Background: The rapid growth of the biomedical literature makes identifying strong evidence a time-consuming task. Applying machine learning to the process could be a viable solution that limits effort while maintaining accuracy.
Objective: The goal of the research was to summarize the nature and comparative performance of machine learning approaches that have been applied to retrieve high-quality evidence for clinical consideration from the biomedical literature.
Methods: We conducted a systematic review of studies that applied machine learning techniques to identify high-quality clinical articles in the biomedical literature. Multiple databases were searched to July 2020. Extracted data focused on the applied machine learning model, steps in the development of the models, and model performance.
Results: From 3918 retrieved studies, 10 met our inclusion criteria. All followed a supervised machine learning approach and applied, from a limited range of options, a high-quality standard for the training of their model. The results show that machine learning can achieve a sensitivity of 95% while maintaining a high precision of 86%.
Conclusions: Machine learning approaches perform well in retrieving high-quality clinical studies. Performance may improve by applying more sophisticated approaches such as active learning and unsupervised machine learning approaches.
Background and Significance
Evidence-based medicine (EBM) is identified by three key elements: the best available clinical evidence, clinician expertise, and application of the evidence with consideration of patients’ circumstances, values, and preferences . EBM complements or reduces reliance on expert opinion with a coherent and structured framework for assessing and applying the best evidence to patient care decisions [ ]. An obvious and worsening barrier to the implementation of EBM is the continuously growing body of medical literature. According to the National Library of Medicine, over 900,000 new citations were indexed in MEDLINE in 2020, very few of which were relevant to or ready for clinical attention [ ]. Searching for the best clinical care evidence is a challenging task for researchers and clinicians, and facilitation of the search process is a necessity [ ].
Search filters, also referred to as hedges, allow researchers, clinicians, and librarians to retrieve evidence from bibliographic databases and journals by filtering searches to return reliable and specific articles to address clinical questions, produce systematic reviews, or inform clinical guidelines . MEDLINE search filters, for example, enable researchers to combine the use of free text with controlled vocabularies like Medical Subject Heading (MeSH) terms and other indexing features to improve search results targeting the clinical question at hand [ , ]. There are search filters that focus on the purpose of a study and its methods or topical content areas [ ]. Topical search filters help identify articles based on particular clinical conditions using terms related to that condition [ ], while methodological search filters comprise terms that identify articles based on their research purpose [ ]. For example, the Hedges project, developed by the Health Information Research Unit at McMaster University, provides search filters for MEDLINE, PsycINFO, and EMBASE using the OVID syntax for a range of purpose categories of articles such as treatment, diagnosis, and prognosis and include methodological terms [ , , ]. For searches seeking articles on a treatment (purpose), the search hedge includes methodological terms related to clinical or randomized controlled trials (RCTs), while the diagnosis search hedge includes methodological terms including sensitivity and specificity [ ].
These search filters were developed to identify high-quality studies based on established critical appraisal criteria for methodological rigor [- ]. This was done by annotating articles as meeting or not meeting criteria and using the annotated dataset to evaluate the performance of search terms to optimally retrieve the high-quality studies. For RCTs, applying the Cochrane risk for bias tool includes assessing randomization method, allocation concealment, follow-up data for at least 80% of participants, blinding of participants, and outcome assessors [ ]. For the Hedges project, the criteria applied to articles by purpose are available online [ ].
Clinical search filters are intended to help clinicians, researchers, and policymakers quickly access relevant studies and systematic reviews in a way that can be tailored to the user’s demand . The filters differ in their sensitivity and specificity according to the terms used, databases searched, and precision of the filter [ ]. Some filters offer high specificity, which limits the proportion of off-target articles that are retrieved. This is useful for busy clinicians who value the most efficient use of their time in finding relevant evidence quickly. Search filters may also have the option to maximize sensitivity and identify all potentially relevant articles at the cost of including a higher proportion of off-target articles [ ], an approach more suited to the conduct of systematic literature reviews.
Although search filters, such as Clinical Queries in PubMed, have been used since 1990 and have continued to work well over the years , they have some limitations. One limitation is their partial dependence on MeSH indexing terms, as the process of indexing of articles within MEDLINE can take up to a year for some articles [ ]. For diagnostic studies, there is large variability in designs and methods, which may result in largely incomplete literature searches [ ]. When applied in the context of conducting a systematic review, the highly specific filters result in missing evidence [ ], and the high sensitivity search filters will only partially reduce the time-consuming task of screening retrieved titles and abstracts [ ].
Overview of Machine Learning Applied for Text Processing
Machine learning is a subset of artificial intelligence that refers to a series of computational methods using experience to improve performance or achieve accurate and precise predictions. Experience, in this context, refers to the information made available to the machine for the analysis . A more detailed definition was provided by Mitchell [ ]: “A computer program is said to learn from experience (E) with respect to some class of tasks (T) and performance measure (P), if its performance at tasks in T, as measured by P, improves with experience E.”
Machine learning applications have become increasingly popular and essential in health care , as the system generates an enormous amount of data every day [ ]. Machine learning can identify relevant relations in large health care–generated datasets and derive algorithms that generate accurate predictions [ , ]. For example, machine learning has been used to predict the risk for nosocomial infection by leveraging data from electronic health records [ - ]. A machine learning classifier is a mathematical procedure responsible for identifying the patterns and performing the prediction task on the dataset, while a machine learning model is the output of the algorithm [ ]. A machine learning model represents the complete learning process including the training of the algorithm and the used set of features [ ].
Another application of machine learning in the health care and biomedical literature is text mining, which refers to the discovery of previously unknown information from unstructured textual data . This is done by converting the text to structured analyzable data using natural language processing (NLP) [ ]. With the exponential increase in the amount of information available for clinicians and researchers, both in biomedical literature and electronic health records [ ], text mining has been applied for text summarization [ ], literature retrieval [ ], and evidence grading [ ]. Machine learning has also been applied to automate the screening process for systematic reviews, identifying relevant articles while decreasing workload and increasing efficiency [ , , ]. Semantic analysis, the process of understanding text by interpreting meanings from the unstructured text [ ], has been applied to information extraction from the biomedical literature [ ].
