Automating data extraction in systematic reviews: a systematic review

Background Automation of the parts of systematic review process, specifically the data extraction step, may be an important strategy to reduce the time necessary to complete a systematic review. However, the state of the science of automatically extracting data elements from full texts has not been well described. This paper performs a systematic review of published and unpublished methods to automate data extraction for systematic reviews. Methods We systematically searched PubMed, IEEEXplore, and ACM Digital Library to identify potentially relevant articles. We included reports that met the following criteria: 1) methods or results section described what entities were or need to be extracted, and 2) at least one entity was automatically extracted with evaluation results that were presented for that entity. We also reviewed the citations from included reports. Results Out of a total of 1190 unique citations that met our search criteria, we found 26 published reports describing automatic extraction of at least one of more than 52 potential data elements used in systematic reviews. For 25 (48 %) of the data elements used in systematic reviews, there were attempts from various researchers to extract information automatically from the publication text. Out of these, 14 (27 %) data elements were completely extracted, but the highest number of data elements extracted automatically by a single study was 7. Most of the data elements were extracted with F-scores (a mean of sensitivity and positive predictive value) of over 70 %. Conclusions We found no unified information extraction framework tailored to the systematic review process, and published reports focused on a limited (1–7) number of data elements. Biomedical natural language processing techniques have not been fully utilized to fully or even partially automate the data extraction step of systematic reviews.

different written resources [7]. Information extraction primarily constitutes concept extraction, also known as named entity recognition, and relation extraction, also known as association extraction. NLP handles written text at level of documents, words, grammar, meaning, and context. NLP techniques have been used to automate extraction of genomic and clinical information from biomedical literature. Similarly, automation of the data extraction step of the systematic review process through NLP may be one strategy to reduce the time necessary to complete and update a systematic review. The data extraction step is one of the most time-consuming steps of a systematic review. Automating or even semiautomating this step could substantially decrease the time taken to complete systematic reviews and thus decrease the time lag for research evidence to be translated into clinical practice. Despite these potential gains from NLP, the state of the science of automating data extraction has not been well described.
To date, there is limited knowledge and methods on how to automate the data extraction phase of the systematic reviews, despite being one of the most timeconsuming steps. To address this gap in knowledge, we sought to perform a systematic review of methods to automate the data extraction component of the systematic review process.

Methods
Our methodology was based on the Standards for Systematic Reviews set by the Institute of Medicine [8]. We conducted our study procedures as detailed below with input from the Cochrane Heart Group US Satellite.

Eligibility criteria
We included a report that met the following criteria: 1) the methods or results section describes what entities were or needed to be extracted, and 2) at least one entity was automatically extracted with evaluation results that were presented for that entity.
We excluded a report that met any of the following criteria: 1) the methods were not applied to the data extraction step of a systematic review; 2) the report was an editorial, commentary, or other non-original research report; or 3) there was no evaluation component.

Information sources and searches
For collecting the initial set of articles for our review, we developed search strategies with the help of the Cochrane Heart Group US Satellite, which includes systematic reviewers and a medical librarian. We refined these strategies using relevant citations from related papers. We searched three datasets: PubMed, IEEExplore, and ACM digital library, and our searches were limited between January 1, 2000 and January 6, 2015 (see Appendix 1). We restricted our search to these dates because biomedical information extraction algorithms prior to 2000 are unlikely to be accurate enough to be used for systematic reviews.
We retrieved articles that dealt with the extraction of various data elements, defined as categories of data that pertained to any information about or deriving from a study, including details of methods, participants, setting, context, interventions, outcomes, results, publications, and investigators [1] from included study reports. After we retrieved the initial set of reports from the search results, we then evaluated reports included in the references of these reports. We also sought expert opinion for additional relevant citations.

