Feature engineering combined with machine learning and rule-based methods for structured information extraction from narrative clinical discharge summaries

Objective

A system that translates narrative text in the medical domain into structured representation is in great demand. The system performs three sub-tasks: concept extraction, assertion classification, and relation identification.

Design

The overall system consists of five steps: (1) pre-processing sentences, (2) marking noun phrases (NPs) and adjective phrases (APs), (3) extracting concepts that use a dosage-unit dictionary to dynamically switch two models based on Conditional Random Fields (CRF), (4) classifying assertions based on voting of five classifiers, and (5) identifying relations using normalized sentences with a set of effective discriminating features.

Measurements

Macro-averaged and micro-averaged precision, recall and F-measure were used to evaluate results.

Results

The performance is competitive with the state-of-the-art systems with micro-averaged F-measure of 0.8489 for concept extraction, 0.9392 for assertion classification and 0.7326 for relation identification.

Conclusions

The system exploits an array of common features and achieves state-of-the-art performance. Prudent feature engineering sets the foundation of our systems. In concept extraction, we demonstrated that switching models, one of which is especially designed for telegraphic sentences, improved extraction of the treatment concept significantly. In assertion classification, a set of features derived from a rule-based classifier were proven to be effective for the classes such as conditional and possible. These classes would suffer from data scarcity in conventional machine-learning methods. In relation identification, we use two-staged architecture, the second of which applies pairwise classifiers to possible candidate classes. This architecture significantly improves performance.     

Biomedical negation scope detection with conditional random fields

Objective

Negation is a linguistic phenomenon that marks the absence of an entity or event. Negated events are frequently reported in both biological literature and clinical notes. Text mining applications benefit from the detection of negation and its scope. However, due to the complexity of language, identifying the scope of negation in a sentence is not a trivial task.

Design

Conditional random fields (CRF), a supervised machine-learning algorithm, were used to train models to detect negation cue phrases and their scope in both biological literature and clinical notes. The models were trained on the publicly available BioScope corpus.

Measurement

The performance of the CRF models was evaluated on identifying the negation cue phrases and their scope by calculating recall, precision and F1-score. The models were compared with four competitive baseline systems.

Results

The best CRF-based model performed statistically better than all baseline systems and NegEx, achieving an F1-score of 98% and 95% on detecting negation cue phrases and their scope in clinical notes, and an F1-score of 97% and 85% on detecting negation cue phrases and their scope in biological literature.

Conclusions

This approach is robust, as it can identify negation scope in both biological and clinical text. To benefit text mining applications, the system is publicly available as a Java API and as an online application at http://negscope.askhermes.org.     

Using rule-based natural language processing to improve disease normalization in biomedical text

Background and objective

In order for computers to extract useful information from unstructured text, a concept normalization system is needed to link relevant concepts in a text to sources that contain further information about the concept. Popular concept normalization tools in the biomedical field are dictionary-based. In this study we investigate the usefulness of natural language processing (NLP) as an adjunct to dictionary-based concept normalization.

Methods

We compared the performance of two biomedical concept normalization systems, MetaMap and Peregrine, on the Arizona Disease Corpus, with and without the use of a rule-based NLP module. Performance was assessed for exact and inexact boundary matching of the system annotations with those of the gold standard and for concept identifier matching.

Results

Without the NLP module, MetaMap and Peregrine attained F-scores of 61.0% and 63.9%, respectively, for exact boundary matching, and 55.1% and 56.9% for concept identifier matching. With the aid of the NLP module, the F-scores of MetaMap and Peregrine improved to 73.3% and 78.0% for boundary matching, and to 66.2% and 69.8% for concept identifier matching. For inexact boundary matching, performances further increased to 85.5% and 85.4%, and to 73.6% and 73.3% for concept identifier matching.

Conclusions

We have shown the added value of NLP for the recognition and normalization of diseases with MetaMap and Peregrine. The NLP module is general and can be applied in combination with any concept normalization system. Whether its use for concept types other than disease is equally advantageous remains to be investigated.     