There are several types of machine learning determined by their mathematical approach . The basic machine learning strategies are supervised learning, unsupervised learning, and reinforced learning [ , ]. Supervised learning relies on a prelabeled training dataset to provide the machine with the necessary input to make accurate predictions [ ]. Decision tree (DT), naïve Bayes (NB), and support vector machine (SVM) are common supervised machine learning algorithms [ ]. Unsupervised learning does not use labeled data and is mainly used for structuring and organizing data rather than classification [ ]. In reinforced learning, the algorithm learns by reacting to its environment and reaches predictions via a reward system [ ]. A common machine learning technique is ensemble learning, which combines more than one classifier to perform an individual prediction task. Boosting is one of the commonly used ensemble learners, which combines multiple weak classifiers and converts them into one strong classifier [ ]. Neural networks are multilayer mathematical structures consisting of an input layer, an output layer, and a hidden layer (commonly more than one layer) in between [ ]. In each layer a series of calculations occurs, leading to better performance [ ]. Due to the multilayer nature of neural networks, their field of study is known as deep learning. Neural networks can be supervised, unsupervised, or reinforced [ ].
Another appealing application of machine learning approaches to the biomedical literature is to improve retrieval of clinically relevant articles, building on and hopefully overcoming the limitation faced by Boolean searching. Several studies have been conducted to assess the performance of machine learning classifiers to identify specific categories of published articles. For example, Marshall and colleagues  applied machine learning to identify RCTs. Del Fiol and colleagues [ ] used machine learning to extract only scientifically sound treatment studies from PubMed. However, no systematic review of studies objectively assessing the performance of such machine learning models, ideally comparing their performance to traditional evidence retrieval methods such as validated Boolean search filters or manual critical appraisal by experts in the field, has been performed to date. Such a systematic review would be of critical value in driving future machine learning research aimed at improving the delivery of relevant evidence to the point of care.
The objective of this systematic review is to summarize the nature (methods and approaches) and comparative performance (eg, recall and precision) of machine learning approaches that have been applied to retrieve high-quality evidence for clinical consideration from the biomedical literature. High-quality is defined as articles that meet established methodological critical appraisal criteria, with annotated datasets that apply these criteria considered the gold standard.
The following subsections describe in detail the steps that were conducted to identify, screen, and abstract data from the included studies.
Nine databases were searched from inception to July 8, 2020, to identify relevant articles: Web of Science (title, abstract); MEDLINE; Embase; PsychINFO (title, abstract, keyword, subject terms); Wiley Online Library; ScienceDirect (title, abstract, keyword); CINAHL; IEEE (title, abstract, keywords), and Association of Computer Machinery digital library (title, abstract). The Multidisciplinary Digital Publishing Institute (title, abstract) database was searched on November 17, 2020. The search strategy was developed with a librarian (TN). Search terms related to 4 concepts—machine learning, literature retrieval, high research quality, and biomedical literature—were combined using the AND Boolean operator. The OVID MEDLINE search included the following terms, which were translated for the other databases (mp = multipurpose, searching within the title, original title, abstract, subject heading, name of substance, and registry word fields):
- Machine learning: (neural networks/ or machine learning/ or natural language processing/ or data mining/ or support vector machine/ or (“text categorization” or “text classification” or “text analysis” or “literature mining” or “text mining”).mp)
- Study objective or goal: (“Abstracting and Indexing”/ or “information storage and retrieval”/ or (“article retrieval” or “literature surveillance” or “literature screening” or “article screening” or “evidence search” or “evidence screening” or “evidence review” or “information retrieval” or “literature survey” or “document classification” or “review efficiency” or “citation screening” or “literature databases”).mp.)
- High-quality: (“Sensitivity and Specificity”/ or evidence-based medicine/ or (“quality” or “evidence” or “high-quality” or “clinical trial” or “random*” or “randomized controlled trial” or “sensitivity or specificity” or “accuracy” or “precision”).mp.)
In the Association of Computer Machinery digital library and Multidisciplinary Digital Publishing Institute search queries, terms related to the biomedical literature were included: (“PubMed” or “MEDLINE” or “medical literature” or “Biomedical literature”).
Articles retrieved by our search queries were collected in a single Research Information Systems file using JabRef software. Deduplication was conducted using both JabRef automatic deduping and Covidence automatic deduplication. We included articles that met the following criteria:
- Reported on the use of a machine learning approach for the retrieval of single studies or systematic reviews concerning the management of health care problems in large biomedical bibliographic databases such as MEDLINE and EMBASE
- Classified retrieved articles based on quality (using a gold standard)
- Used a textual analysis machine learning approach
- Evaluated the performance of the machine learning approach (ie, they present a comparison of retrieval methods or other ways of appraising the performance of the machine learning approach)
- Conducted within the biomedical literature domain
- Published in the English language
Abstract and Full-Text Screening
Titles and abstracts of all the retrieved articles were screened independently in Covidence.org by two members of the study team. Articles were assessed as relevant, irrelevant, or maybe relevant. The full texts of relevant and maybe relevant articles were then reviewed in duplicate, with conflicts adjudicated by a third team member.
A data extraction spreadsheet was developed to gather data regarding the methods of the machine learning approaches as detailed by the survey by Agarwal and Mittal  and included details on preprocessing steps, text representation, feature selection, feature extraction, and classifiers used. Additionally, we extracted data specific to the retrieval of high-quality articles such as the quality gold standard, the comparators used to test the machine learning models, and the performance of the developed algorithms.
Our search queries retrieved 3918 articles after 472 duplicates were removed; 3632 were excluded during the title and abstract screening for not applying a machine learning approach to biomedical articles. A total of 286 were selected for full-text screening, and 10 articles met our eligibility criteria () [ ]. Due to the heterogeneity in the population (retrieved articles), index method (machine learning algorithm used), gold standard, and outcomes (definition of high-quality study), we did not perform a quantitative synthesis of the results.
Quality Gold Standard
Each study used a quality gold standard database of original studies or systematic reviews that were manually reviewed and annotated by experts based on their scientific soundness and clinical relevance (). Datasets of articles that met or did not meet standards for quality and relevance were used to train the machine learning models. Four studies used the American College of Physicians (ACP) Journal Club as their quality gold standard [ - ], 3 studies used the Clinical Hedges dataset [ , , , ], 2 studies considered articles that were included in treatment clinical guidelines as high quality [ , ], and 1 article used the Cochrane Library as their gold standard [ ].
|Author||Quality gold standard|
|Aphinyanaphongs et al ||ACPa Journal Club (treatment class)b|
|Aphinyanaphongs et al ||ACP Journal Club (treatment, diagnosis, etiology, prognosis)b|
|Aphinyanaphongs et al ||ACP Journal Club (treatment, diagnosis, etiology, prognosis)b|
|Kilicoglu et al ||Clinical Hedgesb|
|Lin et al ||ACP Journal Club (unspecified classes of articles)b|
|Afzal et al ||Clinical Hedgesb|
|Bian et al ||Articles cited in 11 clinical guidelines on the treatment of cardiac, autoimmune, and respiratory diseases|
|Del Fiol et al ||Clinical Hedgesb|
|Bian et al ||Articles cited in 11 clinical guidelines on the treatment of cardiac, autoimmune, and respiratory diseases|
|Afzal et al ||Cochrane Library Reviews|
aACP: American College of Physicians.
bHand searches of articles from approximately 125 clinical journals that were assessed by critical appraisal criteria; articles meeting criteria were then judged by clinicians for clinical relevance. ACP Journal Club includes additional reviews by clinicians.
