Shen, Feichen

Precision oncology is expected to improve selection of targeted therapies, tailored to individual patients and ultimately improve cancer patients’ outcomes. Several cancer genetics knowledge databases have been successfully developed for such purposes, including CIViC and OncoKB, with active community-based curations and scoring of genetic-treatment evidences. Although many studies were conducted based on each knowledge base respectively, the integrative analysis across both knowledge bases remains largely unexplored. Thus, there exists an urgent need for a heterogeneous precision oncology knowledge resource with computational power to support drug repurposing discovery in a timely manner, especially for life-threatening cancer. In this pilot study, we built a heterogeneous precision oncology knowledge resource (POKR) by integrating CIViC and OncoKB, in order to incorporate unique information contained in each knowledge base and make associations amongst biomedical entities (e.g., gene, drug, disease) computable and measurable via training POKR graph embeddings. All the relevant codes, database dump files, and pre-trained POKR embeddings can be accessed through the following URL: https://github.com/shenfc/POKR.

Abstract Background: DNIV accounts for 6%–10% of newly diagnosed breast cancer cases. Despite widespread mammography screening, its incidence is increasing in the United States and survival of this disease has only modestly improved since the late 1970s. As patient data accumulates in EHR, it’s promising to generate practice-based evidence through utilization of observational data sources. However, assembly of a DNIV cohort based on EHR data is challenging, as key pathologic and staging information are stored in unstructured clinical narratives and not available as structured data. In this study, we developed a rule-based algorithm to phenotype DNIV using natural language processing (NLP) techniques, and implemented the algorithm on our institutional EHR to extract potential DNIV cases. Methods and Results: We defined DNIV as those with either (1) M1 disease identified at time of initial presentation or M1 disease identified within 4 months after definitive surgery. We first developed a reference case list of DNIV verified by physician chart review of the EHR. We next refined the algorithm on a training dataset containing 51 positive and 38 negative reference cases. Next we tested the performance on the testing data containing 23 positive and 55 negative cases. The phenotyping algorithm identified key data elements using NLP, i.e., stage IV breast cancer, definitive surgery, stage 0-III, recurrent breast cancer and associated dates. To identify DNIV cases, phenotyping algorithm integrated temporal relations among the key data elements. The following steps were conducted in the following sequential order: (1) Identification of patients with breast cancer diagnosis using ICD-9 and ICD-10 codes. (2) Patients are positive cases if there are explicit mentions that delineate DNIV from recurrent metastatic breast cancer, such as “de novo stage IV” or “primary intact” detected by NLP at time of diagnosis. (3) Otherwise, patients with definitive surgery are selected if stage IV was within 5 years before definitive surgery or within 4 months after definitive surgery, along with patients with stage IV but without definitive surgery. (4) We further excluded patients with stage 0-III detected within 5 years after stage IV. (5) We further excluded patients with recurrent breast cancer detected before or at time of detection of stage IV. The remaining patients were left as positive cases. Precision of the algorithm was 70%, recall was 87% and F1, the weighted average of precision and recall, was 77%. We implemented our algorithm to interrogate 10 million clinical documents in a cohort of 56,548 patients with breast cancer diagnosis codes who presented to our institution between 2004 and 2018 and had research authorization. We identified 1918 potential DNIV cases. Conclusion: Our future focus is on algorithm refinement. An algorithm-generated cohort could serve as a data source for further study on outcomes related to DNIV and ideally as automated data abstractor and staging process. Citation Format: Liwei Wang, Karthik Giridhar, Kimberly Corbin, Brenda Ernst, Sadia Choudhery, Emmanuel Gabriel, Feichen Shen, Hongfang Liu, James Jakub. Identifying de novo stage IV breast cancer (DNIV) cases in Electronic Health Records (EHR) using natural language processing [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-050.

Abstract Background: Next-generation sequencing provides comprehensive information about individuals’ genetic makeup and is commonplace in precision oncology practice. Due to the heterogeneity of individual patient’s disease conditions and treatment journeys, not all targeted therapies were initiated despite actionable mutations. To better understand and support the clinical decision-making process in precision oncology, there is a need to examine real-world associations of patients’ genetic information and treatment choice.Methods: To fill the gap of insufficient use of real-world data (RWD) in electronic health records (EHRs), we generated a single Resource Description Framework (RDF) resource, called PO2RDF (precision oncology to RDF) by integrating information regarding gene, variant, disease, and drug from genetic reports and EHRs. Results: There are total 2,309,014 triples contained in the PO2RDF. Among them 32,815 triples are related to Gene, 34,695 triples are related to Variant, 8,787 triples are related to Disease, 26,154 triples are related to Drug. We performed one use case analysis to demonstrate the usability of the PO2RDF: we examined real-world associations between EGFR mutations and targeted therapies to confirm existing knowledge and detect off-label use. Conclusions: In conclusion, our work proposed to use RDF to organize and distribute clinical RWD that is otherwise inaccessible externally. Our work serves as a pilot study that will lead to new clinical applications and could ultimately stimulate progress in the field of precision oncology.

