 3D ResNet-50 models (F1 score [0.54], AUC [0.72] [three-class]; F1 score [0.68], AUC [0.71] [two-class]) (p = 0.017 for three class predictions). We also attempted to identify salient regions within the input 3D image dataset via integrated gradient (IG) techniques to assess the relevance of the tumor surrounding volume for RP stratification. These techniques appeared to indicate the significance of the tumor and surrounding regions in the prediction of RP. Overall, 3D CNNs performed well to predict clinical RP in our cohort based on the provided image sets and radiotherapy dose information.

Physicians often label anatomical structure sets in Digital Imaging and Communications in Medicine (DICOM) images with nonstandard random names. Hence, the standardization of these names for the Organs at Risk (OARs), Planning Target Volumes (PTVs), and ‘Other’ organs is a vital problem. This paper presents novel deep learning methods on structure sets by integrating multimodal data compiled from the radiotherapy centers of the US Veterans Health Administration (VHA) and Virginia Commonwealth University (VCU). These de-identified data comprise 16,290 prostate structures. Our method integrates the multimodal textual and imaging data with Convolutional Neural Network (CNN)-based deep learning approaches such as CNN, Visual Geometry Group (VGG) network, and Residual Network (ResNet) and shows improved results in prostate radiotherapy structure name standardization. Evaluation with macro-averaged F1 score shows that our model with single-modal textual data usually performs better than previous studies. The models perform well on textual data alone, while the addition of imaging data shows that deep neural networks achieve better performance using information present in other modalities. Additionally, using masked images and masked doses along with text leads to an overall performance improvement with the CNN-based architectures than using all the modalities together. Undersampling the majority class leads to further performance enhancement. The VGG network on the masked image-dose data combined with CNNs on the text data performs the best and presents the state-of-the-art in this domain.

Multimodal classification research has been gaining popularity with new datasets in domains such as satellite imagery, biometrics, and medicine. Prior research has shown the benefits of combining data from multiple sources compared to traditional unimodal data that has led to the development of many novel multimodal architectures. However, the lack of consistent terminologies and architectural descriptions makes it difficult to compare different solutions. We address these challenges by proposing a new taxonomy for describing multimodal classification models based on trends found in recent publications. Examples of how this taxonomy could be applied to existing models are presented as well as a checklist to aid in the clear and complete presentation of future models. Many of the most difficult aspects of unimodal classification have not yet been fully addressed for multimodal datasets, including big data, class imbalance, and instance-level difficulty. We also provide a discussion of these challenges and future directions of research.

Abstract Rigorous radiotherapy quality surveillance and comprehensive outcome assessment require electronic capture and automatic abstraction of clinical, radiation treatment planning, and delivery data. We present the design and implementation framework of an integrated data abstraction, aggregation, and storage, curation, and analytics software: the Health Information Gateway and Exchange (HINGE), which collates data for cancer patients receiving radiotherapy. The HINGE software abstracts structured DICOM‐RT data from the treatment planning system (TPS), treatment data from the treatment management system (TMS), and clinical data from the electronic health records (EHRs). HINGE software has disease site‐specific “Smart” templates that facilitate the entry of relevant clinical information by physicians and clinical staff in a discrete manner as part of the routine clinical documentation. Radiotherapy data abstracted from these disparate sources and the smart templates are processed for quality and outcome assessment. The predictive data analyses are done on using well‐defined clinical and dosimetry quality measures defined by disease site experts in radiation oncology. HINGE application software connects seamlessly to the local IT/medical infrastructure via interfaces and cloud services and performs data extraction and aggregation functions without human intervention. It provides tools to assess variations in radiation oncology practices and outcomes and determines gaps in radiotherapy quality delivered by each provider.

Learning from imbalanced data is among the most challenging areas in contemporary machine learning. This becomes even more difficult when considered the context of big data that calls for dedicated architectures capable of high-performance processing. Apache Spark is a highly efficient and popular architecture, but it poses specific challenges for algorithms to be implemented for it. While oversampling algorithms are an effective way for handling class imbalance, they have not been designed for distributed environments. In this paper, we propose a holistic look on oversampling algorithms for imbalanced big data. We discuss the taxonomy of oversampling algorithms and their mechanisms used to handle skewed class distributions. We introduce a Spark library with 14 state-of-the-art oversampling algorithms implemented and evaluate their efficacy via extensive experimental study. Using binary and multi-class massive data sets, we analyze the effectiveness of oversampling algorithms and their relationships with different types of classifiers. We evaluate the trade-off between accuracy and time complexity of oversampling algorithms, as well as their scalability when increasing the size of data. This allows us to gain insight into the usefulness of specific components of oversampling algorithms for big data, as well as formulate guidelines and recommendations for designing future resampling approaches for massive imbalanced data. Our library can be downloaded from https://github.com/fsleeman/spark-class-balancing.git.