Study selection
We first de-duplicated the retrieve citations. For calibration and refinement of the inclusion and exclusion criteria, 100 citations were randomly selected and independently reviewed by a two authors (SRJ and PG). Disagreements were resolved by consensus with a third author (MH). In a second round, another set of 100 randomly selected abstracts was independently reviewed by two study authors (SRJ and PG), whereby we achieved a strong level of agreement (kappa = 0.97). Given the high level of agreement, the remaining studies were reviewed only by one author (PG). In this phase, we identified reports as "not relevant" or "potentially relevant".
Two authors (PG and SRJ) independently reviewed the full text of all citations (N = 74) that were identified as "potentially relevant". We classified included reports into various categories based on the particular data element that they attempted to extract from the original, scientific articles. Example of these data elements might be overall evidence, specific interventions, among others (Table 1). We resolved disagreements between the two reviewers through consensus with a third author (MDH).

Data collection process
Two authors (PG and SRJ) independently reviewed the included articles to extract data, such as the particular entity automatically extracted by the study, algorithm or technique used, and evaluation results into a data abstraction spreadsheet. We resolved disagreements through consensus with a third author (MDH).

Data items
We reviewed the Cochrane Handbook for Systematic Reviews [1], the CONsolidated Standards Of Reporting Trials (CONSORT) [9] statement, the Standards for Reporting of Diagnostic Accuracy (STARD) initiative [10], and PICO [11], PECODR [12], and PIBOSO [13] frameworks to obtain the data elements to be considered. PICO stands for Population, Intervention, Comparison, Outcomes; PECODR stands for Patient-Population-Problem,

Data synthesis and analysis
Because of the large variation in study methods and measurements, a meta-analysis of methodological features and contextual factors associated with the frequency of data extraction methods was not possible. We therefore present a narrative synthesis of our findings. We did not thoroughly assess risk of bias, including reporting bias, for these reports because the study designs did not match domains evaluated in commonly used instruments such as the Cochrane Risk of Bias tool [1] or QUADAS-2 instrument used for systematic reviews of randomized trials and diagnostic test accuracy studies, respectively [14].

Study selection
Of 1190 unique citations retrieved, we selected 75 reports for full-text screening, and we included 26 articles that met our inclusion criteria (Fig. 1). Agreement on abstract and full-text screening was 0.97 and 1.00. Table 1 provides a list of items to be considered in the data extraction process based on the Cochrane Handbook (Appendix 2) [1], CONSORT statement [9], STARD initiative [10], and PICO [11], PECODR [12], and PIBOSO [13] frameworks. We provide the major group for each field and report which standard focused on that field. Finally, we report whether there was a published method to extract that field. Table 1 also identifies the data elements relevant to systematic review process categorized by their domain and the standard from which the element was adopted and was associated with existing automation methods, where present.  [36,37,40] concept/neither: column 5), whether the extraction was done from full-text or abstracts (full text/abstract: column 6) and the main accuracy results reported by the system (results: column 7). The studies are arranged by increasing complexity by ordering studies that classified sentences before those that extracted the concepts and ordering studies that extracted data from abstracts before those that extracted data from full-text reports. The accuracy of most (N = 18, 69 %) studies was measured using a standard text mining metric known as F-score, which is the harmonic mean of precision (positive predictive value) and recall (sensitivity). Some studies (N = 5, 19 %) reported only the precision of their method, while some reported the accuracy values (N = 2, 8 %). One study (4 %) reported P5 precision, which indicates the fraction of positive predictions among the top 5 results returned by the system.

Studies that did not implement a data extraction system
Dawes et al. [12] identified 20 evidence-based medicine journal synopses with 759 extracts in the corresponding PubMed abstracts. Annotators agreed with the identification of an element 85 and 87 % for the evidence-based medicine synopses and PubMed abstracts, respectively. After consensus among the annotators, agreement rose to 97 and 98 %, respectively. The authors proposed various lexical patterns and developed rules to discover each PECODR element from the PubMed abstracts and the corresponding evidence-based medicine journal synopses that might make it possible to partially or fully automate the data extraction process.