A matrix of the preprocessing steps that were applied to the dataset before developing the classifiers as reported in the included studies is presented in. Seven of the included studies provided details of their preprocessing steps [ , , - , , ], which included the conversion of text to lowercase, word-stemming, and removal of stop words. Additionally, 6 studies applied a term weighting method [ , - , , ] to express the importance of a word in each document based on its frequency. Afzal et al [ ] used vocabulary pruning by removing off topic-specific frequent terms and rarely occurring terms. Three studies did not specify the steps for their preprocessing steps [ , , ].
|Author||Text converted to lowercase||Removal of punctuation||Removal of stop words||Porter- stemming||Weighting method||Unique preprocessing considered|
|Aphinyanaphongs et al ||✓a||✓||✓||✓||Log frequency with redundancy||NRb|
|Aphinyanaphongs et al ||✓||✓||✓||✓||Log frequency with redundancy||NR|
|Aphinyanaphongs et al ||✓||✓||✓||✓||Log frequency with redundancy||Removed infrequent words|
|Kilicoglu et al ||✓||NR||✓||✓||Information gain measure||Removed infrequent words|
|Lin et al ||NR||NR||NR||NR||NR||NR|
|Afzal et al ||✓||NR||✓||✓||TF-IDFc||Vocabulary pruning|
|Bian et al ||NR||NR||NR||NR||NR||NR|
|Del Fiol et al ||✓||NR||✓||NR||NR||Removed articles without abstracts, concatenated title, and abstract words|
|Bian et al ||NR||NR||NR||NR||NR||NR|
|Afzal et al ||✓||NR||NR||NR||TF-IDF||Removed articles with missing values|
bNR: not reported.
cTF-IDF: term frequency–inverse document frequency.
Most of the included articles relied on the text as their features (). Seven articles used words from titles and abstracts as their features [ , , - , , ]. Kilicoglu et al [ ] and Afzal et al [ ] used article metadata features, Unified Medical Language System features, SemRep semantic prediction, and MeSH terms in combination with the words of titles and abstracts features. Lin et al [ ] selected specific features from the citation dataset: journal impact factor, MeSH terms, sample size, P value, and confidence intervals. Bian et al [ , ] relied on MEDLINE metadata as well as bibliometric features, which included citation count, journal impact factor, number of comments on PubMed, Altmetric score, study sample size, registration in ClinicalTrials.gov, and article age, and assessed how each feature contributed to the classification. The experiment by Bian et al [ ] used only time-agnostic features (features available at the time of an article’s publication), which are journal impact factor, sample size, number of grants, number of authors, number of clinically useful sentences, scientific impact of authors’ institution, numbers of references, page count, registration in ClinicalTrials.gov, and publication in PubMed Central. Afzal et al [ ] used automatic feature engineering with RapidMiner software for the title and abstract text feature extraction as part of the multilayer perceptron model.
Machine Learning Classifier
The majority of the included studies developed multiple algorithms and selected the top-performing one for their main classification tasks (). Aphinyanaphongs et al [ , ], initially reported their results using SVM, NB, and boosting algorithms in both their 2003 and 2005 experiments; however, they ended up selecting SVM as their top-performing classifier in a separate study [ ]. Bian et al [ , ] and Afzal et al [ ] compared the performance of multiple classifiers (SVM, NB, DT, k-nearest neighbors, random forest, multilayer perceptron) and selected the best performing for their experiment in the context of the same study (NB, DT, and SVM, respectively). We refer to the classifier that was selected for the classification task as the main classifier.
From the included articles, SVM was the most used classifier. Five studies used an SVM algorithm as one of their main experiment classifiers (), 2 studies used a neural network as their main classifier; Del Fiol et al [ ] used a convolutional neural network (CNN), while Afzal et al [ ] used a multilayer feed-forward artificial neural network (ANN). DT algorithms were used in 2 studies for their main text classification function [ , ]. Four of the included studies applied multiple classifying approaches [ , , , ].
|Author||Naïve Bayes||SVMa||Decision tree||Ensemble||Neural network|
|Aphinyanaphongs et al ||✓b||✓||N/Ac||✓||N/A||N/A|
|Aphinyanaphongs et al ||✓||✓||N/A||✓||N/A||N/A|
|Aphinyanaphongs et al ||N/A||✓||N/A||N/A||N/A||N/A|
|Kilicoglu et al ||✓||✓||N/A||✓||✓||N/A|
|Lin et al ||N/A||N/A||✓||N/A||N/A||N/A|
|Afzal et al ||N/A||✓||N/A||N/A||N/A||N/A|
|Bian et al ||✓||N/A||N/A||N/A||N/A||N/A|
|Del Fiol et al ||N/A||N/A||N/A||N/A||N/A||✓|
|Bian et al ||N/A||N/A||✓||N/A||N/A||N/A|
|Afzal et al ||N/A||N/A||N/A||N/A||N/A||✓|
aSVM: support vector machine.
Comparator for Evaluating the Performance of the Classifiers
As per our inclusion criteria, to evaluate the performance of the machine learning method to classify articles appropriately, articles had to report a comparison of their applied machine learning model to a gold standard method such as gold standard high-quality articles retrieval method, for example, search filters, a manually annotated high-quality articles’ dataset, or a baseline machine learning model for high-quality articles retrieval (). Aphinyanaphongs et al [ - ] used Clinical Query filters with sensitivity and specificity optimization [ ]. The experiment conducted by Kilicoglu et al [ ] evaluated their machine learning approach by applying it in a new dataset annotated by experts. The NB high-quality algorithm by Kilicoglu et al [ ] was considered a comparator on its own for its high recall and was used as such by Bian and colleagues [ , ], who also used PubMed’s best match as a comparator. Lin et al [ ] used accuracy and k-value performance metrics in comparison to the results of the critical appraisal process by experts in the field. Also, Lin et al [ ] has applied a comparison between their classifier, which was a DT, to other known text classifiers like SVM and ANN. Afzal et al [ ] have used a SVM model for quality articles retrieval and compared its performance to the SVM model proposed by Sarker et al [ ], reporting that their classifier achieved a higher performance with their reported features selected.
