Towards Passive Dietary Monitoring: Dilated CNN-Based Meal Detection Using Ambulatory High-Resolution Electrogastrography

BackgroundObjective tools for longitudinal dietary monitoring, despite its importance in the health triad of diet, sleep, and exercise, remain limited. Widely available ambulatory options include continuous glucose monitoring, which is a delayed response, and manual logging, which is rife with human error. Clinical tools such as gastric emptying scintigraphy are impractical for everyday use. High-resolution electrogastrography (HR-EGG) offers an alternative by treating gastric myoelectric activity as a biomarker of digestive state. However, its utility for ambulatory meal detection remains unclear. We hypothesize that HR-EGG and accelerometry together encode distinct postprandial gastric signatures to enable automated meal detection, and that postural context represents a relevant source of variation in signal detectability. MethodsHR-EGG and accelerometry data were collected from seven healthy adults across sixteen 150-minute meal sessions under IRB-approved protocol. Each session included a 30-minute fasted baseline, consumption of a standardized meal, and 90-minute postprandial period of sitting, walking, and lying in a randomized order. Features of the gastric slow wave, including raw and normalized bandpower, phase gradient directionality (PGD), wave direction, and wave speed, were extracted alongside triaxial accelerometer magnitude. A dilated one-dimensional convolutional network (1D CNN) was trained to classify meal consumption at five-minute resolution using leave-one-subject-out cross-validation. Postural effects on gastric myoelectric metrics were assessed using the Friedman test. ResultsThe model achieved a mean AUROC of 0.925 (95% CI: [0.857, 0.993]) and mean AUPRC of 0.824 (95% CI: [0.668, 0.980]; null model: 0.20). Feature ablation showed PGD as the most informative input ({Delta}AUPRC = -0.188), with wave propagation speed the least informative ({Delta}AUPRC = -0.105). Walking produced the highest signal-to-noise ratio (9.94 dB), lying had the most stable gastric rhythm (89.1% normogastric), and sitting demonstrated the greatest frequency instability (dominant frequency standard deviation = 0.825 cpm). ConclusionA dilated 1D CNN applied to spatiotemporal HR-EGG features enables temporally aware passive meal detection across ambulatory contexts. This study framework addresses a gap between clinical need for objective dietary monitoring and the limitations of current detection methods.