Integrating existing natural language processing tools for medication extraction from discharge summaries

Objective

To develop an automated system to extract medications and related information from discharge summaries as part of the 2009 i2b2 natural language processing (NLP) challenge. This task required accurate recognition of medication name, dosage, mode, frequency, duration, and reason for drug administration.

Design

We developed an integrated system using several existing NLP components developed at Vanderbilt University Medical Center, which included MedEx (to extract medication information), SecTag (a section identification system for clinical notes), a sentence splitter, and a spell checker for drug names. Our goal was to achieve good performance with minimal to no specific training for this document corpus; thus, evaluating the portability of those NLP tools beyond their home institution. The integrated system was developed using 17 notes that were annotated by the organizers and evaluated using 251 notes that were annotated by participating teams.

Measurements

The i2b2 challenge used standard measures, including precision, recall, and F-measure, to evaluate the performance of participating systems. There were two ways to determine whether an extracted textual finding is correct or not: exact matching or inexact matching. The overall performance for all six types of medication-related findings across 251 annotated notes was considered as the primary metric in the challenge.

Results

Our system achieved an overall F-measure of 0.821 for exact matching (0.839 precision; 0.803 recall) and 0.822 for inexact matching (0.866 precision; 0.782 recall). The system ranked second out of 20 participating teams on overall performance at extracting medications and related information.

Conclusions

The results show that the existing MedEx system, together with other NLP components, can extract medication information in clinical text from institutions other than the site of algorithm development with reasonable performance.     

Use of a support vector machine for categorizing free-text notes: assessment of accuracy across two institutions

Background

Electronic health record (EHR) users must regularly review large amounts of data in order to make informed clinical decisions, and such review is time-consuming and often overwhelming. Technologies like automated summarization tools, EHR search engines and natural language processing have been shown to help clinicians manage this information.

Objective

To develop a support vector machine (SVM)-based system for identifying EHR progress notes pertaining to diabetes, and to validate it at two institutions.

Materials and methods

We retrieved 2000 EHR progress notes from patients with diabetes at the Brigham and Women''s Hospital (1000 for training and 1000 for testing) and another 1000 notes from the University of Texas Physicians (for validation). We manually annotated all notes and trained a SVM using a bag of words approach. We then used the SVM on the testing and validation sets and evaluated its performance with the area under the curve (AUC) and F statistics.

Results

The model accurately identified diabetes-related notes in both the Brigham and Women''s Hospital testing set (AUC=0.956, F=0.934) and the external University of Texas Faculty Physicians validation set (AUC=0.947, F=0.935).

Discussion

Overall, the model we developed was quite accurate. Furthermore, it generalized, without loss of accuracy, to another institution with a different EHR and a distinct patient and provider population.

Conclusions

It is possible to use a SVM-based classifier to identify EHR progress notes pertaining to diabetes, and the model generalizes well.     

Research and applications: Supervised machine learning and active learning in classification of radiology reports

Objective

This paper presents an automated system for classifying the results of imaging examinations (CT, MRI, positron emission tomography) into reportable and non-reportable cancer cases. This system is part of an industrial-strength processing pipeline built to extract content from radiology reports for use in the Victorian Cancer Registry.

Materials and methods

In addition to traditional supervised learning methods such as conditional random fields and support vector machines, active learning (AL) approaches were investigated to optimize training production and further improve classification performance. The project involved two pilot sites in Victoria, Australia (Lake Imaging (Ballarat) and Peter MacCallum Cancer Centre (Melbourne)) and, in collaboration with the NSW Central Registry, one pilot site at Westmead Hospital (Sydney).

Results

The reportability classifier performance achieved 98.25% sensitivity and 96.14% specificity on the cancer registry''s held-out test set. Up to 92% of training data needed for supervised machine learning can be saved by AL.

Discussion

AL is a promising method for optimizing the supervised training production used in classification of radiology reports. When an AL strategy is applied during the data selection process, the cost of manual classification can be reduced significantly.