Del Fiol et al  was the first study to incorporate the use of deep learning in quality articles retrieval, relying on a CNN. Del Fiol and colleagues [ ] compared their proposed classifier to the PubMed Clinical Queries broad filter since it achieves a nearly perfect recall. Also, they compared their proposed model to McMaster textword search and McMaster balanced search filter created by the Clinical Hedges group to evaluate the capabilities of their model of retrieving recently published evidence and achieving a balance between recall and precision [ ]. Afzal et al [ ], in their experiment using ANN, compared their model’s results to the CNN results of Del Fiol et al [ ], the DT results of Bian et al [ ], and their prior experiment using an SVM for quality articles retrieval [ , ]. Also, Afzal et al [ ] compared their proposed ANN to well-known algorithms used in the literature like NB, SVM, DT, and gradient boosted trees.
|Aphinyanaphongs et al [- ]|
|Kilicoglu et al |
|Lin et al |
|Afzal et al |
|Bian et al |
|Del Fiol et al |
|Bian et al |
|Afzal et al |
aSVM: support vector machine.
bNB: naïve Bayes.
cDT: decision tree.
dGBT: gradient boosted trees.
All included articles applied a supervised machine learning model. Validation by applying a resampling k-fold approach was used in 7 studies. Five used 10-fold cross-validation [, , , , ], and 2 studies relied on 5-fold cross-validation [ , ]. The most common performance metrics used in the included studies were sensitivity (recall), specificity, accuracy, area under the curve (AUC), F-measure, and precision ( ). The recall was generally high, above 85%, across all experiment classifiers except the SVM by Kilicoglu et al [ ], and the NB and DT reported by Bian et al [ ] and Bian et al [ ], respectively, as both had a recall below 30%. Precision ranged from 9% to 86%, with the neural network of Afzal et al [ ] and the SVM by Kilicoglu et al [ ] the highest. AUC was measured in all studies and ranged from 0.73 to 0.99. Lin et al [ ] and Bian et al [ , ] used novel performance metrics in their approaches. In the 2 studies by Bian and colleagues [ , ], performance was primarily determined by calculating the top 20 precision which is the measure of the percentage of true positive citations among the first 20 retrieved citations. Lin et al [ ] used Cohen kappa (k-value) as their performance metric, which is the agreement between machine performance (observed value) and gold standard (expected value) [ , ]. Bian et al [ , ] reported a top 20 precision of 34% with their 2017 NB classifier and 24% in their 2019 experiment using a DT classifier. Lin et al [ ] reported a k-value of 0.78 in their experiment.
|Classifier and author||Recalla||Specificityb||Precisionc||F-scored||AUCe||Accuracyf|
|Support vector machine|
|Aphinyanaphongs et al ||0.967||0.87||0.169||0.29g||0.98||0.893|
|Aphinyanaphongs et al ||0.96||0.86||0.18||0.30g||0.97||NRh|
|Aphinyanaphongs et al ||0.98||0.88||0.305||0.47g||0.95||NR|
|Kilicoglu et al ||0.229||NR||0.865||0.36||0.96||NR|
|Afzal et al ||NR||NR||NR||0.87||0.73||0.785|
|Aphinyanaphongs et al ||0.967||0.76||0.091||0.17g||0.95||0.787|
|Aphinyanaphongs et al ||NR||NR||NR||NR||0.95||NR|
|Kilicoglu et al ||0.975||NR||0.138||0.24||0.82||NR|
|Bian et al ||0.23||NR||0.33||0.21||NR||NR|
|Aphinyanaphongs et al ||0.967||0.786||0.099||0.18g||0.96||0.804|
|Aphinyanaphongs et al ||NR||NR||NR||NR||0.94||NR|
|Kilicoglu et al ||0.729||NR||0.823||0.77||0.97||NR|
|Del Fiol et al ||0.969||NR||0.346||0.51||NR||NR|
|Afzal et al ||0.951||NR||0.863||0.9||0.99||0.973|
|Lin et al ||NR||NR||NR||NR||NR||0.854|
|Bian et al ||0.09||NR||0.39||0.14||NR||NR|
|Kilicoglu et al ||0.864||NR||0.747||0.801||0.98||NR|
aRecall: proportion of correctly identified positives among the real positive.
bSpecificity: the proportion of actual negatives, which got predicted as the negative (or true negative).
cPrecision: proportion of correctly identified positives among all classified positives.
dF-score: harmonic mean of the precision and recall. F-score is equivalent to F1-score and used interchangeably.
eAUC: area under the curve traced out by graphing the true positive rate against the false positive rate. The higher the AUC, the better the classifier prediction.
fAccuracy: number of correctly predicted documents out of all classified documents.
gCalculated as F-measure=(2*precision*recall)/(precision+recall) using recall and precision when available from the articles.
hNR: not reported.
To our knowledge, this is the first systematic review of machine learning approaches used to classify scientifically sound and clinically relevant studies from the biomedical literature. All included studies followed a supervised machine learning technique in which the learning algorithm depends on prelabeled data provided for training . Despite the technological advancements from 2003 to 2020 when the studies were published, none reported applying unsupervised or active learning approaches for the classification of articles based on quality. Active learning is a subtype of machine learning in which the learning algorithm is allowed to select the data from which it learns by querying a human operator and can achieve a performance comparable to the standard supervised learning algorithms with fewer labeled data [ ]. For example, active learning was used in the recent work by Gates et al [ ] and Tsou et al [ ], who used Abstrackr, a freely available active machine learning tool that automates the screening of titles and abstracts [ ]. Abstrackr achieved 100% sensitivity after screening only 31.8% of the citations in the dataset [ ].
There is a limited range of quality standards comprising the prelabeled training datasets across the included articles. ACP Journal Club and the Clinical Hedges follow the same inclusion and exclusion criteria for high-quality evidence . Aphinyanaphongs et al [ ] considered an article as high-quality if it were included in ACP Journal Club but considered only those classified as treatment, which limits their results to RCTs. The authors expanded their inclusion to articles tagged as treatment, diagnosis, prognosis, and etiology in their subsequent studies [ , ]. Having consistency across gold standard databases in classifier development strengthens our ability to compare performance. There are, however, limited manually annotated datasets available as these are time consuming and expensive to develop and require consistency and highly skilled people. Using studies that are included in guidelines and systematic reviews, as done by Bian et al [ , ] and Afzal et al [ ], leverages screening work that has already been completed to a high standard; however, citations in guidelines may include lower quality evidence in the training process [ ].
