Simplifying Digital Therapeutics Explanations Using Large Language Models: Randomized Online Experiments on Readability and Perceived Comprehension (Preprint)

Authors:
Preprint
Publisher website
Google Scholar
Abstract
BACKGROUND Digital therapeutics (DTx) are evidence-based software interventions with the potential to treat health conditions. However, uptake remains limited by low public awareness and overly complex patient education materials that exceed recommended readability levels. Large language models (LLMs) may simplify such content; however, their effect on actual comprehension has not been empirically demonstrated. OBJECTIVE To examine whether LLM-based simplification of DTx explanatory materials enhances public comprehension and subjective evaluations of readability, clarity, and comprehensibility compared with manufacturer-provided documents. METHODS We developed a simplification tool using the GPT-4o API, configured for deterministic outputs and guided by structured readability instructions. Original DTx explanatory materials about insomnia and nicotine dependence were obtained from manufacturers and transformed into simplified versions. Two randomized, between-subject online experiments were conducted (N = 1,000; 500 per condition). Participants were stratified by age and sex and screened for relevance (Insomnia Severity Index ≥8 for the insomnia experiment; smoking ≥5 cigarettes/day for the nicotine dependence experiment). Within each experiment, participants were randomly assigned to review either the original or the LLM-simplified explanation. Perceived understanding and post-exposure evaluations of ease, clarity, and comprehensibility were assessed pre- and post-exposure. RESULTS Repeated-measures analysis of variance revealed significant Group × Time interaction effects on perceived understanding in both experiments: insomnia (F₁,₄₉₈ = 24.8; P <.001) and nicotine dependence (F₁,₄₉₈ = 14.1; P < .001), with greater improvements in the LLM-simplified groups. Mann–Whitney U tests further showed that LLM-simplified explanations were rated as significantly easier, clearer, and more comprehensible than the original versions in both experiments (all P < .05). CONCLUSIONS Compared with manufacturer-provided original materials, LLM-simplified DTx explanations led to greater improvements in perceived understanding and subjective readability among lay audiences, even after a single exposure. This finding highlights the scalability of LLM-based simplification as a strategy to address health literacy barriers. Integrating such tools into patient education may enhance access to digital therapeutic information and support broader societal diffusion.

This publication has 30 references indexed in Scilit: