Viability of Applying Large Language Models to Indoor Climate Sensor and Health Data for Scientific Discovery

• Main Authors: Chen, Aidan, Du, Jonathan, Rodriguez, Aaron, Rodriguez, Ryan, Higgins, Jack, Podmore, Robin, Liu, Ryan, Ilao, Emin, Degilla, Sam, Bibiano, James, Chan, Candice, Pickering, Annalicia, Showstark, Mary, Lee, Jarone, Chen, Milton
• Format: Conference Proceeding
• Language: English
• Subjects: Internet; Large language models; Lungs; Medical services; Meteorology; Monitoring; Particle measurements; Pollution; Pollution measurement; Temperature measurement
• ISSN: 2473-5728
• DOI: 10.1109/GHTC62424.2024.10771517

In partnership with IEEE Smart Village, the VSee team developed an indoor climate monitoring system deployed to homes in polluted areas of the Philippines. We collected participants' medical records including smoking history and methods of cooking, as well as measuring their heart and lung health via electrocardiogram and spirometer tests. We found that indoor temperatures can exceed published macroclimate temperatures by 9 degrees, indoor pollution is just as detrimental to health as secondhand smoke, CO2 levels in high-end hotels in the USA may be comparable to the low-income homes next to large trash landfills in Manila, and that the air quality on airplanes and trains is often borderline unhealthy. Additionally, we trained several large language models (LLMs) on our pollution and medical records data to explore the viability of using LLMs to accelerate scientific discoveries by analyzing patterns in largescale datasets. We then tested 3 AI systems-GPT4 (OpenAI), Gemini 1.5 Pro (Google), Claude 3 Opus (Anthropic)-and found that Anthropic's model has a slight edge over that of OpenAI and Google on our datasets, and that modern LLMs in general are just as good as human physicians and scientists in formulating research hypotheses and selecting subjects for experiments.