BreatheSafe: An IoT and Machine-Learning Air Quality Monitoring and Health Advisory System with Evidence-Gated Virtual Sensing

Low-cost air-quality monitors usually measure only particulate matter (PM), while India's Central Pollution Control Board (CPCB) National Air Quality Index (AQI) is defined over up to seven pollutants and requires at least three. This paper presents BreatheSafe, an IoT and machine-learning system that measures PM1.0, PM2.5, PM10, temperature, pressure and total volatile organic compounds with an ESP32 node, streams the readings over secure MQTT to ThingsBoard Cloud, and closes a two-way loop in which a Python service estimates the unmeasured gases (NO2, SO2, CO, O3, NH3), computes the final AQI and person-specific health advice, and writes them back to the cloud. Seven regression families were compared per gas on CPCB hourly station data for Mumbai with a time-ordered split. NO2 was estimated with a test R² of 0.59 and 49% lower error than a median baseline, while SO2 and CO did not generalise to a test period overlapping the 2020 COVID-19 lockdown. Adding every estimated gas to the AQI reduced the AQI category accuracy from 99.9% to 82.6%; an evidence-gated composition step that admits a gas only if it improves validation accuracy preserved 99.9%. A 3-hour PM2.5 forecast reduced the error of a persistence baseline by about 10%. The results show that virtual sensing must be validated at the level of the final index, not only per pollutant. AI Disclosure: An AI assistant (Claude, Anthropic) was used to help draft and edit parts of the literature survey and methodology and to format references. All content, results and references were reviewed and verified by the authors, who take full responsibility for the work.