Megan Kaiser

Web 2.0-Based Crowdsourcing for High-Quality Gold Standard Development in Clinical Natural Language Processing

Background A high-quality gold standard is vital for supervised, machine learning-based, clinical natural language processing (NLP) systems. In clinical NLP projects, expert annotators traditionally create the gold standard. However, traditional annotation is expensive and time-consuming. To reduce the cost of annotation, general NLP projects have turned to crowdsourcing based on Web 2.0 technology, which involves submitting smaller subtasks to a coordinated marketplace of workers on the Internet. Many studies have been conducted in the area of crowdsourcing, but only a few have focused on tasks in the general NLP field and only a handful in the biomedical domain, usually based upon very small pilot sample sizes. In addition, the quality of the crowdsourced biomedical NLP corpora were never exceptional when compared to traditionally-developed gold standards. The previously reported results on medical named entity annotation task showed a 0.68 F-measure based agreement between crowdsourced and traditionally-developed corpora. Objective Building upon previous work from the general crowdsourcing research, this study investigated the usability of crowdsourcing in the clinical NLP domain with special emphasis on achieving high agreement between crowdsourced and traditionally-developed corpora. Methods To build the gold standard for evaluating the crowdsourcing workers’ performance, 1042 clinical trial announcements (CTAs) from the ClinicalTrials.gov website were randomly selected and double annotated for medication names, medication types, and linked attributes. For the experiments, we used CrowdFlower, an Amazon Mechanical Turk-based crowdsourcing platform. We calculated sensitivity, precision, and F-measure to evaluate the quality of the crowd’s work and tested the statistical significance (P<.001, chi-square test) to detect differences between the crowdsourced and traditionally-developed annotations. Results The agreement between the crowd’s annotations and the traditionally-generated corpora was high for: (1) annotations (0.87, F-measure for medication names; 0.73, medication types), (2) correction of previous annotations (0.90, medication names; 0.76, medication types), and excellent for (3) linking medications with their attributes (0.96). Simple voting provided the best judgment aggregation approach. There was no statistically significant difference between the crowd and traditionally-generated corpora. Our results showed a 27.9% improvement over previously reported results on medication named entity annotation task. Conclusions This study offers three contributions. First, we proved that crowdsourcing is a feasible, inexpensive, fast, and practical approach to collect high-quality annotations for clinical text (when protected health information was excluded). We believe that well-designed user interfaces and rigorous quality control strategy for entity annotation and linking were critical to the success of this work. Second, as a further contribution to the Internet-based crowdsourcing field, we will publicly release the JavaScript and CrowdFlower Markup Language infrastructure code that is necessary to utilize CrowdFlower’s quality control and crowdsourcing interfaces for named entity annotations. Finally, to spur future research, we will release the CTA annotations that were generated by traditional and crowdsourced approaches.

Evaluating the impact of pre-annotation on annotation speed and potential bias: natural language processing gold standard development for clinical named entity recognition in clinical trial announcements

Objective To present a series of experiments: (1) to evaluate the impact of pre-annotation on the speed of manual annotation of clinical trial announcements; and (2) to test for potential bias, if pre-annotation is utilized. Methods To build the gold standard, 1400 clinical trial announcements from the clinicaltrials.gov website were randomly selected and double annotated for diagnoses, signs, symptoms, Unified Medical Language System (UMLS) Concept Unique Identifiers, and SNOMED CT codes. We used two dictionary-based methods to pre-annotate the text. We evaluated the annotation time and potential bias through F-measures and ANOVA tests and implemented Bonferroni correction. Results Time savings ranged from 13.85% to 21.5% per entity. Inter-annotator agreement (IAA) ranged from 93.4% to 95.5%. There was no statistically significant difference for IAA and annotator performance in pre-annotations. Conclusions On every experiment pair, the annotator with the pre-annotated text needed less time to annotate than the annotator with non-labeled text. The time savings were statistically significant. Moreover, the pre-annotation did not reduce the IAA or annotator performance. Dictionary-based pre-annotation is a feasible and practical method to reduce the cost of annotation of clinical named entity recognition in the eligibility sections of clinical trial announcements without introducing bias in the annotation process.

