Enhancing Asthma Self-Management with Environmental Passive-Monitoring Data and Machine Learning-Based Predictions

Monitoring enables timely action which is critical in avoiding asthma attacks. With the abundance of local weather and pollution data, when augmented with machine learning, it is becoming possible to replace traditional tedious active monitoring in certain instances for some groups of patients. This analysis on the AAMOS-00 dataset, with 22 severe asthma patients, compared the performance of XGBoost to predict worsening asthma conditions using either passive or active monitoring on two groups of patients: individuals who were and were not triggered by outdoor environmental factors. On patients triggered by environmental factors, the performances of the passive- and the active-monitoring models were close to matching (passive: AUC = 0.83, 2.9-fold increase in precision; active: AUC = 0.89, 3.1-fold increase in precision). On the other hand, the active-monitoring model was significantly superior for patients not triggered by environmental factors (passive: AUC = 0.52, 1.2-fold increase in precision; active: AUC = 0.83, 3.7-fold increase in precision). Robustness evaluation showed it was possible to have good predictions of asthma attacks in the coming 5 to 7 days using 5 to 28 days of data.