Leveraging LLMs and explainable AI to decode citizen complaints for neighborhood respiratory health prediction

Respiratory illnesses pose spatially varying threats to human well-being, yet effective monitoring approaches remain limited. This study introduces a novel, place-based framework for predicting neighborhood-level respiratory health by decoding citizen complaints. Utilizing 2.3 million 311 non-urgent service request records in New York City, we demonstrate that citizen complaints outperform conventional census-derived variables in predicting census tract-level asthma prevalence, achieving higher accuracy and reduced systemic bias. Incorporating large language models (LLMs) yields an additional 8 % performance gain through context-aware weighting of complaint content. Explainable AI techniques reveal the marginal contributions of complaint data, identifying the categories most influential in prediction. Our findings uncover neighborhood respiratory health signals embedded in complaint records, and present the value of AI in translating unstructured urban data into predictive capacity and actionable insights. By pairing publicly available data with cutting-edge AI methods, our approach offers an efficient, scalable tool for assessing respiratory health disparities and informing timely, targeted, place-based interventions.