Conclusions

The most important practical application of the reportability classifier is that it can dramatically reduce human effort in identifying relevant reports from the large imaging pool for further investigation of cancer. The classifier is built on a large real-world dataset and can achieve high performance in filtering relevant reports to support cancer registries.     

Research and applications: N-gram support vector machines for scalable procedure and diagnosis classification,with applications to clinical free text data from the intensive care unit

Background

Existing risk adjustment models for intensive care unit (ICU) outcomes rely on manual abstraction of patient-level predictors from medical charts. Developing an automated method for abstracting these data from free text might reduce cost and data collection times.

Objective

To develop a support vector machine (SVM) classifier capable of identifying a range of procedures and diagnoses in ICU clinical notes for use in risk adjustment.

Materials and methods

We selected notes from 2001–2008 for 4191 neonatal ICU (NICU) and 2198 adult ICU patients from the MIMIC-II database from the Beth Israel Deaconess Medical Center. Using these notes, we developed an implementation of the SVM classifier to identify procedures (mechanical ventilation and phototherapy in NICU notes) and diagnoses (jaundice in NICU and intracranial hemorrhage (ICH) in adult ICU). On the jaundice classification task, we also compared classifier performance using n-gram features to unigrams with application of a negation algorithm (NegEx).

Results

Our classifier accurately identified mechanical ventilation (accuracy=0.982, F₁=0.954) and phototherapy use (accuracy=0.940, F₁=0.912), as well as jaundice (accuracy=0.898, F₁=0.884) and ICH diagnoses (accuracy=0.938, F₁=0.943). Including bigram features improved performance on the jaundice (accuracy=0.898 vs 0.865) and ICH (0.938 vs 0.927) tasks, and outperformed NegEx-derived unigram features (accuracy=0.898 vs 0.863) on the jaundice task.

Discussion

Overall, a classifier using n-gram support vectors displayed excellent performance characteristics. The classifier generalizes to diverse patient populations, diagnoses, and procedures.

Conclusions

SVM-based classifiers can accurately identify procedure status and diagnoses among ICU patients, and including n-gram features improves performance, compared to existing methods.     

Automatic extraction of relations between medical concepts in clinical texts

Objective

A supervised machine learning approach to discover relations between medical problems, treatments, and tests mentioned in electronic medical records.

Materials and methods

A single support vector machine classifier was used to identify relations between concepts and to assign their semantic type. Several resources such as Wikipedia, WordNet, General Inquirer, and a relation similarity metric inform the classifier.

Results

The techniques reported in this paper were evaluated in the 2010 i2b2 Challenge and obtained the highest F1 score for the relation extraction task. When gold standard data for concepts and assertions were available, F1 was 73.7, precision was 72.0, and recall was 75.3. F1 is defined as 2*Precision*Recall/(Precision+Recall). Alternatively, when concepts and assertions were discovered automatically, F1 was 48.4, precision was 57.6, and recall was 41.7.

Discussion

Although a rich set of features was developed for the classifiers presented in this paper, little knowledge mining was performed from medical ontologies such as those found in UMLS. Future studies should incorporate features extracted from such knowledge sources, which we expect to further improve the results. Moreover, each relation discovery was treated independently. Joint classification of relations may further improve the quality of results. Also, joint learning of the discovery of concepts, assertions, and relations may also improve the results of automatic relation extraction.

Conclusion

Lexical and contextual features proved to be very important in relation extraction from medical texts. When they are not available to the classifier, the F1 score decreases by 3.7%. In addition, features based on similarity contribute to a decrease of 1.1% when they are not available.     

MITRE system for clinical assertion status classification

Objective

To describe a system for determining the assertion status of medical problems mentioned in clinical reports, which was entered in the 2010 i2b2/VA community evaluation ‘Challenges in natural language processing for clinical data’ for the task of classifying assertions associated with problem concepts extracted from patient records.

Materials and methods

A combination of machine learning (conditional random field and maximum entropy) and rule-based (pattern matching) techniques was used to detect negation, speculation, and hypothetical and conditional information, as well as information associated with persons other than the patient.

