An AI-driven framework for the prediction of personalised health response to air pollution

TL;DR

This work presents AI-Respire, a modular cloud-based framework that predicts personalised physiological responses to air pollution by integrating wearable-derived health data with real-time environmental exposures. The core model is an Adversarial Autoencoder enhanced with LSTM and CNN components, pre-trained on the INHALE dataset and fine-tuned via transfer learning using smartwatch data to capture individual-specific patterns. Evaluation shows modest but detectable shifts in breathing and heart rate under simulated pollution spikes, with cross-cohort validation in U-BIOPRED linking these shifts to asthma burden and FeNO, supporting physiological relevance. The secure Data Management Platform enables scalable, GDPR-compliant data integration and supports real-world deployments for anticipatory, personalised environmental health monitoring and interventions.

Abstract

Air pollution is a growing global health threat, exacerbated by climate change and linked to cardiovascular and respiratory diseases. While personal sensing devices enable real-time physiological monitoring, their integration with environmental data for individualised health prediction remains underdeveloped. Here, we present a modular, cloud-based framework that predicts personalised physiological responses to pollution by combining wearable-derived data with real-time environmental exposures. At its core is an Adversarial Autoencoder (AAE), initially trained on high-resolution pollution-health data from the INHALE study and fine-tuned using smartwatch data via transfer learning to capture individual-specific patterns. Consistent with changes in pollution levels commonly observed in the real-world, simulated pollution spikes (+100%) revealed modest but measurable increases in vital signs (e.g., +2.5% heart rate, +3.5% breathing rate). To assess clinical relevance, we analysed U-BIOPRED data and found that individuals with such subclinical vital sign elevations had higher asthma burden scores or elevated Fractional Exhaled Nitric Oxide (FeNO), supporting the physiological validity of these AI-predicted responses. This integrative approach demonstrates the feasibility of anticipatory, personalised health modelling in response to environmental challenges, offering a scalable and secure infrastructure for AI-driven environmental health monitoring.

Figures (5)

• Figure 1: Overview of the AI-Respire data pipeline and model development workflow. This schematic illustrates the end-to-end architecture developed in AI-Respire for predicting personalised health responses to air pollution. The pipeline integrates two datasets: (1) an offline dataset from the INHALE project containing air pollution exposure, respiratory health, environmental, and physical activity data; and (2) an online dataset collected in real-time via the BreathBot mobile app, including smartwatch-derived physiological data, GPS data, and pollution data synchronised through OpenWeather. Model #1, an Adversarial Autoencoder (AAE), is pre-trained on the INHALE dataset to learn general patterns of pollution-health interactions. Using transfer learning, Model #1 is then fine-tuned with personalised data (Dataset #2) to yield Model #2, which captures individual-level responses. Both models are deployed within a secure Data Management Platform (DMP), supporting predictions that feed back to the user (e.g., via the mobile app) for potential health insights or interventions. The system architecture ensures continuous data flow via secure AWS channels and supports scalable, ethical AI-driven personalised health monitoring.
• Figure 2: Data Management Platform (DMP) architecture in the AI-Respire project. It includes four core components: a secure web interface, backend processing engine, analytical environment, and data storage. The storage system supports raw and integrated data, metadata, and model outputs, enabling scalable, traceable, and secure AI-driven health analysis.
• Figure 3: Overview of the AAE architecture for time-series prediction of pollution-linked health outcomes. The adversarial training is achieved through a discriminator which takes input either by sampling a prior distribution or from the latent space. Colour indicates the type of layers, the key for which is seen at the bottom of the image.
• Figure 4: Comparison of actual and predicted normalised breathing rate ( br_avg) values over half-hourly intervals.
• Figure 5: Comparison of predicted and actual trends for average breath rate (top) and heart rate (bottom) under baseline and 100% increased pollution levels.