Studies that identified sentences but did not extract data elements from abstracts only
Kim et al. [13] used conditional random fields (CRF) [15] for the task of classifying sentences in one of the PICO categories. The features were based on lexical, syntactic, structural, and sequential information in the data. The authors found that unigrams, section headings, and sequential information from preceding sentences were useful features for the classification task. They used 1000 medical abstracts from PIBOSO corpus and achieved micro-averaged F-scores of 91 and 67 % over datasets of structured and unstructured abstracts, respectively.    Boudin et al. [16] utilized a combination of multiple supervised classification techniques for detecting PICO elements in the medical abstracts. They utilized features such as MeSH semantic types, word overlap with title, number of punctuation marks on random forests (RF), naive Bayes (NB), support vector machines (SVM), and multi-layer perceptron (MLP) classifiers. Using 26,000 abstracts from PubMed, the authors took the first sentence in the structured abstracts and assigned a label automatically to build a large training data. They obtained an F-score of 86 % for identifying participants (P), 67 % for interventions (I) and controls (C), and 56 % for outcomes (O).
Huang et al. [17] used a naive Bayes classifier for the PICO classification task. The training data were generated automatically from the structured abstracts. For instance, all sentences in the section of the structured abstract that started with the term "PATIENT" were used to identify participants (P). In this way, the authors could generate a dataset of 23,472 sentences. Using 23,472 sentences from the structured abstracts, they obtained an F-score of 91 % for identifying participants (P), 75 % for interventions (I), and 88 % for outcomes (O).
Verbeke et al. [18] used a statistical relational learningbased approach (kLog) that utilized relational features for classifying sentences. The authors also used the PIBOSO corpus for evaluation and achieved micro-averaged Fscore of 84 % on structured abstracts and 67 % on unstructured abstracts, which was a better performance than Kim et al. [13].
Huang et al. [19] used 19,854 structured extracts and trained two classifiers: one by taking the first sentences of each section (termed CF by the authors) and the other by taking all the sentences in each section (termed CA by the authors). The authors used the naive Bayes classifier and achieved F-scores of 74, 66, and 73 % for identifying participants (P), interventions (I), and outcomes (O),   [20] used the PIBOSO corpus for the identification of sentences with PIBOSO elements. Using conditional random fields (CRF) with discriminative set of features, they achieved micro-averaged F-score of 91 %.
Robinson [21] used four machine learning models, 1) support vector machines, 2) naive Bayes, 3) naive Bayes multinomial, and 4) logistic regression to identify medical abstracts that contained patient-oriented evidence or not. These data included morbidity, mortality, symptom severity, and health-related quality of life. On a dataset of 1356 PubMed abstracts, the authors achieved the highest accuracy using a support vector machines learning model and achieved an F-measure of 86 %.
Chung [22] utilized a full sentence parser to identify the descriptions of the assignment of treatment arms in clinical trials. The authors used predicate-argument structure along with other linguistic features with a maximum entropy classifier. They utilized 203 abstracts from randomized trials for training and 124 abstracts for testing and achieved an F-score of 76 %.
Hara and Matsumoto [23] dealt with the problem of extracting "patient population" and "compared treatments" from medical abstracts. Given a sentence from the abstract, the authors first performed base noun-phrase chunking and then categorized the base noun-phrase into one of the five classes: "disease", "treatment", "patient", "study", and "others" using support vector machine and conditional random field models. After categorization, the authors used regular expression to extract the target words for patient population and comparison. The authors used 200 abstracts including terms such as "neoplasms" and "clinical trial, phase III" and obtained 91 % accuracy for the task of noun phrase classification. For sentence classification, the authors obtained a precision of 80 % for patient population and 82 % for comparisons.

