Aligning text mining and machine learning algorithms with best practices for study selection in systematic literature reviews
Bibliographic record
Abstract
BACKGROUND: Despite existing research on text mining and machine learning for title and abstract screening, the role of machine learning within systematic literature reviews (SLRs) for health technology assessment (HTA) remains unclear given lack of extensive testing and of guidance from HTA agencies. We sought to address two knowledge gaps: to extend ML algorithms to provide a reason for exclusion-to align with current practices-and to determine optimal parameter settings for feature-set generation and ML algorithms. METHODS: We used abstract and full-text selection data from five large SLRs (n = 3089 to 12,769 abstracts) across a variety of disease areas. Each SLR was split into training and test sets. We developed a multi-step algorithm to categorize each citation into the following categories: included; excluded for each PICOS criterion; or unclassified. We used a bag-of-words approach for feature-set generation and compared machine learning algorithms using support vector machines (SVMs), naïve Bayes (NB), and bagged classification and regression trees (CART) for classification. We also compared alternative training set strategies: using full data versus downsampling (i.e., reducing excludes to balance includes/excludes because machine learning algorithms perform better with balanced data), and using inclusion/exclusion decisions from abstract versus full-text screening. Performance comparisons were in terms of specificity, sensitivity, accuracy, and matching the reason for exclusion. RESULTS: The best-fitting model (optimized sensitivity and specificity) was based on the SVM algorithm using training data based on full-text decisions, downsampling, and excluding words occurring fewer than five times. The sensitivity and specificity of this model ranged from 94 to 100%, and 54 to 89%, respectively, across the five SLRs. On average, 75% of excluded citations were excluded with a reason and 83% of these citations matched the reviewers' original reason for exclusion. Sensitivity significantly improved when both downsampling and abstract decisions were used. CONCLUSIONS: ML algorithms can improve the efficiency of the SLR process and the proposed algorithms could reduce the workload of a second reviewer by identifying exclusions with a relevant PICOS reason, thus aligning with HTA guidance. Downsampling can be used to improve study selection, and improvements using full-text exclusions have implications for a learn-as-you-go approach.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.682 | 0.871 |
| Meta-epidemiology (narrow) | 0.004 | 0.003 |
| Meta-epidemiology (broad) | 0.011 | 0.014 |
| Bibliometrics | 0.041 | 0.034 |
| Science and technology studies | 0.003 | 0.006 |
| Scholarly communication | 0.016 | 0.013 |
| Open science | 0.008 | 0.010 |
| Research integrity | 0.004 | 0.005 |
| Insufficient payload (model declined to judge) | 0.003 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; the direct Gemma label and the distilled Codex classifier agree on what is shown here.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".