The limited availability of high-quality dataset options was highlighted by Afzal et al , and finding the ideal gold standard training dataset was the most reported limitation in the included studies. In our opinion, the ideal gold standard training dataset should cover some criteria to overcome the limitations reported in the articles. First, the gold standard should be defined by precise criteria for methodological rigor that is created and recognized by experts in the field [ ]. Selection criteria for the gold standard should be unbiased. Aphinyanaphongs et al [ ] described their concern toward the possibility of a selection bias by the ACP Journal Club editors in a particular year toward a certain topic. Second, the gold standard training dataset should cover a large enough sample of the high-quality class to properly train the model and overcome the class imbalance bias toward the majority class of studies that are not of high quality [ , ]. Third, the gold standard training set should cover multiple health care domains, as Lin et al [ ] reported their high-quality dataset was limited only to cardiovascular diseases and would not perform as well if applied to another medical domain. Fourth, the gold standard training dataset should be up to date as much as possible, which was a limitation reported in both studies by Bian et al [ ] and Afzal et al [ ].
Another possible constraint affecting accurate prediction is the feature selection process. Del Fiol et al  stated that using MeSH-based features instead of the sole reliance on text features in their experiment could have improved the precision of their neural network. In consensus with the recommendation of Del Fiol et al [ ], some of the included studies provided evidence that the use of a combination of features improves the overall performance of the classifiers. For example, in the experiment by Afzal et al [ ], the combination of publication type and MeSH term features in addition to title and abstract features produced the best and the most stable results. Also, Kilicoglu et al [ ] proved that the incorporation of MEDLINE citation metadata and Unified Medical Language System features in addition to words of titles and abstracts yielded the best performance. Such important features may not be immediately available at the time of indexing in MEDLINE [ ], which poses a challenge in identifying recently published evidence [ , ].
There was a higher rate of incorporating SVM algorithms in the experiments by the study authors. SVMs are known for their high accuracy  and their low classification error [ ], making them ideal for linear classification. Afzal et al [ ] developed an ANN algorithm that had higher accuracy when compared with their previous SVM classifier [ ]. Further applications of newer machine learning approaches will advance the knowledge base on these quickly evolving methods. While SVMs currently have good accuracy and low error rates, emerging approaches may well outperform them.
The main purpose of using machine learning in the classification of high-quality articles is to decrease the workload on those performing manual classification without losing relevant articles in the process. Recall, the proportion of correctly identified high-quality articles from the high-quality pool, is the most important metric to be used, followed by precision, the proportion of correctly identified positive articles among all those classified as positive. The included studies reported a range of recall and precision some of which would not meet the objective of identifying the high-quality articles correctly. For example, the NB classifier developed by Bian et al  performed significantly less than the NB by Kilicoglu et al [ ] and PubMed Best Match in terms of recall (23% vs 55% and 65%, respectively). Despite performing worse in recall, their classifier achieved a higher precision (33% vs 5% and 4%) [ ].
Additionally, accuracy, the number of correctly predicted documents out of all classified documents, is considered a common metric for evaluating classifiers; however, its use is considered inappropriate to evaluate imbalanced dataset classification . For example, a classifier labeling all entries as false (given that false is the majority class) would have high accuracy but would fail to perform the needed task of accurately classifying the passing articles (rare class), making it useless [ ]. The harmonic mean of the recall and precision measurements is the F-score, and it is used to evaluate the machine learning algorithms implemented on unbalanced datasets [ ]. F-score was first used in the study by Kilicoglu et al [ ] where the performance of the classifiers was reported using recall, precision, F-score, and AUC, without including accuracy. Additionally, Afzal et al [ ] did not rely on recall to compare between multiple classifiers; instead, they used the F-score, precision, and accuracy. Also, they have applied a novel approach to compare between the classifiers, in which they summed the metrics for a classifier with a higher sum reflecting better performance [ ].
The highest reported recall in our review was 98% with the SVM developed by Aphinyanaphongs and Aliferis , however, the algorithm had low precision of 30.5%. The best balance between recall and precision was achieved by the ANN approach used by Afzal et al [ ], which reported a high recall of 95.1% and a high precision of 86.3%, thereby achieving the target of not losing quality literature while decreasing the manual classification workload.
The experiment by Kilicoglu et al  assesses the effect of applying 3 different machine learning classifiers (SVM, NB, boosting, and ensemble) trained using the same Clinical Hedges dataset on the overall performance of the resulting models. Using multiple feature set combinations, the highest recall was achieved by the NB classifier, and the highest F-scores were achieved by ensemble (0.80) and text-boosting (0.77) based models [ ]. Only the studies by Aphinyanaphongs and colleagues [ , ] and Kilicoglu et al [ ] incorporated ensemble techniques in the development of their main classifiers, and their results suggest that using multiple classifiers in combination can improve the balance between recall and precision (the F-score).
Strengths and Limitations
This is the first systematic review to characterize the machine learning approaches in high-quality article retrieval. When narrowing our research question, we excluded other text summarization and text categorization approaches being used in the biomedical literature. These include but are not limited to studies concerned with the automation of the systematic review process [, ], biomedical literature summarization [ ], and semantic models’ applications in the biomedical literature [ ]. Given the technical nature of the application of machine learning approaches for text classification, we expanded our search beyond clinical bibliographic databases to include those which index technical articles.
Across the included studies, some steps were not fully reported in the methods, including preprocessing steps, cross-validation folds, and features selected. To our knowledge, there are no reporting guidelines for machine learning approaches being applied for literature retrieval. The Equator Network includes 6 reporting guidelines for machine learning approaches; however, all 6 are focused on articles applying machine learning in clinical settings . For example, the most recently published guideline focuses on the reporting of interventions involving artificial intelligence in clinical trial protocols [ ]. The lack of reporting guidance for the NLP component of machine learning being applied in the biomedical literature creates a noticeable gap in reporting the steps of the applied approach, features used and justification for their use, and inconsistency in the reported performance achieved by the machine. As a result, there was a lack of consistency in the reporting of results and methods provided by the authors, which also limits our ability to compare the performance of the classifiers. Also, one of the limitations developing the review was the inability to directly compare the performance of the models across the included studies because of the different training datasets and the applied settings. Finally, a challenge with machine learning is that the algorithms are considered as being derived in a black box; an enigmatic interpretation that the machines provide findings and predictions without any accompanying explanation [ ].
