Performance of Zero-Shot Classifiers for Categorizing RCT Abstracts by Intervention Type: Validation Study

Artificial intelligence has gained relevance due to its potential to reduce the workload in evidence synthesis or bibliometric projects. While the main focus has been lately on the use of instruction-tuned large language models, zero-shot classification models have not been tested for such task. These models are large language models trained on large datasets of labeled data able to categorize text among proposed labels, irrespective of the text domain or the topic. They are relatively small, able to run on consumer-grade computers, and almost hyperparameter-free.

In our study, we use abstracts of randomized clinical trials in rheumatology as a case example to evaluate the performance of openly available, generalist, zero-shot classification models in classifying types of interventions against a human gold standard.

We classified all rheumatology RCT abstracts published between 2009 and 2022 (n=1,054) as "drug" or "non-drug" using two zero-shot text classification models (DeBERTa and BART) and few-shot prompting using Llama3 8B. Different labeling of categories provided to the zero-shot classification models and different prompts provided to Llama3 8B were tested. Performance was evaluated using accuracy and predictive value of both categories against a human gold standard.

Most randomized controlled trials, RCTs (452/1054, 42.9%) assessed drug interventions. The DeBERTa model achieved the highest accuracy (929/1054, 88.1%; 95% CI 86%-90%) when using the "drug" and "non-drug" labels. Llama3 8B and few-shot prompting had slightly higher accuracy and predictive values. Both zero-shot and Llama3 8B models had performance on par with a human without experience in evidence synthesis (905/1054, 85.9%; 95% CI 83.6%-87.8% accuracy). Misclassifications occurred for trials where the intervention was harder to classify, such as procedures (eg, intra-articular injections), food compounds, vitamins, supplements, or biological treatments.

This study shows the potential of zero-shot classification models for simple classification tasks, demonstrating accuracy comparable to that of an untrained human. These models are potential tools to streamline systematic review tasks for bibliometric studies in classifying abstracts by supplementing one reviewer.