ROBoto2: An Interactive System and Dataset for LLM-assisted Clinical Trial Risk of Bias Assessment

AI summary

Overview

• Research area: Natural Language Processing applied to evidence synthesis and clinical trial quality assessment (specifically Cochrane Risk of Bias version 2, or ROB2).
• Technical level: Intermediate. The system design and evaluation are described plainly, but familiarity with retrieval-augmented generation, F1 metrics, and inter-rater reliability statistics helps.
• Scope: The paper introduces ROBoto2, an open-source web platform and released dataset for LLM-assisted, human-in-the-loop ROB2 assessment of clinical trial reports, plus a benchmark of retrieval methods and four LLMs on that dataset.

What This Paper Is About

Assessing risk of bias in clinical trials with the Cochrane ROB2 tool is a required step in systematic reviews, but it takes trained reviewers 30+ minutes per trial report and does not scale to reviews covering hundreds or thousands of trials. The authors build ROBoto2, a web-based tool that parses a trial PDF, retrieves relevant passages, prompts an LLM with each of the ROB2 signaling questions, and lets an expert accept, edit, or override each suggestion. They also release a dataset of 521 pediatric trial assessments and benchmark four LLMs to characterize how well current models handle this task.

Key Contributions

1. The ROBoto2 system: A public, open-source web tool (code and API available) that integrates document preprocessing, within-document passage retrieval, LLM prompting, and interactive expert review into a human-AI collaborative ROB2 pipeline.
2. A released dataset: 521 ROB2 assessments comprising 8954 signaling questions and 1202 evidence passages, including both manual annotations and LLM-assisted annotations by medical experts, collected during a real systematic review of pediatric clinical trial literature.
3. A benchmark and analysis: The first evaluation of LLM-assisted ROB2 assessment, covering retrieval strategies (BM25 vs. Sentence-Transformers) and four LLMs (Llama-3.3-70B-Instruct, GPT-3.5-Turbo, GPT-4o, Claude 3.5-Sonnet) across oracle-evidence, retrieved-evidence, and full-paper settings.
4. Practical system usage statistics: Reported acceptance and correction rates for model answers, rationales, and retrieved evidence passages from the 276 assessments completed inside ROBoto2.

Main Findings

• Moderate best-case accuracy: The best performing configuration, Claude 3.5-Sonnet with the full paper as context, reached a micro-F1 of 0.71, which the authors describe as leaving considerable room for improvement.
• Domain difficulty varies: All evaluated models do well on Domain 1, where questions tend to be answerable directly from the text. Performance is mixed on D2 and D3, which can require interpreting numerical data, and D5 is difficult because it may require knowledge of external clinical resources and guidelines.
• Context helps, with one exception: Increasing context generally improves performance for most models, but reduces accuracy for GPT-3.5-Turbo. Retrieval-based settings sometimes surpass oracle-evidence performance, likely because annotator-selected evidence passages are incomplete.
• Few-shot prompting does not help: Few-shot prompting did not outperform zero-shot prompting when models received the same number of context passages.
• Over-conservative automation: Human raters assessed 47 of 276 trials as high risk, while the LLM-only pipeline assessed 101 as high risk. Error analysis shows the strongest models tend to over-select "No Information," which the authors suggest may reflect cautiousness from safety and alignment training.
• Humans still correct the model frequently: In the 276 ROBoto2 assessments, annotators supplied their own answer 42.4% of the time and edited rationales 28.7% of the time, rather than using the model's answer (57.6%) or rationale (71.3%).
• Evidence retrieval quality: Of 3370 retrieved passages, 615 up/downvotes were collected, 78.0% of which were positive feedback. Sentence-Transformers (S-BERT) outperformed BM25 on retrieval: recall@1/3/5/10 was 0.268/0.455/0.519/0.678 for S-BERT versus 0.140/0.272/0.367/0.533 for BM25.
• Inter-rater agreement is limited but consistent with prior work: On 20 papers (440 signaling questions) independently double-annotated in ROBoto2, four-class Cohen's Kappa was 0.40, described as fair to moderate agreement and consistent with prior reports of slight to moderate agreement among experienced raters (Fleiss' Kappa of 0.45 at the domain level).