Despite the longevity of research for the identification of high-quality literature using machine learning, evidence is still scarce and slowly progressing over time, and determining the most reliable approach is difficult as the field is quickly evolving. This slow progression in the field may have been caused by the lack of publicly available standard benchmarks for the identification of high-quality articles biomedical literature to compare the performance of the proposed methods. A similar problem was addressed in the molecular machine learning domain by creating MolecularNet, a large-scale, open-source, and high-quality benchmark for molecular learning algorithms . Our review provides a summary of current approaches and performance of machine learning models applied to retrieve high-quality evidence for clinical consideration from the biomedical literature and highlights the importance of selecting optimal quality gold standard data for training. The findings include that the use of different feature sets in combination with text features is likely to improve the performance of machine learning models. There is a lack of reporting consistency in the literature which makes replication of the experiments difficult. Supervised machine learning has been the focus to date. The rapid development in the field of NLP and the availability of new state of the art techniques such as Bidirectional Encoder Representations from Transformers (BERT) for language understanding [ ] and bio-BERT for biomedical text mining [ ] hold promise for future advances in the field of information extraction from the biomedical literature. Considering the increasingly available data to apply these approaches to, we anticipate that the performance of classifiers to identify high-quality evidence will continue to grow.
We thank Rita Jezrawi and Hamza Issa for assistance screening articles for inclusion. We thank Dr Mehrdad Roham for proofreading the final version of the manuscript.
All authors contributed to the design of the study. TN and WA developed the search strategies. WA ran the searches and led the screening and data abstraction. All authors contributed to the interpretation of the data. WA and CL drafted early versions of the manuscript. All authors supervised the study and reviewed and provided revisions to the manuscript. All authors approved the final manuscript.
Conflicts of Interest
Features selected by the included articles.PDF File (Adobe PDF File), 187 KB
- Guyatt G, Jaeschke R, Wilson M, Montori V, Richardson W. What Is evidence-based medicine? In: Guyatt G, Rennie D, Meade MO, Cook DJ, editors. Users’ Guides to the Medical Literature: A Manual for Evidence-Based Clinical Practice, 3rd Edition. New York: McGraw-Hill Education; 2015:7-14.
- Kamath S, Guyatt G. Importance of evidence-based medicine on research and practice. Indian J Anaesth 2016 Sep;60(9):622-625 [FREE Full text] [CrossRef] [Medline]
- MEDLINE PubMed Production Statistics. URL: https://www.nlm.nih.gov/bsd/medline_pubmed_production_stats.html [accessed 2021-08-23]
- Wilczynski NL, Morgan D, Haynes RB, Hedges Team. An overview of the design and methods for retrieving high-quality studies for clinical care. BMC Med Inform Decis Mak 2005 Jun 21;5:20 [FREE Full text] [CrossRef] [Medline]
- Beale S, Duffy S, Glanville J, Lefebvre C, Wright D, McCool R, et al. Choosing and using methodological search filters: searchers' views. Health Info Libr J 2014 Apr 23;31(2):133-147. [CrossRef]
- Wilczynski NL, Haynes RB, Hedges Team. Developing optimal search strategies for detecting clinically sound causation studies in MEDLINE. AMIA Annu Symp Proc 2003:719-723 [FREE Full text] [Medline]
- Leeflang M, Scholten R, Rutjes A, Reitsma J, Bossuyt P. Use of methodological search filters to identify diagnostic accuracy studies can lead to the omission of relevant studies. J Clin Epidemiol 2006 Mar;59(3):234-240. [CrossRef] [Medline]
- Damarell RA, May N, Hammond S, Sladek RM, Tieman JJ. Topic search filters: a systematic scoping review. Health Info Libr J 2019 Mar 21;36(1):4-40. [CrossRef] [Medline]
- McKibbon K, Wilczynski NL, Haynes R, Hedges Team. Retrieving randomized controlled trials from medline: a comparison of 38 published search filters. Health Info Libr J 2009 Sep;26(3):187-202 [FREE Full text] [CrossRef] [Medline]
- Miller PA, McKibbon KA, Haynes RB. A quantitative analysis of research publications in physical therapy journals. Phys Ther 2003 Feb;83(2):123-131. [Medline]
- Search filters for MEDLINE in Ovid Syntax and the PubMed translation. Health Information Research Unit, McMaster University. URL: https://hiru.mcmaster.ca/hiru/HIRU_Hedges_MEDLINE_Strategies.aspx [accessed 2021-01-31]
- Hedges. Health Information Research Unit, McMaster University. URL: https://hiru.mcmaster.ca/hiru/HIRU_Hedges_home.aspx [accessed 2021-05-31]
- Jadad AR, Moore RA, Carroll D, Jenkinson C, Reynolds DJ, Gavaghan DJ, et al. Assessing the quality of reports of randomized clinical trials: is blinding necessary? Control Clin Trials 1996 Feb;17(1):1-12. [CrossRef] [Medline]
- Higgins JPT, Altman DG, Gøtzsche PC, Jüni P, Moher D, Oxman AD, Cochrane Bias Methods Group, Cochrane Statistical Methods Group. The Cochrane Collaboration's tool for assessing risk of bias in randomised trials. BMJ 2011 Oct 18;343(oct18 2):d5928-d5928 [FREE Full text] [CrossRef] [Medline]
- Inclusion Criteria. Health Information Research Unit, McMaster University. 2019. URL: https://hiru.mcmaster.ca/hiru/InclusionCriteria.html [accessed 2021-05-31]
- Wong SS, Wilczynski NL, Haynes RB, Hedges Team. Developing optimal search strategies for detecting clinically relevant qualitative studies in MEDLINE. Stud Health Technol Inform 2004;107(Pt 1):311-316. [Medline]
- Gill PJ, Roberts NW, Wang KY, Heneghan C. Development of a search filter for identifying studies completed in primary care. Fam Pract 2014 Dec 18;31(6):739-745. [CrossRef] [Medline]
- Wilczynski NL, McKibbon KA, Walter SD, Garg AX, Haynes RB. MEDLINE clinical queries are robust when searching in recent publishing years. J Am Med Inform Assoc 2013;20(2):363-368 [FREE Full text] [CrossRef] [Medline]
- Irwin AN, Rackham D. Comparison of the time-to-indexing in PubMed between biomedical journals according to impact factor, discipline, and focus. Res Social Adm Pharm 2017;13(2):389-393. [CrossRef] [Medline]
- Tsafnat G, Glasziou P, Karystianis G, Coiera E. Automated screening of research studies for systematic reviews using study characteristics. Syst Rev 2018 Apr 25;7(1):64 [FREE Full text] [CrossRef] [Medline]
- Mohri M, Rostamizadeh A, Talwalkar A. In: Bach F, editor. Foundations of Machine Learning, Second edition. Cambridge: MIT Press; 2018.
- Mitchell T. Machine Learning, 1st ed. New York: McGraw-Hill Science; 1997:432.
