Jadwal Sholat

Memuat jadwal sholatโ€ฆ

Kedokteran & Kesehatan editorial

Open AccessOA2023

AutoTrial: Prompting Language Models for Clinical Trial Design

AutoTrial uses a hybrid of discrete and neural prompting to generate clinical trial eligibility criteria, achieving a 60% win rate against GPT-3.5 in human evaluations despite a much smaller parameter size.
Zifeng Wang; Cao Xiao; Jimeng Sunยท Conference on Empirical Methods in Natural Language Processingยท 2023ยท DOI 10.48550/arXiv.2305.11366

The core problem

Clinical trials are essential for drug development, and constructing appropriate eligibility criteria (inclusion/exclusion criteria for patient recruitment) is critical to trial success. Proper design should consider similar precedent trials and their eligibility criteria to ensure sufficient patient coverage. The authors present AutoTrial, a method to aid the design of clinical eligibility criteria using language models. The method enables (1) controllable generation under instructions via a hybrid of discrete and neural prompting, (2) scalable knowledge incorporation via in-context learning, and (3) explicit reasoning chains to provide rationales for understanding the outputs.

Innovation

AutoTrial leverages language models with a hybrid prompting strategy that combines discrete prompts (e.g., task instructions and examples) and neural prompts (learned continuous representations). This allows controllable generation under instructions. In-context learning is used to incorporate scalable knowledge from precedent trials, enabling the model to adapt to new trials without retraining. Additionally, the method generates explicit reasoning chains that provide rationales for the generated criteria, enhancing interpretability. The approach is evaluated on over 70K clinical trials.
Introduction
Clinical trials are essential for drug development, and constructing appropriate eligibility criteria (inclusion/exclusion criteria for patient recruitment) is critical to trial success. Proper design should consider similar precedent trials and their eligibility criteria to ensure sufficient patient coverage. The authors present AutoTrial, a method to aid the design of clinical eligibility criteria using language models. The method enables (1) controllable generation under instructions via a hybrid of discrete and neural prompting, (2) scalable knowledge incorporation via in-context learning, and (3) explicit reasoning chains to provide rationales for understanding the outputs.

Why it matters

The results demonstrate that AutoTrial can effectively assist in clinical trial design by generating relevant and coherent eligibility criteria. The hybrid prompting approach and in-context learning allow the model to incorporate domain knowledge and produce controllable outputs. The explicit reasoning chains provide transparency, which is crucial for clinical applications. The strong performance against GPT-3.5, despite a smaller model size, suggests that the proposed prompting strategy is efficient and effective. Future work could explore broader applications and further refinements.

Who should read this

Pembaca Research Digest dan praktisi rumpun ini.

Opening member contentโ€ฆ