Methodology in Plain English

The authors model ROB2 assessment as a document-level question-answering task. A user uploads a clinical trial PDF, which is converted to structured JSON using the S2ORC-doc2json library. Each paragraph is embedded with Sentence-Transformers all-MiniLM-L6-v2 and stored in a key-value index. For each of the ROB2 signaling questions, the question text is embedded with the same model, the top-k most similar paragraphs are retrieved by cosine similarity, and an LLM is prompted with the question, its elaboration text, and the retrieved paragraphs to produce an answer plus a rationale. The system also supports giving the full paper as context to models with large context windows. The published web interface uses all-MiniLM-L6-v2 with k=3 as a balance between performance and cost, and is built with React and TypeScript for client-side embedding and retrieval, plus a Python/FastAPI back end for parsing and LLM calls. Domain-level and overall risk judgments are produced by implementing the ROB2 flowcharts in code.

For evaluation, the authors used 245 reports annotated manually by five expert annotators (graduate degrees in public health, epidemiology, medical sciences, or clinical practice, with systematic review experience), drawn from an initial corpus of 2334 matching pediatric trial reports published 1991-2020. An additional 276 reports were annotated with LLM assistance using a version of ROBoto2 that used retrieval with gpt-3.5-turbo-0125, and were withheld from the final evaluation but included in the dataset. Answers were aggregated into three classes (Yes/Probably Yes, No/Probably No, No Information) for micro- and macro-averaged F1 scoring.

Why This Matters

• Impact on research: Risk of bias assessment is a documented bottleneck in systematic reviews, and this work provides both an operational tool and the first benchmark for LLM-assisted ROB2, giving future work a reproducible starting point. The dataset includes both model outputs and expert corrections, which the authors explicitly release to support future model development.
• Real-world applications:
  • Systematic review teams screening large bodies of pediatric trial literature for evidence quality.
  • Clinical guideline developers who need structured quality judgments across many trials.
  • Evidence synthesis groups repurposing saved assessment time into gathering and vetting supporting evidence passages.
  • Researchers studying human-AI collaboration, since ROBoto2 logs acceptance, correction, and passage-voting behavior alongside model outputs.
• Industry relevance: The system is open source with a public web interface and API, making it deployable by review organizations and tool vendors. The finding that fully automated assessment is over-conservative implies that commercial or institutional deployments should keep a human in the loop rather than advertising end-to-end automation.

Future Directions

• Measuring and mitigating anchoring bias: The authors note that exposing model outputs may bias annotators, and leave measurement of this effect in the ROB setting to future work.
• Reasoning models: The experiments exclude reasoning models such as OpenAI's o3, Anthropic's Claude 4-Sonnet-Thinking, and DeepSeek-R1; the authors suggest testing whether these improve answer classification accuracy and rationale quality.
• Addressing dataset imbalance and sparse annotations: The dataset is unbalanced toward "some concerns" labels, and evidence paragraphs exist only for a small, domain-skewed subset of questions (D1 has the most, D5 very few), so retrieval and oracle results may not represent sparsely annotated questions.
• Verifying efficiency and quality gains: The annotation team qualitatively reported that time saved on assessment was repurposed into judging and retrieving evidence, but the impact of this repurposing on final assessment quality was not explicitly measured and should be studied.

Target Audience

This paper is most useful to evidence synthesis researchers and systematic review methodologists, clinical trialists and guideline developers evaluating trial quality, and NLP researchers working on human-in-the-loop LLM systems for biomedical document understanding. It is also relevant to tool builders interested in deployment lessons from a real annotation campaign, since the paper reports what experts accepted, edited, and rejected in practice.

Authors’ abstract

We present ROBOTO2, an open-source, web-based platform for large language model (LLM)-assisted risk of bias (ROB) assessment of clinical trials. ROBOTO2 streamlines the traditionally labor-intensive ROB v2 (ROB2) annotation process via an interactive interface that combines PDF parsing, retrieval-augmented LLM prompting, and human-in-the-loop review. Users can upload clinical trial reports, receive preliminary answers and supporting evidence for ROB2 signaling questions, and provide real-time feedback or corrections to system suggestions. ROBOTO2 is publicly available at https://roboto2.vercel.app/, with code and data released to foster reproducibility and adoption. We construct and release a dataset of 521 pediatric clinical trial reports (8954 signaling questions with 1202 evidence passages), annotated using both manually and LLM-assisted methods, serving as a benchmark and enabling future research. Using this dataset, we benchmark ROB2 performance for 4 LLMs and provide an analysis into current model capabilities and ongoing challenges in automating this critical aspect of systematic review.