Results

The best submission obtained an overall micro-averaged F-score of 0.9343.

Conclusions

Using semantic attributes of concepts and information about document structure as features for statistical classification of assertions is a good way to leverage rule-based and statistical techniques. In this task, the choice of features may be more important than the choice of classifier algorithm.     

Research and applications: Identifying medical terms in patient-authored text: a crowdsourcing-based approach

Background and objective

As people increasingly engage in online health-seeking behavior and contribute to health-oriented websites, the volume of medical text authored by patients and other medical novices grows rapidly. However, we lack an effective method for automatically identifying medical terms in patient-authored text (PAT). We demonstrate that crowdsourcing PAT medical term identification tasks to non-experts is a viable method for creating large, accurately-labeled PAT datasets; moreover, such datasets can be used to train classifiers that outperform existing medical term identification tools.

Materials and methods

To evaluate the viability of using non-expert crowds to label PAT, we compare expert (registered nurses) and non-expert (Amazon Mechanical Turk workers; Turkers) responses to a PAT medical term identification task. Next, we build a crowd-labeled dataset comprising 10 000 sentences from MedHelp. We train two models on this dataset and evaluate their performance, as well as that of MetaMap, Open Biomedical Annotator (OBA), and NaCTeM''s TerMINE, against two gold standard datasets: one from MedHelp and the other from CureTogether.

Results

When aggregated according to a corroborative voting policy, Turker responses predict expert responses with an F1 score of 84%. A conditional random field (CRF) trained on 10 000 crowd-labeled MedHelp sentences achieves an F1 score of 78% against the CureTogether gold standard, widely outperforming OBA (47%), TerMINE (43%), and MetaMap (39%). A failure analysis of the CRF suggests that misclassified terms are likely to be either generic or rare.

Conclusions

Our results show that combining statistical models sensitive to sentence-level context with crowd-labeled data is a scalable and effective technique for automatically identifying medical terms in PAT.     

A supervised framework for resolving coreference in clinical records

Objective

A method for the automatic resolution of coreference between medical concepts in clinical records.

Materials and methods

A multiple pass sieve approach utilizing support vector machines (SVMs) at each pass was used to resolve coreference. Information such as lexical similarity, recency of a concept mention, synonymy based on Wikipedia redirects, and local lexical context were used to inform the method. Results were evaluated using an unweighted average of MUC, CEAF, and B³ coreference evaluation metrics. The datasets used in these research experiments were made available through the 2011 i2b2/VA Shared Task on Coreference.

Results

The method achieved an average F score of 0.821 on the ODIE dataset, with a precision of 0.802 and a recall of 0.845. These results compare favorably to the best-performing system with a reported F score of 0.827 on the dataset and the median system F score of 0.800 among the eight teams that participated in the 2011 i2b2/VA Shared Task on Coreference. On the i2b2 dataset, the method achieved an average F score of 0.906, with a precision of 0.895 and a recall of 0.918 compared to the best F score of 0.915 and the median of 0.859 among the 16 participating teams.

Discussion

Post hoc analysis revealed significant performance degradation on pathology reports. The pathology reports were characterized by complex synonymy and very few patient mentions.

Conclusion

The use of several simple lexical matching methods had the most impact on achieving competitive performance on the task of coreference resolution. Moreover, the ability to detect patients in electronic medical records helped to improve coreference resolution more than other linguistic analysis.     

Research and applications: Imaging informatics for consumer health: towards a radiology patient portal

Objective

With the increased routine use of advanced imaging in clinical diagnosis and treatment, it has become imperative to provide patients with a means to view and understand their imaging studies. We illustrate the feasibility of a patient portal that automatically structures and integrates radiology reports with corresponding imaging studies according to several information orientations tailored for the layperson.

Methods

The imaging patient portal is composed of an image processing module for the creation of a timeline that illustrates the progression of disease, a natural language processing module to extract salient concepts from radiology reports (73% accuracy, F₁ score of 0.67), and an interactive user interface navigable by an imaging findings list. The portal was developed as a Java-based web application and is demonstrated for patients with brain cancer.