Studies that identified only sentences but did not extract data elements from full-text reports
Zhao et al. [24] used two classification tasks to extract study data including patient details, including one at the sentence level and another at the keyword level. The authors first used a five-class scheme including 1) patient, 2) result, 3) intervention, 4) study design, and 5) research goal and tried to classify sentences into one of these five classes. They further used six classes for keywords such as sex (e.g., male, female), age (e.g., 54-year-old), race (e.g., Chinese), condition (e.g., asthma), intervention, and study design (e.g., randomized trial). They utilized conditional random fields for the classification task. Using 19,893 medical abstracts and full-text articles from 17 journal websites, they achieved F-scores of 75 % for identifying patients, 61 % for intervention, 91 % for results, 79 % for study design, and 76 % for research goal.
Hsu et al. [25] attempted to classify whether a sentence contains the "hypothesis", "statistical method", "outcomes", or "generalizability" of the study and then extracted the values. Using 42 full-text papers, the authors obtained F-scores of 86 % for identifying hypothesis, 84 % for statistical method, 90 % for outcomes, and 59 % for generalizability.
Song et al. [26] used machine learning-based classifiers such as maximum entropy classifier (MaxEnt), support vector machines (SVM), multi-layer perceptron (MLP), naive Bayes (NB), and radial basis function network (RBFN) to classify the sentences into categories such as analysis (statistical facts found by clinical experiment), general (generally accepted scientific facts, process, and methodology), recommendation (recommendations about interventions), and rule (guidelines). They utilized the principle of information gain (IG) as well as genetic algorithm (GA) for feature selection. They used 346 sentences from the clinical guideline document and obtained an Fscore of 98 % for classifying sentences.
Marshall et al. [27] used soft-margin support vector machines in a joint model for risk of bias assessment along with supporting sentences for random sequence generation, allocation concealment, blinding of participants and personnel, and blinding of outcome assessment, among others. They utilized presence of unigrams in the supporting sentences as features in their model. Working with full text of 2200 clinical trials, the joint model achieved Fscores of 56, 48, 35, and 38 % for identifying sentences corresponding to random sequence generation, allocation concealment, blinding of participants and personnel, and blinding of outcome assessment, respectively.

Studies that identified data elements only from abstracts but not from full texts
Demner-Fushman and Lin [28] used a rule-based approach to identify sentences containing PICO. Using 275 manually annotated abstracts, the authors achieved an accuracy of 80 % for population extraction and 86 % for problem extraction. They also utilized a supervised classifier for outcome extraction and achieved accuracy from 64 to 95 % across various experiments.
Kelly and Yang [29] used regular expressions and gazetteer to extract the number of participants, participant age, gender, ethnicity, and study characteristics. The authors utilized 386 abstracts from PubMed obtained with the query "soy and cancer" and achieved F-scores of 96 % for identifying the number of participants, 100 % for age of participants, 100 % for gender of participants, 95 % for ethnicity of participants, 91 % for duration of study, and 87 % for health status of participants.
Hansen et al. [30] used support vector machines [31] to extract number of trial participants from abstracts of the randomized control trials. The authors utilized features such as part-of-speech tag of the previous and next words and whether the sentence is grammatically complete (contained a verb). Using 233 abstracts from PubMed, they achieved an F-score of 86 % for identifying participants.
Xu et al. [32] utilized text classifications augmented with hidden Markov models [33] to identify sentences about subject demographics. These sentences were then parsed to extract information regarding participant descriptors (e.g., men, healthy, elderly), number of trial participants, disease/symptom name, and disease/symptom descriptors. After testing over 250 RCT abstracts, the authors obtained an accuracy of 83 % for participant descriptors: 83 %, 93 % for number of trial participants, 51 % for diseases/symptoms, and 92 % for descriptors of diseases/symptoms.
Summerscales et al. [34] used a conditional random field-based approach to identify various named entities such as treatments (drug names or complex phrases) and outcomes. The authors extracted 100 abstracts of randomized trials from the BMJ and achieved F-scores of 49 % for identifying treatment, 82 % for groups, and 54 % for outcomes.
Summerscales et al. [35] also proposed a method for automatic summarization of results from the clinical trials. The authors first identified the sentences that contained at least one integer (group size, outcome numbers, etc.). They then used the conditional random field classifier to find the entity mentions corresponding to treatment groups or outcomes. The treatment groups, outcomes, etc. were then treated as various "events." To identify all the relevant information for these events, the authors utilized templates with slots. The slots were then filled using a maximum entropy classifier. They utilized 263 abstracts from the BMJ and achieved F-scores of 76 % for identifying groups, 42 % for outcomes, 80 % for group sizes, and 71 % for outcome numbers.

