Existing systems for automating the assessment of risk-of-bias (RoB) in medical studies are supervised approaches that require substantial training data to work well. However, recent revisions to RoB guidelines have resulted in a scarcity of available training data. In this study, we investigate the effectiveness of generative large language models (LLMs) for assessing RoB. Their application requires little or no training data and, if successful, could serve as a valuable tool to assist human experts during the construction of systematic reviews. Following Cochrane's latest guidelines (RoB2) designed for human reviewers, we prepare instructions that are fed as input to LLMs, which then infer the risk associated with a trial publication. We distinguish between two modelling tasks: directly predicting RoB2 from text; and employing decomposition, in which a RoB2 decision is made after the LLM responds to a series of signalling questions. We curate new testing data sets and evaluate the performance of four general- and medical-domain LLMs. The results fall short of expectations, with LLMs seldom surpassing trivial baselines. On the direct RoB2 prediction test set (n = 5993), LLMs perform akin to the baselines (F1: 0.1-0.2). In the decomposition task setup (n = 28,150), similar F1 scores are observed. Our additional comparative evaluation on RoB1 data also reveals results substantially below those of a supervised system. This testifies to the difficulty of solving this task based on (complex) instructions alone. Using LLMs as an assisting technology for assessing RoB2 thus currently seems beyond their reach.

BACKGROUND:Assessment of the quality of medical evidence available on the web is a critical step in the preparation of systematic reviews. Existing tools that automate parts of this task validate the quality of individual studies but not of entire bodies of evidence and focus on a restricted set of quality criteria.OBJECTIVE:We proposed a quality assessment task that provides an overall quality rating for each body of evidence (BoE), as well as finer-grained justification for different quality criteria according to the Grading of Recommendation, Assessment, Development, and Evaluation formalization framework. For this purpose, we constructed a new data set and developed a machine learning baseline system (EvidenceGRADEr).METHODS:We algorithmically extracted quality-related data from all summaries of findings found in the Cochrane Database of Systematic Reviews. Each BoE was defined by a set of population, intervention, comparison, and outcome criteria and assigned a quality grade (high, moderate, low, or very low) together with quality criteria (justification) that influenced that decision. Different statistical data, metadata about the review, and parts of the review text were extracted as support for grading each BoE. After pruning the resulting data set with various quality checks, we used it to train several neural-model variants. The predictions were compared against the labels originally assigned by the authors of the systematic reviews.RESULTS:Our quality assessment data set, Cochrane Database of Systematic Reviews Quality of Evidence, contains 13,440 instances, or BoEs labeled for quality, originating from 2252 systematic reviews published on the internet from 2002 to 2020. On the basis of a 10-fold cross-validation, the best neural binary classifiers for quality criteria detected risk of bias at 0.78 F1 (P=.68; R=0.92) and imprecision at 0.75 F1 (P=.66; R=0.86), while the performance on inconsistency, indirectness, and publication bias criteria was lower (F1 in the range of 0.3-0.4). The prediction of the overall quality grade into 1 of the 4 levels resulted in 0.5 F1. When casting the task as a binary problem by merging the Grading of Recommendation, Assessment, Development, and Evaluation classes (high+moderate vs low+very low-quality evidence), we attained 0.74 F1. We also found that the results varied depending on the supporting information that is provided as an input to the models.CONCLUSIONS:Different factors affect the quality of evidence in the context of systematic reviews of medical evidence. Some of these (risk of bias and imprecision) can be automated with reasonable accuracy. Other quality dimensions such as indirectness, inconsistency, and publication bias prove more challenging for machine learning, largely because they are much rarer. This technology could substantially reduce reviewer workload in the future and expedite quality assessment as part of evidence synthesis.

Gleb Kuzmin, Artem Vazhentsev, Artem Shelmanov, Xudong Han, Simon Suster, Maxim Panov, Alexander Panchenko, Timothy Baldwin. Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Automatically rating the quality of published research is a critical step in medical evidence synthesis. While several methods have been proposed, their algorithmic fairness has been overlooked even though significant risks may follow when such systems are deployed in biomedical contexts. In this work, we study fairness on two systems along two sensitive attributes, participant sex and medical area. In some cases, we find important inequalities, leading us to apply various debiasing methods. Upon examining an interplay of systems’ predictive performance, fairness, as well as medically critical selective classification capabilities and calibration performance, we find that fairness can sometimes improve through debiasing, but at a cost in other performance measures.