Phenotyping for patient safety: algorithm development for electronic health record based automated adverse event and medical error detection in neonatal intensive care

Background Although electronic health records (EHRs) have the potential to provide a foundation for quality and safety algorithms, few studies have measured their impact on automated adverse event (AE) and medical error (ME) detection within the neonatal intensive care unit (NICU) environment. Objective This paper presents two phenotyping AE and ME detection algorithms (ie, IV infiltrations, narcotic medication oversedation and dosing errors) and describes manual annotation of airway management and medication/fluid AEs from NICU EHRs. Methods From 753 NICU patient EHRs from 2011, we developed two automatic AE/ME detection algorithms, and manually annotated 11 classes of AEs in 3263 clinical notes. Performance of the automatic AE/ME detection algorithms was compared to trigger tool and voluntary incident reporting results. AEs in clinical notes were double annotated and consensus achieved under neonatologist supervision. Sensitivity, positive predictive value (PPV), and specificity are reported. Results Twelve severe IV infiltrates were detected. The algorithm identified one more infiltrate than the trigger tool and eight more than incident reporting. One narcotic oversedation was detected demonstrating 100% agreement with the trigger tool. Additionally, 17 narcotic medication MEs were detected, an increase of 16 cases over voluntary incident reporting. Conclusions Automated AE/ME detection algorithms provide higher sensitivity and PPV than currently used trigger tools or voluntary incident-reporting systems, including identification of potential dosing and frequency errors that current methods are unequipped to detect.

Mining FDA drug labels for medical conditions

BackgroundCincinnati Children’s Hospital Medical Center (CCHMC) has built the initial Natural Language Processing (NLP) component to extract medications with their corresponding medical conditions (Indications, Contraindications, Overdosage, and Adverse Reactions) as triples of medication-related information ([(1) drug name]-[(2) medical condition]-[(3) LOINC section header]) for an intelligent database system, in order to improve patient safety and the quality of health care. The Food and Drug Administration’s (FDA) drug labels are used to demonstrate the feasibility of building the triples as an intelligent database system task.MethodsThis paper discusses a hybrid NLP system, called AutoMCExtractor, to collect medical conditions (including disease/disorder and sign/symptom) from drug labels published by the FDA. Altogether, 6,611 medical conditions in a manually-annotated gold standard were used for the system evaluation. The pre-processing step extracted the plain text from XML file and detected eight related LOINC sections (e.g. Adverse Reactions, Warnings and Precautions) for medical condition extraction. Conditional Random Fields (CRF) classifiers, trained on token, linguistic, and semantic features, were then used for medical condition extraction. Lastly, dictionary-based post-processing corrected boundary-detection errors of the CRF step. We evaluated the AutoMCExtractor on manually-annotated FDA drug labels and report the results on both token and span levels.ResultsPrecision, recall, and F-measure were 0.90, 0.81, and 0.85, respectively, for the span level exact match; for the token-level evaluation, precision, recall, and F-measure were 0.92, 0.73, and 0.82, respectively.ConclusionsThe results demonstrate that (1) medical conditions can be extracted from FDA drug labels with high performance; and (2) it is feasible to develop a framework for an intelligent database system.

Preparing an annotated gold standard corpus to share with extramural investigators for de-identification research

OBJECTIVE The current study aims to fill the gap in available healthcare de-identification resources by creating a new sharable dataset with realistic Protected Health Information (PHI) without reducing the value of the data for de-identification research. By releasing the annotated gold standard corpus with Data Use Agreement we would like to encourage other Computational Linguists to experiment with our data and develop new machine learning models for de-identification. This paper describes: (1) the modifications required by the Institutional Review Board before sharing the de-identification gold standard corpus; (2) our efforts to keep the PHI as realistic as possible; (3) and the tests to show the effectiveness of these efforts in preserving the value of the modified data set for machine learning model development. MATERIALS AND METHODS In a previous study we built an original de-identification gold standard corpus annotated with true Protected Health Information (PHI) from 3503 randomly selected clinical notes for the 22 most frequent clinical note types of our institution. In the current study we modified the original gold standard corpus to make it suitable for external sharing by replacing HIPAA-specified PHI with newly generated realistic PHI. Finally, we evaluated the research value of this new dataset by comparing the performance of an existing published in-house de-identification system, when trained on the new de-identification gold standard corpus, with the performance of the same system, when trained on the original corpus. We assessed the potential benefits of using the new de-identification gold standard corpus to identify PHI in the i2b2 and PhysioNet datasets that were released by other groups for de-identification research. We also measured the effectiveness of the i2b2 and PhysioNet de-identification gold standard corpora in identifying PHI in our original clinical notes. RESULTS Performance of the de-identification system using the new gold standard corpus as a training set was very close to training on the original corpus (92.56 vs. 93.48 overall F-measures). Best i2b2/PhysioNet/CCHMC cross-training performances were obtained when training on the new shared CCHMC gold standard corpus, although performances were still lower than corpus-specific trainings. DISCUSSION AND CONCLUSION We successfully modified a de-identification dataset for external sharing while preserving the de-identification research value of the modified gold standard corpus with limited drop in machine learning de-identification performance.

