PuffConv: A System for Online and On-device Puff Detection for Smoking Cessation

Smoking Cessation is a vital wellness application as smoking has health issues pertaining to cancer, cardio-pulmonary diseases, hypertension, and diabetes. This paper presents a method for online and on-device puff detection on a microcontroller-based wearable device. We design a specialized Convolutional Neural Network (CNN) based model for puff detection from Respirational Inductance Photoplethysmogram (RIP) with a 6-axis IMU signal achieving 81% F1-score and provide an algorithm to quantify episodes based on the certainty of detected puffs. We use model reduction techniques, e.g., pruning, quantization, and intelligent data manipulation, to reduce our model and fit it to one target hardware. However, we find that creating models with different accuracy-size trade-offs for varying target platforms is often a brute force and tedious process. To address this, we present an automated model generation approach that takes the above-mentioned dataset and platform constraints as input and generates tailor-made, extremely small models for target micro-controller platforms. In summary, we provide a framework for rapidly developing sub-250 kB, accurate smoking puff detection models for wearable platforms with varying configurations.