Objectives: A major obstacle in deployment of models for automated quality assessment is their reliability. To analyze their calibration and selective classification performance.Study Design and Setting: We examine two systems for assessing the quality of medical evidence, EvidenceGRADEr and RobotReviewer, both developed from Cochrane Database of Systematic Reviews (CDSR) to measure strength of bodies of evidence and risk of bias (RoB) of individual studies, respectively. We report their calibration error and Brier scores, present their reliability diagrams, and analyze the risk-coverage trade-off in selective classification.Results: The models are reasonably well calibrated on most quality criteria (expected calibration error [ECE] 0.04-0.09 for EvidenceGRADEr, 0.03-0.10 for RobotReviewer). However, we discover that both calibration and predictive performance vary significantly by medical area. This has ramifications for the application of such models in practice, as average performance is a poor indicator of group-level performance (e.g., health and safety at work, allergy and intolerance, and public health see much worse performance than cancer, pain, and anesthesia, and Neurology). We explore the reasons behind this disparity. Conclusion: Practitioners adopting automated quality assessment should expect large fluctuations in system reliability and predictive performance depending on the medical area. Prospective indicators of such behavior should be further researched. & COPY; 2023 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Background Publication of registered clinical trials is a critical step in the timely dissemination of trial findings. However, a significant proportion of completed clinical trials are never published, motivating the need to analyze the factors behind success or failure to publish. This could inform study design, help regulatory decision-making, and improve resource allocation. It could also enhance our understanding of bias in the publication of trials and publication trends based on the research direction or strength of the findings. Although the publication of clinical trials has been addressed in several descriptive studies at an aggregate level, there is a lack of research on the predictive analysis of a trial’s publishability given an individual (planned) clinical trial description. Objective We aimed to conduct a study that combined structured and unstructured features relevant to publication status in a single predictive approach. Established natural language processing techniques as well as recent pretrained language models enabled us to incorporate information from the textual descriptions of clinical trials into a machine learning approach. We were particularly interested in whether and which textual features could improve the classification accuracy for publication outcomes. Methods In this study, we used metadata from ClinicalTrials.gov (a registry of clinical trials) and MEDLINE (a database of academic journal articles) to build a data set of clinical trials (N=76,950) that contained the description of a registered trial and its publication outcome (27,702/76,950, 36% published and 49,248/76,950, 64% unpublished). This is the largest data set of its kind, which we released as part of this work. The publication outcome in the data set was identified from MEDLINE based on clinical trial identifiers. We carried out a descriptive analysis and predicted the publication outcome using 2 approaches: a neural network with a large domain-specific language model and a random forest classifier using a weighted bag-of-words representation of text. Results First, our analysis of the newly created data set corroborates several findings from the existing literature regarding attributes associated with a higher publication rate. Second, a crucial observation from our predictive modeling was that the addition of textual features (eg, eligibility criteria) offers consistent improvements over using only structured data (F1-score=0.62-0.64 vs F1-score=0.61 without textual features). Both pretrained language models and more basic word-based representations provide high-utility text representations, with no significant empirical difference between the two. Conclusions Different factors affect the publication of a registered clinical trial. Our approach to predictive modeling combines heterogeneous features, both structured and unstructured. We show that methods from natural language processing can provide effective textual features to enable more accurate prediction of publication success, which has not been explored for this task previously.

We explore whether state-of-the-art BERT models encode sufficient domain knowledge to correctly perform domain-specific inference. Although BERT implementations such as BioBERT are better at domain-based reasoning than those trained on general-domain corpora, there is still a wide margin compared to human performance on these tasks. To bridge this gap, we explore whether supplementing textual domain knowledge in the medical NLI task: a) by further language model pretraining on the medical domain corpora, b) by means of lexical match algorithms such as the BM25 algorithm, c) by supplementing lexical retrieval with dependency relations, or d) by using a trained retriever module, can push this performance closer to that of humans. We do not find any significant difference between knowledge supplemented classification as opposed to the baseline BERT models, however. This is contrary to the results for evidence retrieval on other tasks such as open domain question answering (QA). By examining the retrieval output, we show that the methods fail due to unreliable knowledge retrieval for complex domain-specific reasoning. We conclude that the task of unsupervised text retrieval to bridge the gap in existing information to facilitate inference is more complex than what the state-of-the-art methods can solve, and warrants extensive research in the future.

Several previous studies on explanation for recurrent neural networks focus on approaches that find the most important input segments for a network as its explanations. In that case, the manner in which these input segments combine with each other to form an explanatory pattern remains unknown. To overcome this, some previous work tries to find patterns (called rules) in the data that explain neural outputs. However, their explanations are often insensitive to model parameters, which limits the scalability of text explanations. To overcome these limitations, we propose a pipeline to explain RNNs by means of decision lists (also called rules) over skipgrams. For evaluation of explanations, we create a synthetic sepsis-identification dataset, as well as apply our technique on additional clinical and sentiment analysis datasets. We find that our technique persistently achieves high explanation fidelity and qualitatively interpretable rules.

Effective representation of biomedical names for downstream NLP tasks requires the encoding of both lexical as well as domain-specific semantic information. Ideally, the synonymy and semantic relatedness of names should be consistently reflected by their closeness in an embedding space. To achieve such robustness, prior research has considered multi-task objectives when training neural encoders. In this paper, we take a next step towards truly robust representations, which capture more domain-specific semantics while remaining universally applicable across different biomedical corpora and domains. To this end, we use conceptual grounding constraints which more effectively align encoded names to pretrained embeddings of their concept identifiers. These constraints are effective even when using a Deep Averaging Network, a simple feedforward encoding architecture that allows for scaling to large corpora while remaining sufficiently expressive. We empirically validate our approach using multiple tasks and benchmarks, which assess both literal synonymy as well as more general semantic relatedness.

Neural encoders of biomedical names are typically considered robust if representations can be effectively exploited for various downstream NLP tasks. To achieve this, encoders need to model domain-specific biomedical semantics while rivaling the universal applicability of pretrained self-supervised representations. Previous work on robust representations has focused on learning low-level distinctions between names of fine-grained biomedical concepts. These fine-grained concepts can also be clustered together to reflect higher-level, more general semantic distinctions, such as grouping the names nettle sting and tick-borne fever together under the description puncture wound of skin. It has not yet been empirically confirmed that training biomedical name encoders on fine-grained distinctions automatically leads to bottom-up encoding of such higher-level semantics. In this paper, we show that this bottom-up effect exists, but that it is still relatively limited. As a solution, we propose a scalable multi-task training regime for biomedical name encoders which can also learn robust representations using only higher-level semantic classes. These representations can generalise both bottom-up as well as top-down among various semantic hierarchies. Moreover, we show how they can be used out-of-the-box for improved unsupervised detection of hypernyms, while retaining robust performance on various semantic relatedness benchmarks.

While solving math word problems automatically has received considerable attention in the NLP community, few works have addressed probability word problems specifically. In this paper, we employ and analyse various neural models for answering such word problems. In a two-step approach, the problem text is first mapped to a formal representation in a declarative language using a sequence-to-sequence model, and then the resulting representation is executed using a probabilistic programming system to provide the answer. Our best performing model incorporates general-domain contextualised word representations that were finetuned using transfer learning on another in-domain dataset. We also apply end-to-end models to this task, which bring out the importance of the two-step approach in obtaining correct solutions to probability problems.

We present COVID-SEE, a system for medical literature discovery based on the concept of information exploration, which builds on several distinct text analysis and natural language processing methods to structure and organise information in publications, and augments search through a visual overview of a collection enabling exploration to identify key articles of interest. We developed this system over COVID-19 literature to help medical professionals and researchers explore the literature evidence, and improve findability of relevant information. COVID-SEE is available at http://covid-see.com.

Recent research on robust representations of biomedical names has focused on modeling large amounts of fine-grained conceptual distinctions using complex neural encoders. In this paper, we explore the opposite paradigm: training a simple encoder architecture using only small sets of names sampled from high-level biomedical concepts. Our encoder post-processes pretrained representations of biomedical names, and is effective for various types of input representations, both domain-specific or unsupervised. We validate our proposed few-shot learning approach on multiple biomedical relatedness benchmarks, and show that it allows for continual learning, where we accumulate information from various conceptual hierarchies to consistently improve encoder performance. Given these findings, we propose our approach as a low-cost alternative for exploring the impact of conceptual distinctions on robust biomedical name representations.

