Assessing Trustworthiness in LLMs-Assisted Literature Review: A Case Study of Multimodal Machine Learning in Medicine

Large Language Models (LLMs) are increasingly being used to support the structured extraction of methodological evidence from scientific literature and to build reviews and surveys. However, their use for scientific evidence extraction raises methodological questions concerning the consistency, correctness, and traceability of the extracted information. To address these aspects, this study assesses the trustworthiness of an LLM-assisted framework for structured methodological evidence extraction. A large-scale full-text corpus of multimodal machine learning in medicine, constructed through a structured literature search and automated open-access full-text retrieval, constitutes a case study for the evaluation. In this study, trustworthiness is operationalised through three complementary dimensions: extraction consistency, correctness, and traceability. The proposed pipeline was applied to 1371open-access full-text articles focused on multimodal machine learning in medicine. These articles were independently analysed using two locally deployed open-weight Large Language Models, Qwen and Gemma 3, through the same constrained prompt and predefined JSON schema. The extraction process was constrained through explicit anti-inference rules, while each structured item was linked to source-level textual evidence to support traceability and subsequent inspection. For selected research questions, extraction consistency was assessed through human inter-annotator agreement, whereas correctness was evaluated against an adjudicated human reference. Additional error analyses examined omissions, negative and Not specified outputs, while a targeted analysis conducted on a restricted subset of articles with Calibration-positive predictions examined the availability, textual fidelity, and methodological relevance of the associated source-linked evidence. Resultsare twofold: (i) methodologically, extraction reliability varied across tasks, with finer-grained categories showing lower positive recovery and more omissions; cross-model agreement alone was not sufficient to establish correctness; (ii) in the application case study, both models identified broadly consistent patterns across the analysed corpus, while datasets, fusion strategies, and reliability-oriented methods were frequently assigned to the Not specified category. Overall, the findings support the use of LLM-assisted extraction for large-scale methodological mapping, while highlighting the value of explicit assessment and human validation for interpreting model outputs.