Your Watch Knows You’re Gasping: AI Turns Health Data into Pollution Warnings

Researchers from Imperial College London, including the Data Science Institute, the Department of Earth Science & Engineering, the Centre for AI-Physics Modelling at Imperial-X, the National Heart & Lung Institute, and the Department of Materials, have unveiled a groundbreaking AI-based framework to predict individual health responses to air pollution. Built under the AI-Respire project and leveraging the earlier INHALE initiative, this pioneering work brings together wearable health data, environmental exposure metrics, and deep learning models to create a powerful tool for personal health monitoring. Funded by EPSRC, NERC, and Innovate UK, among others, this system combines physiological data from smart devices with high-resolution pollution and weather data to forecast health responses tailored to each user.

The Pollution Crisis Demands Personalised Solutions

Air pollution is a persistent global health threat, contributing to an estimated seven million premature deaths each year. It worsens or triggers various diseases, notably asthma, COPD, and heart disease, and the situation is aggravated by climate change, which leads to more intense wildfires and hotter summers, increasing the levels of particulate matter and noxious gases. In the UK alone, man-made air pollution is responsible for as many as 36,000 deaths annually. While strategies exist to reduce general exposure, what’s missing is the ability for individuals to assess their personal risk in real time. The researchers identified that artificial intelligence, trained on personal health data and environmental inputs, could offer precisely this level of insight, forecasting how a specific person’s body reacts to changing pollution levels.

BreathBot and the Real-Time Health Data Revolution

At the core of this new system lies BreathBot, a mobile app developed to collect real-time physiological, environmental, and location-based data. The app connects with popular wearable devices such as Garmin and Fitbit, and integrates data from Apple Health and Android Health Connect. It collects key metrics like heart rate, breathing rate, step count, blood pressure, and user-reported health information. BreathBot also continuously logs GPS location data and synchronises with OpenWeather APIs to fetch air quality and meteorological details. The data pipeline is fortified by strong privacy and encryption protocols, including GDPR compliance, secure HTTPS transmission, and cloud storage via Amazon Web Services. This ensures the sensitive health information of each user remains safe while enabling high-frequency, high-resolution data capture crucial for health prediction.

Building the AI Model: From Population to Personal

To train their predictive engine, the researchers used data from the INHALE project, which had monitored 59 individuals using wearable devices across the summer and winter seasons. For the AI-Respire model, a subset of 10 healthy participants provided hourly physiological and environmental readings, creating a rich dataset that captured temporal and seasonal variations. The AI model itself is an Adversarial Autoencoder (AAE), further enhanced with convolutional layers to capture spatial patterns and LSTM layers for modelling time-based trends. This hybrid neural network can learn and reconstruct health signals from noisy, irregular, or partially missing data using a technique called inpainting. The model achieved a Mean Squared Error (MSE) of 0.0029 when predicting breathing rate values, indicating strong performance in understanding and forecasting individual physiological responses.

Testing Pollution Impact and Enhancing the Model

To evaluate how pollution influences health metrics, the researchers manipulated the input data to simulate increased and decreased pollution levels. Minor increases (20% or 50%) in particulate matter and gases like NO₂ and O₃ had little effect on predicted breathing and heart rates. However, doubling the pollution resulted in clear changes; breathing rate increased by about 3.5%, while heart rate went up by 2.5%. This suggests that for healthy individuals, the body might not respond drastically to short-term pollution fluctuations unless thresholds are significantly breached. When pollution levels were decreased, even a 100% drop did not lead to measurable improvements in respiratory metrics, possibly reflecting the body’s slower recovery rate or the influence of other ambient stressors. The team also applied polynomial feature transformation, expanding the dataset from 29 to 462 features, allowing the model to detect complex, nonlinear relationships between environmental exposure and physiological outcomes.

Personalisation with Transfer Learning

One of the most powerful elements of the framework is its use of transfer learning to personalise predictions. A model trained on INHALE data was adapted using eight months of smartwatch data from a single user. Only the final layers of the network were fine-tuned, while early layers remained unchanged, preserving the foundational knowledge. The retrained model, tested on unseen data, achieved an impressively low MSE of 4.24 × 10⁵. It accurately captured daily and weekly behavioural cycles, physiological trends, and their interactions with environmental conditions. This shows that the AI-Respire system can adapt seamlessly to new individuals, making it a practical and efficient tool for real-world deployment. It also opens the door to scaling this framework across diverse demographics and health profiles without retraining the entire model from scratch.

The AI-Respire framework stands out as a highly promising tool for precision health monitoring in an era of rising environmental risks. It combines real-time data capture, privacy-preserving cloud architecture, and powerful machine learning algorithms to offer individuals a window into how their bodies respond to the air they breathe. The ability to personalise predictions, quantify health risks from pollution, and potentially trigger preventive actions offers a transformative advancement in public health. Future research will aim to apply the model to broader populations, identify pollution-specific biomarkers, and improve risk-scoring through uncertainty modelling. As air pollution continues to threaten health worldwide, tools like AI-Respire may be the key to bridging the gap between environmental exposure and individual well-being.