Results and discussion

The system was exhibited at an international radiology conference to solicit feedback from a diverse group of healthcare professionals. There was wide support for educating patients about their imaging studies, and an appreciation for the informatics tools used to simplify images and reports for consumer interpretation. Primary concerns included the possibility of patients misunderstanding their results, as well as worries regarding accidental improper disclosure of medical information.

Conclusions

Radiologic imaging composes a significant amount of the evidence used to make diagnostic and treatment decisions, yet there are few tools for explaining this information to patients. The proposed radiology patient portal provides a framework for organizing radiologic results into several information orientations to support patient education.     

Automatic extraction of medication information from medical discharge summaries

Objective

This article describes a system developed for the 2009 i2b2 Medication Extraction Challenge. The purpose of this challenge is to extract medication information from hospital discharge summaries.

Design

The system explored several linguistic natural language processing techniques (eg, term-based and token-based rule matching) to identify medication-related information in the narrative text. A number of lexical resources was constructed to profile lexical or morphological features for different categories of medication constituents.

Measurements

Performance was evaluated in terms of the micro-averaged F-measure at the horizontal system level.

Results

The automated system performed well, and achieved an F-micro of 80% for the term-level results and 81% for the token-level results, placing it sixth in exact matches and fourth in inexact matches in the i2b2 competition.

Conclusion

The overall results show that this relatively simple rule-based approach is capable of tackling multiple entity identification tasks such as medication extraction under situations in which few training documents are annotated for machine learning approaches, and the entity information can be characterized with a set of feature tokens.     

Research and applications: Syntactic parsing of clinical text: guideline and corpus development with handling ill-formed sentences

Objective

To develop, evaluate, and share: (1) syntactic parsing guidelines for clinical text, with a new approach to handling ill-formed sentences; and (2) a clinical Treebank annotated according to the guidelines. To document the process and findings for readers with similar interest.

Methods

Using random samples from a shared natural language processing challenge dataset, we developed a handbook of domain-customized syntactic parsing guidelines based on iterative annotation and adjudication between two institutions. Special considerations were incorporated into the guidelines for handling ill-formed sentences, which are common in clinical text. Intra- and inter-annotator agreement rates were used to evaluate consistency in following the guidelines. Quantitative and qualitative properties of the annotated Treebank, as well as its use to retrain a statistical parser, were reported.

Results

A supplement to the Penn Treebank II guidelines was developed for annotating clinical sentences. After three iterations of annotation and adjudication on 450 sentences, the annotators reached an F-measure agreement rate of 0.930 (while intra-annotator rate was 0.948) on a final independent set. A total of 1100 sentences from progress notes were annotated that demonstrated domain-specific linguistic features. A statistical parser retrained with combined general English (mainly news text) annotations and our annotations achieved an accuracy of 0.811 (higher than models trained purely with either general or clinical sentences alone). Both the guidelines and syntactic annotations are made available at https://sourceforge.net/projects/medicaltreebank.

Conclusions

We developed guidelines for parsing clinical text and annotated a corpus accordingly. The high intra- and inter-annotator agreement rates showed decent consistency in following the guidelines. The corpus was shown to be useful in retraining a statistical parser that achieved moderate accuracy.     

Automatic discourse connective detection in biomedical text

Objective

Relation extraction in biomedical text mining systems has largely focused on identifying clause-level relations, but increasing sophistication demands the recognition of relations at discourse level. A first step in identifying discourse relations involves the detection of discourse connectives: words or phrases used in text to express discourse relations. In this study supervised machine-learning approaches were developed and evaluated for automatically identifying discourse connectives in biomedical text.