Radiotherapy treatment for cancer patients involves a complex workflow involving radiation physicists, therapists, dosimetrists, physicians and nurses. Multiple hand-offs between these care team members often lead to errors varying in severity levels. Such errors are logged in incident reports stored in the Radiation Oncology Incident Learning System. Here, we present an automated incident triage and severity determination pipeline that can predict high and low severity incidents. Incident reports are collected from the US Veterans Health Affairs (VHA) and Virginia Commonwealth University (VCU) radiation oncology centers. Natural language processing (NLP) and deep learning (DL) methods, like CNN and BiLSTM, are used to predict severity using the 'Incident Description' information. Other features like 'Incident Type', 'Action taken by reporter' and 'Incident discovered at' are used to infer the best performing model. Random oversampling and minority class oversampling are employed to address large class imbalance ratios in the data. We observed that CNN performs best on both VHA data (0.83 F1-score) and the combined VCU+VHA data (0.83 F1-score) while CNN with minority sampling performs better on VCU data (0.60 F1-score) using the 'Incident Description' feature. Different feature combinations suggest that the two feature model using 'Incident Description' and 'Action taken by reporter' performs better with CNN on both the VHA (0.84 F1-score) and combined VCU+VHA data (0.81 F1-score). Multiple features were considered for the first time where the two feature model using CNNs emerge as the best suited for automating the radiotherapy incident triage and prioritization process.

Significant growth in Electronic Health Records (EHR) over the last decade has provided an abundance of clinical text that is mostly unstructured and untapped. This huge amount of clinical text data has motivated the development of new information extraction and text mining techniques. Named Entity Recognition (NER) and Relationship Extraction (RE) are key components of information extraction tasks in the clinical domain. In this paper, we highlight the present status of clinical NER and RE techniques in detail by discussing the existing proposed NLP models for the two tasks and their performances and discuss the current challenges. Our comprehensive survey on clinical NER and RE encompass current challenges, state-of-the-art practices, and future directions in information extraction from clinical text. This is the first attempt to discuss both of these interrelated topics together in the clinical context. We identified many research articles published based on different approaches and looked at applications of these tasks. We also discuss the evaluation metrics that are used in the literature to measure the effectiveness of the two these NLP methods and future research directions.

Standardization of radiotherapy structure names is essential for developing data-driven personalized radiotherapy treatment plans. Different types of data are associated with radiotherapy structures, such as the physician-given text labels, geometric (image) data, and Dose-Volume Histograms (DVH). Prior work on structure name standardization used just one type of data. We present novel approaches to integrate complementary types (views) of structure data to build better-performing machine learning models. We present two methods, namely (a) intermediate integration and (b) late integration, to combine physician-given textual structure name features and geometric information of structures. The dataset consisted of 709 prostate cancer and 752 lung cancer patients across 40 radiotherapy centers administered by the U.S. Veterans Health Administration (VA) and the Department of Radiation Oncology, Virginia Commonwealth University (VCU). We used randomly selected data from 30 centers for training and ten centers for testing. We also used the VCU data for testing. We observed that the intermediate integration approach outperformed the models with a single view of the dataset, while late integration showed comparable performance with single-view results. Thus, we demonstrate that combining different views (types of data) helps build better models for structure name standardization to enable big data analytics in radiation oncology.

Purpose: To present a Machine Learning pipeline for automatically relabeling anatomical structure sets in the Digital Imaging and Communications in Medicine (DICOM) format to a standard nomenclature that will enable data abstraction for research and quality improvement. Methods: DICOM structure sets from approximately 1200 lung and prostate cancer patients across 40 treatment centers were used to build predictive models to automate the relabeling of clinically specified structure labels to standardized labels as defined by the American Association of Physics in Medicine's (AAPM) Task Group 263 (TG-263). Volumetric bitmaps were created based on the delineated volumes and were combined with associated bony anatomy data to build feature vectors. Feature reduction was performed with singular value decomposition and the resulting vectors were used for predicting the label of each structure using five different classifier algorithms on the Apache Spark platform with 5-fold cross-validation. Undersampling methods were used to deal with underlying class imbalance that hindered the performance of classifiers. Experiments were performed on both a curated version of the data, which included only annotated structures, and the non-curated data that included all structures from the original treatment plans. Results: Random Forest provided the highest accuracies with F, scores of 98.77 for lung and 95.06 for prostate on the curated data sets. Scores were lower with 95.67 for lung and 90.22 for prostate on the non-curated data sets, highlighting some of the challenges of classifying real clinical data. Including bony anatomy data and pooling information from all structures for the same patient both increased accuracies. In some cases, undersampling with k-Means clustering for class balancing improved classifier accuracy but in all experiments it significantly reduced run time compared to random undersampling. Conclusion: This work shows that structure sets can be relabeled using our approach with accuracies over 95% for many structure types when presented with curated data. Although accuracies dropped when using the full non-curated data sets, some structure types were still correctly labeled over 90% of the time. With similar results obtained on an external test data set, we can infer that the proposed models are likely to work on other clinical data sets.