Studies that identified data elements from full-text reports
Kiritchenko et al. [36] developed ExaCT, a tool that assists users with locating and extracting key trial characteristics such as eligibility criteria, sample size, drug dosage, and primary outcomes from full-text journal articles. The authors utilized a text classifier in the first stage to recover the relevant sentences. In the next stage, they utilized extraction rules to find the correct solutions. The authors evaluated their system using 50 full-text articles describing randomized trials with 1050 test instances and achieved a P5 precision of 88 % for identifying the classifier. Precision and recall of their extraction rules was found to be 93 and 91 %, respectively.
Restificar et al. [37] utilized latent Dirichlet allocation [38] to infer the latent topics in the sample documents and then used logistic regression to compute the probability that a given candidate criterion belongs to a particular topic. Using 44,203 full-text reports of randomized trials, the authors achieved accuracies of 75 and 70 % for inclusion and exclusion criteria, respectively.
Lin et al. [39] used linear-chain conditional random field for extracting various metadata elements such as number of patients, age group of the patients, geographical area, intervention, and time duration of the study. Using 93 full-text articles, the authors achieved a threefold cross validation precision of 43 % for identifying number of patients, 63 % for age group, 44 % for geographical area, 40 % for intervention, and 83 % for time period.
De Bruijn et al. [40] used support vector machine classifier to first identify sentences describing information elements such as eligibility criteria, sample size, etc. The authors then used manually crafted weak extraction rules to extract various information elements. Testing this twostage architecture on 88 randomized trial reports, they obtained a precision of 69 % for identifying eligibility criteria, 62 % for sample size, 94 % for treatment duration, 67 % for intervention, 100 % for primary outcome estimates, and 67 % for secondary outcomes.
Zhu et al. [41] also used manually crafted rules to extract various subject demographics such as disease, age, gender, and ethnicity. The authors tested their method on 50 articles and for disease extraction obtained an F-score of 64 and 85 % for exactly matched and partially matched cases, respectively.

Risk of bias across studies
In general, many studies have a high risk of selection bias because the gold standards used in the respective studies were not randomly selected. The risk of performance bias is also likely to be high because the investigators were not blinded. For the systems that used rule-based approaches, it was unclear whether the gold standard was used to train the rules or if there were a separate training set. The risk of attrition bias is unclear based on the study design of these non-randomized studies evaluating the performance of NLP methods. Lastly, the risk of reporting bias is unclear because of the lack of protocols in the development, implementation, and evaluation of NLP methods.

Related reviews and studies
Previous reviews on the automation of systematic review processes describe technologies for automating the overall process or other steps. Tsafnat et al. [43] surveyed the informatics systems that automate some of the tasks of systematic review and report systems for each stage of systematic review. Here, we focus on data extraction. None of the existing reviews [43][44][45][46][47] focus on the data extraction step. For example, Tsafnat et al. [43] presented a review of techniques to automate various aspects of systematic reviews, and while data extraction has been described as a task in their review, they only highlighted three studies as an acknowledgement of the ongoing work. In comparison, we identified 26 studies and critically examined their contribution in relation to all the data elements that need to be extracted to fully support the data extraction step. Thomas et al. [44] described the application of text mining technologies such as automatic term recognition, document clustering, classification, and summarization to support the identification of relevant studies in systematic reviews. The authors also pointed out the potential of these technologies to assist at various stages of the systematic review. Slaughter et al. [45] discussed necessary next steps towards developing "living systematic reviews" rather than a static publication, where the systematic reviews can be continuously updated with the latest knowledge available. The authors mentioned the need for development of new tools for reporting on and searching for structured data from clinical trials.
Tsafnat et al. [46] described four main tasks in systematic review: identifying the relevant studies, evaluating risk of bias in selected trials, synthesis of the evidence, and publishing the systematic reviews by generating human-readable text from trial reports. They mentioned text extraction algorithms for evaluating risk of bias and evidence synthesis but remain limited to one particular method for extraction of PICO elements.
Most natural language processing research has focused on reducing the workload for the screening step of systematic reviews (Step 3). Wallace et al. [48,49] and Miwa et al. [50] proposed an active learning framework to reduce the workload in citation screening for inclusion in the systematic reviews. Jonnalagadda et al. [51] designed a distributional semantics-based relevance feedback model to semi-automatically screen citations. Cohen et al. [52] proposed a module for grouping studies that are closely related and an automated system to rank publications according to the likelihood for meeting the inclusion criteria of a systematic review. Choong et al. [53] proposed an automated method for automatic citation snowballing to recursively pursue relevant literature for helping in evidence retrieval for systematic reviews. Cohen et al. [54] constructed a voting perceptronbased automated citation classification system to classify each article as to whether it contains high-quality, drugspecific evidence. Adeva et al. [55] also proposed a classification system for screening articles for systematic review. Shemilt et al. [56] also discussed the use of text mining to reduce screening workload in systematic reviews.

