The model takes an uploaded image or video and provides specific, actionable advice supported by evidence-based resources. The proposed architecture successfully applies both additive and subtractive modifications to images, providing users with a concrete visual representation of a safer environment.

We find that overall, our raters preferred the “Steadi” system output of both image and text (40/54 times (74.1%), Table 1). We demonstrate that contemporary LLMs produce relatively safe recommendations, regardless of the prompting system, with only one set of recommendations rated as unsafe in the baseline prompting system and none in the enhanced. We find that when given specific issues to visualize, image editing LLMs produce edits with good visualization fidelity (46/54 times, 85.2% for baseline and 43/54 times, 79.6% for enhanced; Table 1), and low rates of implausible/hazard-producing edits (6/54 times, 11.1% for both systems; Table 1). Text recommendations and visualized outputs ranged from generally “somewhat actionable” for the baseline system to highly actionable for our enhanced system.

This study introduces a novel application of multimodal LLMs, leveraging their image-generation capabilities for visualizing personalized home safety recommendations. We demonstrate that enhanced frameworks, such as structured prompting and grounding using trusted resources, produce safe, clinically useful, and actionable outputs that categorically rate better than outputs from baseline LLMs. The inherent flexibility of LLMs supports diverse interaction methods, uniquely enabling users to interact with their “consultant” in their preferred mode. LLMs may mitigate delays caused by insurance authorizations and restore autonomy to users.

The visual output capability is also key: generating suggestions directly onto uploaded images offers more intuitive, actionable guidance than abstract text instructions alone. The drive to protect the familiarity of their home from change was identified to be a major motive for older adults rejecting modification advice from OT [5]; direct visualization of user-fed images may help overcome this hurdle and increase acceptance. There are still limitations to the technology, namely outputs may be illogical such as the recommended soap placement, and movement of furniture and door in Multimedia Appendix 1. However, overall, this study demonstrates that LLMs generally produce visual outputs with high fidelity, low hazard introduction rates, and high actionability.

To adhere to HIPAA (Health Insurance Portability and Accountability Act) compliance, future work should consider working with LLM providers to sign a HIPAA Business Associate Amendment or other HIPAA-compliant program. Ethical considerations, such as disclosure of privacy and data protection, should be implemented in accordance with WHO guidance on AI in health [9].

Further testing must be conducted against the current standard for in-home assessments to discover if the proposed model provides comparable advice to professionals. Implementation trials will be needed to mitigate concerns such as the digital divide and ensure accessibility among varying cognitive/visual functions. Implementations must comply with FDA digital health guidance, and characterization and limitation of unsafe output generation must be explored. This model is designed as a supplemental service to be integrated with OT rather than a replacement.

Multimodal LLMs that integrate image generation offer a novel, innovative approach to increasing end users’ accessibility to personalized home environment recommendations for fall prevention. This capability represents a potential supplement to current care services that may enhance patient understanding, motivation, and adherence, serving as a valuable resource to patients who defer or cannot access in-home safety assessments. Rigorous validation of clinical efficacy and user acceptance is essential to translate this technological potential into improved patient outcomes.