The COVID-19 pandemic has driven ever-greater demand for tools which enable efficient exploration of biomedical literature. Although semi-structured information resulting from concept recognition and detection of the defining elements of clinical trials (e.g. PICO criteria) has been commonly used to support literature search, the contributions of this abstraction remain poorly understood. In this study, we compare the results retrieved by a standard search engine with those filtered using clinically-relevant concepts and their relations. Most importantly, with an analysis based on the TREC-COVID dataset, we find that the relational concept selection filters the original retrieved collection in a way that decreases the proportion of unjudged documents and increases the precision, which means that the user is likely to be exposed to a larger number of relevant documents.

Several previous studies on explanation for recurrent neural networks focus on approaches that find the most important input segments for a network as its explanations. In that case, the manner in which these input segments combine with each other to form an explanatory pattern remains unknown. To overcome this, some previous work tries to find patterns (called rules) in the data that explain neural outputs. However, their explanations are often insensitive to model parameters, which limits the scalability of text explanations. To overcome these limitations, we propose a pipeline to explain RNNs by means of decision lists (also called rules) over skipgrams. For evaluation of explanations, we create a synthetic sepsis-identification dataset, as well as apply our technique on additional clinical and sentiment analysis datasets. We find that our technique persistently achieves high explanation fidelity and qualitatively interpretable rules.

Efficient discovery and exploration of biomedical literature has grown in importance in the context of the COVID-19 pandemic, and topic-based methods such as latent Dirichlet allocation (LDA) are a useful tool for this purpose. In this study we compare traditional topic models based on word tokens with topic models based on medical concepts, and propose several ways to improve topic coherence and specificity.

Concept extraction is an important step in clinical natural language processing. Once extracted, the use of concepts can improve the accuracy and generalization of downstream systems. We present a new unsupervised system for the extraction of concepts from clinical text. The system creates representations of concepts from the Unified Medical Language System (UMLS®) by combining natural language descriptions of concepts with word representations, and composing these into higher-order concept vectors. These concept vectors are then used to assign labels to candidate phrases which are extracted using a syntactic chunker. Our approach scores an exact F-score of.32 and an inexact F-score of.45 on the well-known I2b2-2010 challenge corpus, outperforming the only other unsupervised concept extraction method. As our approach relies only on word representations and a chunker, it is completely unsupervised. As such, it can be applied to languages and corpora for which we do not have prior annotations. All our code is open-source and can be found at www.github.com/clips/conch.

Memory networks have been a popular choice among neural architectures for machine reading comprehension and question answering. While recent work revealed that memory networks can't truly perform multi-hop reasoning, we show in the present paper that vanilla memory networks are ineffective even in single-hop reading comprehension. We analyze the reasons for this on two cloze-style datasets, one from the medical domain and another including children's fiction. We find that the output classification layer with entity-specific weights, and the aggregation of passage information with relatively flat attention distributions are the most important contributors to poor results. We propose network adaptations that can serve as simple remedies. We also find that the presence of unseen answers at test time can dramatically affect the reported results, so we suggest controlling for this factor during evaluation.

Understanding the behavior of a trained network and finding explanations for its outputs is important for improving the network's performance and generalization ability, and for ensuring trust in automated systems. Several approaches have previously been proposed to identify and visualize the most important features by analyzing a trained network. However, the relations between different features and classes are lost in most cases. We propose a technique to induce sets of if-then-else rules that capture these relations to globally explain the predictions of a network. We first calculate the importance of the features in the trained network. We then weigh the original inputs with these feature importance scores, simplify the transformed input space, and finally fit a rule induction model to explain the model predictions. We find that the output rule-sets can explain the predictions of a neural network trained for 4-class text classification from the 20 newsgroups dataset to a macro-averaged F-score of 0.80. We make the code available at https://github.com/clips/interpret_with_rules.

Recently, segment convolutional neural networks have been proposed for end-to-end relation extraction in the clinical domain, achieving results comparable to or outperforming the approaches with heavy manual feature engineering. In this paper, we analyze the errors made by the neural classifier based on confusion matrices, and then investigate three simple extensions to overcome its limitations. We find that including ontological association between drugs and problems, and data-induced association between medical concepts does not reliably improve the performance, but that large gains are obtained by the incorporation of semantic classes to capture relation triggers.