Developing automated meta-research approaches in the preclinical Alzheimer's disease literature
Alzheimer’s disease is a devastating neurodegenerative disorder for which there is no cure. A crucial part of the drug development pipeline involves testing therapeutic interventions in animal disease models. However, promising findings in preclinical experiments have not translated into clinical trial success. Reproducibility has often been cited as a major issue affecting biomedical research, where experimental results in one laboratory cannot be replicated in another. By using meta-research (research on research) approaches such as systematic reviews, researchers aim to identify and summarise all available evidence relating to a specific research question. By conducting a meta-analysis, researchers can also combine the results from different experiments statistically to understand the overall effect of an intervention and to explore reasons for variations seen across different publications. Systematic reviews of the preclinical Alzheimer’s disease literature could inform decision making, encourage research improvement, and identify gaps in the literature to guide future research. However, due to the vast amount of potentially useful evidence from animal models of Alzheimer’s disease, it remains difficult to make sense of and utilise this data effectively. Systematic reviews are common practice within evidence based medicine, yet their application to preclinical research is often limited by the time and resources required. In this thesis, I develop, build-upon, and implement automated meta-research approaches to collect, curate, and evaluate the preclinical Alzheimer’s literature. I searched several biomedical databases to obtain all research relevant to Alzheimer’s disease. I developed a novel deduplication tool to automatically identify and remove duplicate publications identified across different databases with minimal human effort. I trained a crowd of reviewers to annotate a subset of the publications identified and used this data to train a machine learning algorithm to screen through the remaining publications for relevance. I developed text-mining tools to extract model, intervention, and treatment information from publications and I improved existing automated tools to extract reported measures to reduce the risk of bias. Using these tools, I created a categorised database of research in transgenic Alzheimer’s disease animal models and created a visual summary of this dataset on an interactive, openly accessible online platform. Using the techniques described, I also identified relevant publications within the categorised dataset to perform systematic reviews of two key outcomes of interest in transgenic Alzheimer’s disease models: (1) synaptic plasticity and transmission in hippocampal slices and (2) motor activity in the open field test. Over 400,000 publications were identified across biomedical research databases, with 230,203 unique publications. In a performance evaluation across different preclinical datasets, the automated deduplication tool I developed could identify over 97% of duplicate citations and a had an error rate similar to that of human performance. When evaluated on a test set of publications, the machine learning classifier trained to identify relevant research in transgenic models performed was highly sensitive (captured 96.5% of relevant publications) and excluded 87.8% of irrelevant publications. Tools to identify the model(s) and outcome measure(s) within the full-text of publications may reduce the burden on reviewers and were found to be more sensitive than searching only the title and abstract of citations. Automated tools to assess risk of bias reporting were highly sensitive and could have the potential to monitor research improvement over time. The final dataset of categorised Alzheimer’s disease research contained 22,375 publications which were then visualised in the interactive web application. Within the application, users can see how many publications report measures to reduce the risk of bias and how many have been classified as using each transgenic model, testing each intervention, and measuring each outcome. Users can also filter to obtain curated lists of relevant research, allowing them to perform systematic reviews at an accelerated pace with reduced effort required to search across databases, and a reduced number of publications to screen for relevance. Both systematic reviews and meta-analyses highlighted failures to report key methodological information within publications. Poor transparency of reporting limited the statistical power I had to understand the sources of between-study variation. However, some variables were found to explain a significant proportion of the heterogeneity. Transgenic animal model had a significant impact on results in both reviews. For certain open field test outcomes, wall colour of the open field arena and the reporting of measures to reduce the risk of bias were found to impact results. For in vitro electrophysiology experiments measuring synaptic plasticity, several electrophysiology parameters, including magnesium concentration of the recording solution, were found to explain a significant proportion of the heterogeneity. Automated meta-research approaches and curated web platforms summarising preclinical research could have the potential to accelerate the conduct of systematic reviews and maximise the potential of existing evidence to inform translation.