A sequence labeling approach to link medications and their attributes in clinical notes and clinical trial announcements for information extraction

Objective The goal of this work was to evaluate machine learning methods, binary classification and sequence labeling, for medication–attribute linkage detection in two clinical corpora. Data and methods We double annotated 3000 clinical trial announcements (CTA) and 1655 clinical notes (CN) for medication named entities and their attributes. A binary support vector machine (SVM) classification method with parsimonious feature sets, and a conditional random fields (CRF)-based multi-layered sequence labeling (MLSL) model were proposed to identify the linkages between the entities and their corresponding attributes. We evaluated the systems performance against the human-generated gold standard. Results The experiments showed that the two machine learning approaches performed statistically significantly better than the baseline rule-based approach. The binary SVM classification achieved 0.94 F-measure with individual tokens as features. The SVM model trained on a parsimonious feature set achieved 0.81 F-measure for CN and 0.87 for CTA. The CRF MLSL method achieved 0.80 F-measure on both corpora. Discussion and conclusions We compared the novel MLSL method with a binary classification and a rule-based method. The MLSL method performed statistically significantly better than the rule-based method. However, the SVM-based binary classification method was statistically significantly better than the MLSL method for both the CTA and CN corpora. Using parsimonious feature sets both the SVM-based binary classification and CRF-based MLSL methods achieved high performance in detecting medication name and attribute linkages in CTA and CN.

Increasing the efficiency of trial-patient matching: automated clinical trial eligibility Pre-screening for pediatric oncology patients

BackgroundManual eligibility screening (ES) for a clinical trial typically requires a labor-intensive review of patient records that utilizes many resources. Leveraging state-of-the-art natural language processing (NLP) and information extraction (IE) technologies, we sought to improve the efficiency of physician decision-making in clinical trial enrollment. In order to markedly reduce the pool of potential candidates for staff screening, we developed an automated ES algorithm to identify patients who meet core eligibility characteristics of an oncology clinical trial.MethodsWe collected narrative eligibility criteria from ClinicalTrials.gov for 55 clinical trials actively enrolling oncology patients in our institution between 12/01/2009 and 10/31/2011. In parallel, our ES algorithm extracted clinical and demographic information from the Electronic Health Record (EHR) data fields to represent profiles of all 215 oncology patients admitted to cancer treatment during the same period. The automated ES algorithm then matched the trial criteria with the patient profiles to identify potential trial-patient matches. Matching performance was validated on a reference set of 169 historical trial-patient enrollment decisions, and workload, precision, recall, negative predictive value (NPV) and specificity were calculated.ResultsWithout automation, an oncologist would need to review 163 patients per trial on average to replicate the historical patient enrollment for each trial. This workload is reduced by 85% to 24 patients when using automated ES (precision/recall/NPV/specificity: 12.6%/100.0%/100.0%/89.9%). Without automation, an oncologist would need to review 42 trials per patient on average to replicate the patient-trial matches that occur in the retrospective data set. With automated ES this workload is reduced by 90% to four trials (precision/recall/NPV/specificity: 35.7%/100.0%/100.0%/95.5%).ConclusionBy leveraging NLP and IE technologies, automated ES could dramatically increase the trial screening efficiency of oncologists and enable participation of small practices, which are often left out from trial enrollment. The algorithm has the potential to significantly reduce the effort to execute clinical research at a point in time when new initiatives of the cancer care community intend to greatly expand both the access to trials and the number of available trials.