Materials and Methods

Two supervised machine-learning models (support vector machines and conditional random fields) were explored for identifying discourse connectives in biomedical literature. In-domain supervised machine-learning classifiers were trained on the Biomedical Discourse Relation Bank, an annotated corpus of discourse relations over 24 full-text biomedical articles (∼112 000 word tokens), a subset of the GENIA corpus. Novel domain adaptation techniques were also explored to leverage the larger open-domain Penn Discourse Treebank (∼1 million word tokens). The models were evaluated using the standard evaluation metrics of precision, recall and F1 scores.

Results and Conclusion

Supervised machine-learning approaches can automatically identify discourse connectives in biomedical text, and the novel domain adaptation techniques yielded the best performance: 0.761 F1 score. A demonstration version of the fully implemented classifier BioConn is available at: http://bioconn.askhermes.org.     

An end-to-end system to identify temporal relation in discharge summaries: 2012 i2b2 challenge

Objective

To create an end-to-end system to identify temporal relation in discharge summaries for the 2012 i2b2 challenge. The challenge includes event extraction, timex extraction, and temporal relation identification.

Design

An end-to-end temporal relation system was developed. It includes three subsystems: an event extraction system (conditional random fields (CRF) name entity extraction and their corresponding attribute classifiers), a temporal extraction system (CRF name entity extraction, their corresponding attribute classifiers, and context-free grammar based normalization system), and a temporal relation system (10 multi-support vector machine (SVM) classifiers and a Markov logic networks inference system) using labeled sequential pattern mining, syntactic structures based on parse trees, and results from a coordination classifier. Micro-averaged precision (P), recall (R), averaged P&R (P&R), and F measure (F) were used to evaluate results.

Results

For event extraction, the system achieved 0.9415 (P), 0.8930 (R), 0.9166 (P&R), and 0.9166 (F). The accuracies of their type, polarity, and modality were 0.8574, 0.8585, and 0.8560, respectively. For timex extraction, the system achieved 0.8818, 0.9489, 0.9141, and 0.9141, respectively. The accuracies of their type, value, and modifier were 0.8929, 0.7170, and 0.8907, respectively. For temporal relation, the system achieved 0.6589, 0.7129, 0.6767, and 0.6849, respectively. For end-to-end temporal relation, it achieved 0.5904, 0.5944, 0.5921, and 0.5924, respectively. With the F measure used for evaluation, we were ranked first out of 14 competing teams (event extraction), first out of 14 teams (timex extraction), third out of 12 teams (temporal relation), and second out of seven teams (end-to-end temporal relation).

Conclusions

The system achieved encouraging results, demonstrating the feasibility of the tasks defined by the i2b2 organizers. The experiment result demonstrates that both global and local information is useful in the 2012 challenge.     

A Novel Hybrid Approach to Automated Negation Detection in Clinical Radiology Reports

Objective

Negation is common in clinical documents and is an important source of poor precision in automated indexing systems. Previous research has shown that negated terms may be difficult to identify if the words implying negations (negation signals) are more than a few words away from them. We describe a novel hybrid approach, combining regular expression matching with grammatical parsing, to address the above limitation in automatically detecting negations in clinical radiology reports.

Design

Negations are classified based upon the syntactical categories of negation signals, and negation patterns, using regular expression matching. Negated terms are then located in parse trees using corresponding negation grammar.

Measurements

A classification of negations and their corresponding syntactical and lexical patterns were developed through manual inspection of 30 radiology reports and validated on a set of 470 radiology reports. Another 120 radiology reports were randomly selected as the test set on which a modified Delphi design was used by four physicians to construct the gold standard.

Results

In the test set of 120 reports, there were a total of 2,976 noun phrases, of which 287 were correctly identified as negated (true positives), along with 23 undetected true negations (false negatives) and 4 mistaken negations (false positives). The hybrid approach identified negated phrases with sensitivity of 92.6% (95% CI 90.9–93.4%), positive predictive value of 98.6% (95% CI 96.9–99.4%), and specificity of 99.87% (95% CI 99.7–99.9%).

Conclusion

This novel hybrid approach can accurately locate negated concepts in clinical radiology reports not only when in close proximity to, but also at a distance from, negation signals.     