- Koh HC, Tan G. Data mining applications in healthcare. J Healthc Inf Manag 2005;19(2):64-72. [Medline]
- Bahri S, Zoghlami N, Abed M, Tavares JMRS. Big data for healthcare: a survey. IEEE Access 2019;7:7397-7408. [CrossRef]
- Beam AL, Kohane IS. Big data and machine learning in health care. JAMA 2018 Apr 03;319(13):1317-1318. [CrossRef] [Medline]
- Wiens J, Shenoy ES. Machine learning for healthcare: on the verge of a major shift in healthcare epidemiology. Clin Infect Dis 2018 Jan 06;66(1):149-153 [FREE Full text] [CrossRef] [Medline]
- Wiens J, Guttag J, Horvitz E. Patient risk stratification with time-varying parameters: a multitask learning approach. J Mach Learning Res 2016;17:1-23 [FREE Full text] [CrossRef]
- Wiens J, Guttag J, Horvitz E. A study in transfer learning: leveraging data from multiple hospitals to enhance hospital-specific predictions. J Am Med Inform Assoc 2014;21(4):699-706 [FREE Full text] [CrossRef] [Medline]
- Wiens J, Horvitz E, Guttag J. Patient risk stratification for hospital-associated C. diff as a time-series classification task. In: Pereira F, Burges C, Bottou L, Weinberger K, editors. Advances in Neural Information Processing Systems. Red Hook: Curran Associates, Inc; 2012:467-475.
- Burkov A. The Hundred-Page Machine Learning Book, 1st ed. Quebec City: Andriy Burkov; 2019.
- Hearst M. Text Data Mining. In: Mitkov R, editor. The Oxford Handbook of Computational Linguistics (1st ed). Oxford: Oxford University Press; 2012.
- Gong L. Application of biomedical text mining. In: Artificial Intelligence—Emerging Trends and Applications InTech. London: IntechOpen; 2018.
- Davidoff F, Miglus J. Delivering clinical evidence where it's needed: building an information system worthy of the profession. JAMA 2011 May 11;305(18):1906-1907. [CrossRef] [Medline]
- Mishra R, Bian J, Fiszman M, Weir CR, Jonnalagadda S, Mostafa J, et al. Text summarization in the biomedical domain: a systematic review of recent research. J Biomed Inform 2014 Dec;52:457-467 [FREE Full text] [CrossRef] [Medline]
- Del Fiol G, Michelson M, Iorio A, Cotoi C, Haynes RB. A deep learning method to automatically identify reports of scientifically rigorous clinical research from the biomedical literature: comparative analytic study. J Med Internet Res 2018 Jun 25;20(6):e10281 [FREE Full text] [CrossRef] [Medline]
- Afzal M, Hussain M, Haynes RB, Lee S. Context-aware grading of quality evidences for evidence-based decision-making. Health Informatics J 2019 Jun;25(2):429-445 [FREE Full text] [CrossRef] [Medline]
- Wallace BC, Small K, Brodley CE, Lau J, Schmid CH, Bertram L, et al. Toward modernizing the systematic review pipeline in genetics: efficient updating via data mining. Genet Med 2012 Jul;14(7):663-669 [FREE Full text] [CrossRef] [Medline]
- Wallace B, Small K, Brodley C, Lau J, Trikalinos T. Deploying an interactive machine learning system in an evidence-based practice center: Abstrackr. Proc 2nd ACM SIGHIT Int Health Informat Symp 2012:819-824 [FREE Full text] [CrossRef]
- Rindflesch T, Fiszman M, Libbus B. Semantic interpretation for the biomedical research literature. Med Informatics 2006:399-422. [CrossRef]
- Holzinger A. Machine learning for health informatics. LNCS 2016;9605:1-24. [CrossRef]
- Dey A. Machine learning algorithms: a review. Int J Comput Sci Inf Technol 2016;7(3):1174-1179 [FREE Full text]
- Cohen S. The basics of machine learning: strategies and techniques. In: Artificial Intelligence and Deep Learning in Pathology. Philadelphia: Elsevier; 2021:13-40.
- Welling M. A First Encounter with Machine Learning. University of California Irvine. 2011. URL: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.441.6238&rep=rep1&type=pdf [accessed 2021-08-23]
- Wang S. Artificial neural network. In: Interdisciplinary Computing in Java Programming. Boston: Springer US; 2003:81-100.
- Hiregoudar SB, Manjunath K, Patil KS. A survey: research summary on neural networks. Int J Res Engineer Technol 2014 May 25;03(15):385-389. [CrossRef]
- Marshall IJ, Noel-Storr A, Kuiper J, Thomas J, Wallace BC. Machine learning for identifying randomized controlled trials: an evaluation and practitioner's guide. Res Synth Methods 2018 Dec;9(4):602-614 [FREE Full text] [CrossRef] [Medline]
- Agarwal B, Mittal N. Text classification using machine learning methods: a survey. 2014 Presented at: Proceedings of the Second International Conference on Soft Computing for Problem Solving; 2014; New Delhi p. 701-709. [CrossRef]
- Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021 Mar 29;372:n71 [FREE Full text] [CrossRef] [Medline]
- Aphinyanaphongs Y, Aliferis CF. Text categorization models for retrieval of high quality articles in internal medicine. AMIA Annu Symp Proc 2003:31-35 [FREE Full text] [Medline]
- Aphinyanaphongs Y, Tsamardinos I, Statnikov A, Hardin D, Aliferis CF. Text categorization models for high-quality article retrieval in internal medicine. J Am Med Inform Assoc 2005;12(2):207-216 [FREE Full text] [CrossRef] [Medline]
- Aphinyanaphongs Y, Aliferis C. Prospective validation of text categorization filters for identifying high-quality, content-specific articles in MEDLINE. AMIA Annu Symp Proc 2006:6-10 [FREE Full text] [Medline]
- Lin J, Chang C, Lin M, Ebell MH, Chiang J. Automating the process of critical appraisal and assessing the strength of evidence with information extraction technology. J Eval Clin Pract 2011 Aug;17(4):832-838. [CrossRef] [Medline]
- Kilicoglu H, Demner-Fushman D, Rindflesch TC, Wilczynski NL, Haynes RB. Towards automatic recognition of scientifically rigorous clinical research evidence. J Am Med Inform Assoc 2009;16(1):25-31 [FREE Full text] [CrossRef] [Medline]
- Bian J, Morid MA, Jonnalagadda S, Luo G, Del Fiol G. Automatic identification of high impact articles in PubMed to support clinical decision making. J Biomed Inform 2017 Sep;73:95-103 [FREE Full text] [CrossRef] [Medline]