Automated detection of medication administration errors in neonatal intensive care

OBJECTIVE To improve neonatal patient safety through automated detection of medication administration errors (MAEs) in high alert medications including narcotics, vasoactive medication, intravenous fluids, parenteral nutrition, and insulin using the electronic health record (EHR); to evaluate rates of MAEs in neonatal care; and to compare the performance of computerized algorithms to traditional incident reporting for error detection. METHODS We developed novel computerized algorithms to identify MAEs within the EHR of all neonatal patients treated in a level four neonatal intensive care unit (NICU) in 2011 and 2012. We evaluated the rates and types of MAEs identified by the automated algorithms and compared their performance to incident reporting. Performance was evaluated by physician chart review. RESULTS In the combined 2011 and 2012 NICU data sets, the automated algorithms identified MAEs at the following rates: fentanyl, 0.4% (4 errors/1005 fentanyl administration records); morphine, 0.3% (11/4009); dobutamine, 0 (0/10); and milrinone, 0.3% (5/1925). We found higher MAE rates for other vasoactive medications including: dopamine, 11.6% (5/43); epinephrine, 10.0% (289/2890); and vasopressin, 12.8% (54/421). Fluid administration error rates were similar: intravenous fluids, 3.2% (273/8567); parenteral nutrition, 3.2% (649/20124); and lipid administration, 1.3% (203/15227). We also found 13 insulin administration errors with a resulting rate of 2.9% (13/456). MAE rates were higher for medications that were adjusted frequently and fluids administered concurrently. The algorithms identified many previously unidentified errors, demonstrating significantly better sensitivity (82% vs. 5%) and precision (70% vs. 50%) than incident reporting for error recognition. CONCLUSIONS Automated detection of medication administration errors through the EHR is feasible and performs better than currently used incident reporting systems. Automated algorithms may be useful for real-time error identification and mitigation.

An end-to-end hybrid algorithm for automated medication discrepancy detection.

BackgroundIn this study we implemented and developed state-of-the-art machine learning (ML) and natural language processing (NLP) technologies and built a computerized algorithm for medication reconciliation. Our specific aims are: (1) to develop a computerized algorithm for medication discrepancy detection between patients’ discharge prescriptions (structured data) and medications documented in free-text clinical notes (unstructured data); and (2) to assess the performance of the algorithm on real-world medication reconciliation data.MethodsWe collected clinical notes and discharge prescription lists for all 271 patients enrolled in the Complex Care Medical Home Program at Cincinnati Children’s Hospital Medical Center between 1/1/2010 and 12/31/2013. A double-annotated, gold-standard set of medication reconciliation data was created for this collection. We then developed a hybrid algorithm consisting of three processes: (1) a ML algorithm to identify medication entities from clinical notes, (2) a rule-based method to link medication names with their attributes, and (3) a NLP-based, hybrid approach to match medications with structured prescriptions in order to detect medication discrepancies. The performance was validated on the gold-standard medication reconciliation data, where precision (P), recall (R), F-value (F) and workload were assessed.ResultsThe hybrid algorithm achieved 95.0%/91.6%/93.3% of P/R/F on medication entity detection and 98.7%/99.4%/99.1% of P/R/F on attribute linkage. The medication matching achieved 92.4%/90.7%/91.5% (P/R/F) on identifying matched medications in the gold-standard and 88.6%/82.5%/85.5% (P/R/F) on discrepant medications. By combining all processes, the algorithm achieved 92.4%/90.7%/91.5% (P/R/F) and 71.5%/65.2%/68.2% (P/R/F) on identifying the matched and the discrepant medications, respectively. The error analysis on algorithm outputs identified challenges to be addressed in order to improve medication discrepancy detection.ConclusionBy leveraging ML and NLP technologies, an end-to-end, computerized algorithm achieves promising outcome in reconciling medications between clinical notes and discharge prescriptions.

Pre-annotating Clinical Notes and Clinical Trial Announcements for Gold Standard Corpus Development: Evaluating the Impact on Annotation Speed and Potential Bias

In this study our aim was to present a series of experiments to evaluate the impact of pre-annotation: (1) on the speed of manual annotation of clinical notes and clinical trial announcements; and (2) test for potential bias if pre-annotation is utilized. The gold standard was 900 clinical trial announcements from clinicaltrials.gov website and 1655 clinical notes annotated for diagnoses, signs, symptoms, UMLS CUI and SNOMED CT codes. Two dictionary-based methods were used to pre-annotate the text. Annotation time savings ranged from 2.89% to 29.1% per entity. The pre-annotation did not reduce the IAA or annotator performance but reduced the time to annotate in every experiment. Dictionary-based pre-annotation is a feasible and practical method to reduce cost of annotation without introducing bias in the process.