Estimating Caloric Intake in Bedridden Hospital Patients with Audio and Neck-Worn Sensors

We present an approach for estimating calorie intake given a limited number of foods provided to patients in an in-bed setting. Data collected from a proximity sensor, inertial measurement unit, ambient light, and audio sensor placed around the neck are used to classify food-type consumed by second using a random forest classifier. A multiple linear regression model is then developed for each food-type to map second-level features to calories per second. We conducted a user study in a patient simulated lab setting, where 10 participants were asked to eat while sitting on a patient bed. A user-independent analysis demonstrated food-type detection at 97.2% F1-Score, and an average Absolute Error of 3.0 kCal per food-type. Our system shows promise in distinguishing food items and predicting calorie intake in a bedridden participant setting given a limited set of food items.