Clinical factors, including T-stage, Gleason score, and baseline prostate-specific antigen, are used to stratify patients with prostate cancer (PCa) into risk groups. This provides prognostic information for a heterogeneous disease such as PCa and guides treatment selection. In this article, we hypothesize that nonclinical factors may also impact treatment selection and their adherence to treatment guidelines. A total of 552 patients with intermediate- and high-risk PCa treated with definitive radiation with or without androgen deprivation therapy (ADT) between 2010 and 2017 were identified from 34 medical centers within the Veterans Health Administration. Medical charts were manually reviewed, and details regarding each patient's clinical history and treatment were extracted. Support Vector Machine and Random forest-based classification was used to identify clinical and nonclinical predictors of adherence to the treatment guidelines from the National Comprehensive Cancer Network (NCCN). We created models for predicting both initial treatment intent and treatment alterations. Our results demonstrate that besides clinical factors, the center in which the patient was treated (nonclinical factor) played a significant role in adherence to NCCN guidelines. Furthermore, the treatment center served as an important predictor to decide on whether or not to prescribe ADT; however, it was not associated with ADT duration and weakly associated with treatment alterations. Such center-bias motivates further investigation on details of center-specific barriers to both NCCN guideline adherence and on oncological outcomes. In addition, we demonstrate that publicly available data sets, for example, that from Surveillance, Epidemiology, and End Results (SEERs), may not be well equipped to build such predictive models on treatment plans.

The Radiotherapy Incident Reporting and Analysis System (RIRAS) receives incident reports from Radiation Oncology facilities across the US Veterans Health Affairs (VHA) enterprise and Virginia Commonwealth University (VCU). In this work, we propose a computational pipeline for analysis of radiation oncology incident reports. Our pipeline uses machine learning (ML) and natural language processing (NLP) based methods to predict the severity of the incidents reported in the RIRAS platform using the textual description of the reported incidents. These incidents in RIRAS are reviewed by a radiation oncology subject matter expert (SME), who initially triages some incidents based on the salient elements in the incident report. To automate the triage process, we used the data from the VHA treatment centers and the VCU radiation oncology department. We used NLP combined with traditional ML algorithms, including support vector machine (SVM) with linear kernel, and compared it against the transfer learning approach with the universal language model fine-tuning (ULMFiT) algorithm. In RIRAS, severities are divided into four categories; A, B, C, and D, with A being the most severe to D being the least. In this work, we built models to predict High (A & B) vs. Low (C & D) severity instead of all the four categories. Models were evaluated with macro-averaged precision, recall, and F1-Score. The Traditional ML machine learning (SVM-linear) approach did well on the VHA dataset with 0.78 F1-Score but performed poorly on the VCU dataset with 0.5 F1-Score. The transfer learning approach did well on both datasets with 0.81 F1-Score on VHA dataset and 0.68 F1-Score on the VCU dataset. Overall, our methods show promise in automating the triage and severity determination process from radiotherapy incident reports.

The lack of standardized structure names in radiotherapy (RT) data limits interoperability, data sharing, and the ability to perform big data analysis. To standardize radiotherapy structure names, we developed an integrated natural language processing (NLP) and machine learning (ML) based system that can map the physician-given structure names to American Association of Physicists in Medicine (AAPM) Task Group 263 (TG-263) standard names. The dataset consist of 794 prostate and 754 lung cancer patients across the 40 different radiation therapy centers managed by the Veterans Health Administration (VA). Additionally, data from the Radiation Oncology department at Virginia Commonwealth University (VCU) was collected to serve as a test set. Domain experts identified as anatomically significant nine prostate and ten lung organs-at-risk (OAR) structures and manually labeled them according to the TG-263 standards, and remaining structures were labeled as Non_OAR. We experimented with six different classification algorithms and three feature vector methods, and the final model was built with fastText algorithm. Multiple validation techniques are used to assess the robustness of the proposed methodology. The macro-averaged F 1 score was used as the main evaluation metric. The model achieved an F 1 score of 0.97 on prostate structures and 0.99 for lung structures from the VA dataset. The model also performed well on the test (VCU) dataset, achieving an F 1 score of 0.93 for prostate structures and 0.95 on lung structures. In this work, we demonstrate that NLP and ML based approaches can used to standardize the physician-given RT structure names with high fidelity. This standardization can help with big data analytics in the radiation therapy domain using population-derived datasets, including standardization of the treatment planning process, clinical decision support systems, treatment quality improvement programs, and hypothesis-driven clinical research.