Abstract Defining patient-to-patient similarity is essential for the development of precision medicine in clinical care and research. Conceptually, the identification of similar patient cohorts appears straightforward; however, universally accepted definitions remain elusive. Simultaneously, an explosion of vendors and published algorithms have emerged and all provide varied levels of functionality in identifying patient similarity categories. To provide clarity and a common framework for patient similarity, a workshop at the American Medical Informatics Association 2019 Annual Meeting was convened. This workshop included invited discussants from academics, the biotechnology industry, the FDA, and private practice oncology groups. Drawing from a broad range of backgrounds, workshop participants were able to coalesce around 4 major patient similarity classes: (1) feature, (2) outcome, (3) exposure, and (4) mixed-class. This perspective expands into these 4 subtypes more critically and offers the medical informatics community a means of communicating their work on this important topic.

Abstract Objective As coronavirus disease 2019 (COVID-19) started its rapid emergence and gradually transformed into an unprecedented pandemic, the need for having a knowledge repository for the disease became crucial. To address this issue, a new COVID-19 machine-readable dataset known as the COVID-19 Open Research Dataset (CORD-19) has been released. Based on this, our objective was to build a computable co-occurrence network embeddings to assist association detection among COVID-19–related biomedical entities. Materials and Methods Leveraging a Linked Data version of CORD-19 (ie, CORD-19-on-FHIR), we first utilized SPARQL to extract co-occurrences among chemicals, diseases, genes, and mutations and build a co-occurrence network. We then trained the representation of the derived co-occurrence network using node2vec with 4 edge embeddings operations (L1, L2, Average, and Hadamard). Six algorithms (decision tree, logistic regression, support vector machine, random forest, naïve Bayes, and multilayer perceptron) were applied to evaluate performance on link prediction. An unsupervised learning strategy was also developed incorporating the t-SNE (t-distributed stochastic neighbor embedding) and DBSCAN (density-based spatial clustering of applications with noise) algorithms for case studies. Results The random forest classifier showed the best performance on link prediction across different network embeddings. For edge embeddings generated using the Average operation, random forest achieved the optimal average precision of 0.97 along with a F1 score of 0.90. For unsupervised learning, 63 clusters were formed with silhouette score of 0.128. Significant associations were detected for 5 coronavirus infectious diseases in their corresponding subgroups. Conclusions In this study, we constructed COVID-19–centered co-occurrence network embeddings. Results indicated that the generated embeddings were able to extract significant associations for COVID-19 and coronavirus infectious diseases.

Abstract Objective The n2c2/UMass Lowell spin-off shared task focused on medical concept normalization (MCN) in clinical records. This task aimed to assess state-of-the-art methods for matching salient medical concepts from clinical records to a controlled vocabulary. We describe the task and the dataset used, compare the participating systems, and identify the strengths and limitations of the current approaches and directions for future research. Materials and Methods Participating teams were asked to link preselected text spans in discharge summaries (henceforth referred to as concept mentions) to the corresponding concepts in the SNOMED CT (Systematized Nomenclature of Medicine Clinical Terms) and RxNorm vocabularies from the Unified Medical Language System. The shared task used the MCN corpus created by the organizers, which maps all mentions of problems, treatments, and tests in the 2010 i2b2/VA challenge data to the Unified Medical Language System concepts. Submitted systems represented 4 broad categories of approaches: cascading dictionary matching, cosine distance, deep learning, and retrieve-and-rank systems. Disambiguation modules were common across all approaches. Results A total of 33 teams participated in the shared task. The best-performing team achieved an accuracy of 0.8526. The median and mean performances among all teams were 0.7733 and 0.7426, respectively. Conclusions Overall performance among the top 10 teams was high. However, particularly challenging for all teams were mentions requiring disambiguation of misspelled words, acronyms, abbreviations, and mentions with more than 1 possible semantic type. Complex mentions of long, multiword terms were also challenging and, in the future, will require better methods for learning contextualized representations of concept mentions and better use of domain knowledge.

Abstract Objective To design, develop, and evaluate a scalable clinical data normalization pipeline for standardizing unstructured electronic health record (EHR) data leveraging the HL7 Fast Healthcare Interoperability Resources (FHIR) specification. Methods We established an FHIR-based clinical data normalization pipeline known as NLP2FHIR that mainly comprises: (1) a module for a core natural language processing (NLP) engine with an FHIR-based type system; (2) a module for integrating structured data; and (3) a module for content normalization. We evaluated the FHIR modeling capability focusing on core clinical resources such as Condition, Procedure, MedicationStatement (including Medication), and FamilyMemberHistory using Mayo Clinic’s unstructured EHR data. We constructed a gold standard reusing annotation corpora from previous NLP projects. Results A total of 30 mapping rules, 62 normalization rules, and 11 NLP-specific FHIR extensions were created and implemented in the NLP2FHIR pipeline. The elements that need to integrate structured data from each clinical resource were identified. The performance of unstructured data modeling achieved F scores ranging from 0.69 to 0.99 for various FHIR element representations (0.69–0.99 for Condition; 0.75–0.84 for Procedure; 0.71–0.99 for MedicationStatement; and 0.75–0.95 for FamilyMemberHistory). Conclusion We demonstrated that the NLP2FHIR pipeline is feasible for modeling unstructured EHR data and integrating structured elements into the model. The outcomes of this work provide standards-based tools of clinical data normalization that is indispensable for enabling portable EHR-driven phenotyping and large-scale data analytics, as well as useful insights for future developments of the FHIR specifications with regard to handling unstructured clinical data.