- Bian J, Abdelrahman S, Shi J, Del Fiol G. Automatic identification of recent high impact clinical articles in PubMed to support clinical decision making using time-agnostic features. J Biomed Inform 2019 Jan;89:1-10 [FREE Full text] [CrossRef] [Medline]
- Afzal M, Park BJ, Hussain M, Lee S. Deep learning based biomedical literature classification using criteria of scientific rigor. Electronics (Switzerland) 2020 Aug 05;9(8):1-12. [CrossRef]
- PubMed Clinical Queries Search Filter. National Library of Medicine. URL: https://pubmed.ncbi.nlm.nih.gov/clinical/ [accessed 2021-05-01]
- Sarker A, Mollá D, Paris C. Automatic evidence quality prediction to support evidence-based decision making. Artif Intell Med 2015 Jun;64(2):89-103. [CrossRef] [Medline]
- Chiang J, Lin J, Yang C. Automated evaluation of electronic discharge notes to assess quality of care for cardiovascular diseases using Medical Language Extraction and Encoding System (MedLEE). J Am Med Inform Assoc 2010 May 01;17(3):245-252. [CrossRef]
- Rau G, Shih Y. Evaluation of Cohen's kappa and other measures of inter-rater agreement for genre analysis and other nominal data. J Engl Acad Purposes 2021 Sep;53:101026. [CrossRef]
- Gates A, Johnson C, Hartling L. Technology-assisted title and abstract screening for systematic reviews: a retrospective evaluation of the Abstrackr machine learning tool. Syst Rev 2018 Mar 12;7(1):45 [FREE Full text] [CrossRef] [Medline]
- Tsou AY, Treadwell JR, Erinoff E, Schoelles K. Machine learning for screening prioritization in systematic reviews: comparative performance of Abstrackr and EPPI-Reviewer. Syst Rev 2020 Apr 02;9(1):73 [FREE Full text] [CrossRef] [Medline]
- Wallace BC, Trikalinos TA, Lau J, Brodley C, Schmid CH. Semi-automated screening of biomedical citations for systematic reviews. BMC Bioinformatics 2010 Jan 26;11(1):55 [FREE Full text] [CrossRef] [Medline]
- Venus C, Jamrozik E. Evidence-poor medicine: just how evidence-based are Australian clinical practice guidelines? Intern Med J 2020 Jan 14;50(1):30-37. [CrossRef] [Medline]
- Lanera C, Berchialla P, Sharma A, Minto C, Gregori D, Baldi I. Screening PubMed abstracts: is class imbalance always a challenge to machine learning? Syst Rev 2019 Dec 06;8(1):317 [FREE Full text] [CrossRef] [Medline]
- Fernández-Delgado M, Cernadas E, Barro S, Amorim D. Do we need hundreds of classifiers to solve real world classification problems? J Mach Learn Res 2014;15(1):3181.
- Bekkar M, Djema H, Alitouche T. Evaluation measures for models assessment over imbalanced data sets. J Inf Engineer Applic 2013;3:27-38 [FREE Full text]
- Tang Y, Zhang YQ, Chawla N, Krasser S. SVMs modeling for highly imbalanced classification. IEEE Trans Syst Man Cybern B 2009 Feb;39(1):281-288. [CrossRef]
- O'Mara-Eves A, Thomas J, McNaught J, Miwa M, Ananiadou S. Using text mining for study identification in systematic reviews: a systematic review of current approaches. Syst Rev 2015 Jan 14;4:5 [FREE Full text] [CrossRef] [Medline]
- Shakeel Y, Krüger J, Nostitz-Wallwitz IV, Saake G, Leich T. Automated selection and quality assessment of primary studies. J Data Inf Qual 2020 Jan 23;12(1):1-26. [CrossRef]
- Yoo I, Hu X, Song I. A coherent graph-based semantic clustering and summarization approach for biomedical literature and a new summarization evaluation method. BMC Bioinformatics 2007 Nov 27;8 Suppl 9:S4. [CrossRef] [Medline]
- Arguello Casteleiro M, Maseda Fernandez D, Demetriou G, Read W, Fernandez Prieto MJ, Des Diz J, et al. A case study on sepsis using PubMed and deep learning for ontology learning. Stud Health Technol Inform 2017;235:516-520. [Medline]
- EQUATOR Network: Reporting Guidelines. URL: https://www.equator-network.org/reporting-guidelines/ [accessed 2021-08-24]
- Rivera SC, Liu X, Chan A, Denniston AK, Calvert MJ, SPIRIT-AI and CONSORT-AI Working Group. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI Extension. BMJ 2020 Sep 09;370:m3210 [FREE Full text] [CrossRef] [Medline]
- Watson DS, Krutzinna J, Bruce IN, Griffiths CE, McInnes IB, Barnes MR, et al. Clinical applications of machine learning algorithms: beyond the black box. BMJ 2019 Mar 12;364:l886. [CrossRef] [Medline]
- Wu Z, Ramsundar B, Feinberg E, Gomes J, Geniesse C, Pappu AS, et al. MoleculeNet: a benchmark for molecular machine learning. Chem Sci 2018 Jan 14;9(2):513-530 [FREE Full text] [CrossRef] [Medline]
- Devlin J, Chang M, Lee K, Toutanova K. BERT: pre-training of deep bidirectional transformers for language understanding. ArXiv. Preprint posted online on October 10, 2018. [FREE Full text]
- Lee J, Yoon W, Kim S, Kim D, Kim S, So C, et al. BioBERT: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics 2020 Feb 15;36(4):1234-1240 [FREE Full text] [CrossRef] [Medline]
|AUC: area under the receiver operating characteristic curve|
|ACP: American College of Physicians|
|ANN: artificial neural network|
|BERT: Bidirectional Encoder Representations from Transformers|
|CNN: convolutional neural network|
|DT: decision tree|
|EBM: evidence-based medicine|
|MeSH: Medical Subject Heading|
|NB: naïve Bayes|
|NLP: natural language processing|
|RCT: randomized controlled trial|
|SVM: support vector machine|
Edited by C Lovis; submitted 13.05.21; peer-reviewed by H Kalicoglu, M Afzal; comments to author 02.07.21; revised version received 15.07.21; accepted 25.07.21; published 09.09.21Copyright
©Wael Abdelkader, Tamara Navarro, Rick Parrish, Chris Cotoi, Federico Germini, Alfonso Iorio, R Brian Haynes, Cynthia Lokker. Originally published in JMIR Medical Informatics (https://medinform.jmir.org), 09.09.2021.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Medical Informatics, is properly cited. The complete bibliographic information, a link to the original publication on https://medinform.jmir.org/, as well as this copyright and license information must be included.