AI-driven IMU-based monitoring system for detecting dry powder inhaler misuse

Abstract

Inhaler misuse has been a persistent challenge for achieving effective pulmonary drug delivery with dry powder inhalers. Among the various misuse patterns, improper device maneuvering, especially incorrect device orientation, is one frequently overlooked yet clinically relevant misuse that impairs therapeutic efficacy. Currently, effective methods for timely and reliable detection of such misuse remain limited. To address this gap, we developed an AI-powered, motion-based digital monitoring system that could detect orientation-related misuse and other commonly observed inhaler-use errors. Specifically, a digital monitoring module was integrated into a Breezhaler ® device using inertial measurement unit (IMU) sensors to capture orientation and movement signals during inhaler use. The collected data were processed and analyzed using two feature selection strategies, Boruta and Recursive Feature Elimination with Cross-Validation (RFECV), which were, for the first time, jointly applied to IMU-derived inhaler motion data to identify the most discriminative features. These selected features were then modeled using five machine learning algorithms: AdaBoostClassifier, ExtraTreesClassifier, Support Vector Classifier (SVC), XGBClassifier, and LogisticRegression. The analyses identified distinct optimal feature sets under Boruta (100 features) and RFECV (66 features) strategies, and revealed differences in model performance across predefined inhalation-related activities. The best-performing models, ExtraTrees classifier with Boruta and SVC with RFECV, illustrated excellent performance with accuracy of 0.967 and 0.950, respectively, as evaluated under a five-fold cross-validation scheme. Furthermore, SHAP-based interpretability analysis was conducted to elucidate feature importance and model decision patterns. In summary, this study demonstrates the feasibility of leveraging machine learning for accurate recognition and classification of inhalation activities and highlights the potential of this approach for future application in improving inhaler use and adherence.