Research implications No standard gold standards or dataset
Among the 26 studies included in this systematic review, only three of them use a common corpus, namely 1000 medical abstracts from the PIBOSO corpus. Unfortunately, even that corpus facilitates only classification of sentences into whether they contain one of the data elements corresponding to the PIBOSO categories. No two other studies shared the same gold standard or dataset for evaluation. This limitation made it impossible for us to compare and assess the relative significance of the reported accuracy measures.

Separate systems for each data element
Few data elements, which are also relatively straightforward to extract automatically, such as the total number of participants (14 overall and 5 for extracting the actual data elements), have a relatively higher number of studies aiming towards extracting the same data element. This is not the case with other data elements. There are 27 out of 52 potential data elements that have not been explored for automated extraction, even if for highlighting the sentences containing them; seven more data elements were explored just by one study. There are 38 out of 52 potential data elements (>70 %) that have not been explored for automated extraction of the actual data elements; three more data elements were explored just by one study. The highest number of data elements extracted by a single study is only seven (14 %). This finding means that not only are more studies needed to explore the remaining 70 % data elements, but that there is an urgent need for a unified framework or system to extract all necessary data elements. The current state of informatics research for data extraction is exploratory, and multiple studies need to be conducted using the same gold standard and on the extraction of the same data elements for effective comparison.

Limitations
Our study has limitations. First, there is a possibility that data extraction algorithms were not published in journals or that our search might have missed them. We sought to minimize this limitation by searching in multiple bibliographic databases, including PubMed, IEEExplore, and ACM Digital Library. However, investigators may have also failed to publish algorithms that had lower F-scores than were previously reported, which we would not have captured. Second, we did not publish a protocol a priori, and our initial findings may have influenced our methods. However, we performed key steps, including screening, full-text review, and data extraction in duplicate to minimize potential bias in our systematic review.

Future work
"On demand" access to summarized evidence and best practices has been considered a sound strategy to satisfy clinicians' information needs and enhance decisionmaking [57][58][59][60][61][62][63][64][65]. A systematic review of 26 studies concluded that information-retrieval technology produces positive impact on physicians in terms of decision enhancement, learning, recall, reassurance, and confirmation [62]. Slaughter et al. [45] discussed necessary next steps towards developing "living systematic reviews" rather than a static publication, where the systematic reviews can be continuously updated with the latest knowledge available. The authors mention the need for development of new tools for reporting on and searching for structured data from published literature. Automated information extraction framework that extract data elements have the potential to assist the systematic reviewers and to eventually automate the screening and data extraction steps.
Medical science is currently witnessing a rapid pace at which medical knowledge is being created-75 clinical trials a day [66]. Evidence-based medicine [67] requires clinicians to keep up with published scientific studies and use them at the point of care. However, it has been shown that it is practically impossible to do that even within a narrow specialty [68]. A critical barrier is that finding relevant information, which may be located in several documents, takes an amount of time and cognitive effort that is incompatible with the busy clinical workflow [69,70]. Rapid systematic reviews using automation technologies will enable clinicians with up-todate and systematic summaries of the latest evidence.

Conclusions
Our systematic review describes previously reported methods to identify sentences containing some of the data elements for systematic reviews and only a few studies that have reported methods to extract these data elements. However, most of the data elements that would need to be considered for systematic reviews have been insufficiently explored to date, which identifies a major scope for future work. We hope that these automated extraction approaches might first act as checks for manual data extraction currently performed in duplicate; then serve to validate manual data extraction done by a single reviewer; then become the primary source for data element extraction that would be validated by a human; and eventually completely automate data extraction to enable living systematic reviews.

Search strategies
Below, we provide the search strategies used in PubMed, ACM Digital Library, and IEEExplore. The search was conducted on January 6, 2015. PubMed