Automatically extracting sentences from Medline citations to support clinicians’ information needs

Objective

Online health knowledge resources contain answers to most of the information needs raised by clinicians in the course of care. However, significant barriers limit the use of these resources for decision-making, especially clinicians’ lack of time. In this study we assessed the feasibility of automatically generating knowledge summaries for a particular clinical topic composed of relevant sentences extracted from Medline citations.

Methods

The proposed approach combines information retrieval and semantic information extraction techniques to identify relevant sentences from Medline abstracts. We assessed this approach in two case studies on the treatment alternatives for depression and Alzheimer''s disease.

Results

A total of 515 of 564 (91.3%) sentences retrieved in the two case studies were relevant to the topic of interest. About one-third of the relevant sentences described factual knowledge or a study conclusion that can be used for supporting information needs at the point of care.

Conclusions

The high rate of relevant sentences is desirable, given that clinicians’ lack of time is one of the main barriers to using knowledge resources at the point of care. Sentence rank was not significantly associated with relevancy, possibly due to most sentences being highly relevant. Sentences located closer to the end of the abstract and sentences with treatment and comparative predications were likely to be conclusive sentences. Our proposed technical approach to helping clinicians meet their information needs is promising. The approach can be extended for other knowledge resources and information need types.     

A new algorithm for reducing the workload of experts in performing systematic reviews

Objective

To determine whether a factorized version of the complement naïve Bayes (FCNB) classifier can reduce the time spent by experts reviewing journal articles for inclusion in systematic reviews of drug class efficacy for disease treatment.

Design

The proposed classifier was evaluated on a test collection built from 15 systematic drug class reviews used in previous work. The FCNB classifier was constructed to classify each article as containing high-quality, drug class-specific evidence or not. Weight engineering (WE) techniques were added to reduce underestimation for Medical Subject Headings (MeSH)-based and Publication Type (PubType)-based features. Cross-validation experiments were performed to evaluate the classifier''s parameters and performance.

Measurements

Work saved over sampling (WSS) at no less than a 95% recall was used as the main measure of performance.

Results

The minimum workload reduction for a systematic review for one topic, achieved with a FCNB/WE classifier, was 8.5%; the maximum was 62.2% and the average over the 15 topics was 33.5%. This is 15.0% higher than the average workload reduction obtained using a voting perceptron-based automated citation classification system.

Conclusion

The FCNB/WE classifier is simple, easy to implement, and produces significantly better results in reducing the workload than previously achieved. The results support it being a useful algorithm for machine-learning-based automation of systematic reviews of drug class efficacy for disease treatment.     

A system for coreference resolution for the clinical narrative

Objective

To research computational methods for coreference resolution in the clinical narrative and build a system implementing the best methods.

Methods

The Ontology Development and Information Extraction corpus annotated for coreference relations consists of 7214 coreferential markables, forming 5992 pairs and 1304 chains. We trained classifiers with semantic, syntactic, and surface features pruned by feature selection. For the three system components—for the resolution of relative pronouns, personal pronouns, and noun phrases—we experimented with support vector machines with linear and radial basis function (RBF) kernels, decision trees, and perceptrons. Evaluation of algorithms and varied feature sets was performed using standard metrics.

Results

The best performing combination is support vector machines with an RBF kernel and all features (MUC score=0.352, B³=0.690, CEAF=0.486, BLANC=0.596) outperforming a traditional decision tree baseline.

Discussion

The application showed good performance similar to performance on general English text. The main error source was sentence distances exceeding a window of 10 sentences between markables. A possible solution to this problem is hinted at by the fact that coreferent markables sometimes occurred in predictable (although distant) note sections. Another system limitation is failure to fully utilize synonymy and ontological knowledge. Future work will investigate additional ways to incorporate syntactic features into the coreference problem.

Conclusion

We investigated computational methods for coreference resolution in the clinical narrative. The best methods are released as modules of the open source Clinical Text Analysis and Knowledge Extraction System and Ontology Development and Information Extraction platforms.