Bachelor's DegreesBachelor of ArtsBachelor of EngineeringDual-Degree ProgramUndergraduate Admissions
Doctoral DegreesDoctor of PhilosophyPhD Innovation ProgramPhD + Doctor of MedicineGraduate Admissions
All Thayer Events
MS Thesis Defense: Dhanashree Vaidya
1:00pm - 2:00pm ET
For info on how to attend this videoconference, please email dhanashree.vaidya.TH@dartmouth.edu
"Towards a wearable dietary monitoring system: Nutrition classification and user gesture detection using machine learning"
Monitoring nutritional intake is becoming more important by the day due to the rise in obesity and other eating related disorders. Understanding the relationship between diet and various health disorders is of the utmost importance in terms of prevention and treatment. This prompts the need for a method to closely monitor eating habits and nutritional intake. Traditionally, manual food logging has been used for monitoring; however, this approach is prone to errors and bias. Most automatic diet monitoring approaches developed to date focus on detecting the amount and duration of food intake and fail to track the actual nutritional content of food consumed. In this thesis, an electrical impedance spectroscopy-based method is proposed for automatic nutrition monitoring. A wearable is designed to conduct bioimpedance spectroscopy by making measurements around the neck to detect the nutrition of food being swallowed. In addition, user interaction with the wearable is explored.
This thesis primarily focuses on investigating the application of electrical impedance spectroscopy for nutritional analysis of food. Of the two frameworks proposed, the first categorizes a food sample into 7 nutritional categories using ensemble classification models. The second approach estimates the amount of different nutrients in a particular food sample to give a nutritional breakdown. These two frameworks are used to classify the dominant nutritional category and predict the nutritional breakdown of in-vitro food bolus formed using 52 distinct food items. The dominant nutritional categories can be predicted with an F1 score in the 90's for fats, alcohol, and fruits and vegetables. For the other categories, an F1 score in the 70's is achieved. The nutritional breakdown can be estimated with high accuracies ranging from 79% to 97% for different nutrients.
For automatic nutrition monitoring a prototype for the wearable was developed that can make bioimpedance measurements around the neck. The best electrode placement selection and impedance change estimation is verified using finite element simulations. Additionally, a template matching algorithm is analyzed for user interaction with the wearable using several head-based gestures which can detect head nods and head shakes with high accuracy.
- Ryan Halter, PhD (Chair)
- Kofi Odame, PhD
- Eric Hansen, PhD
For more information, contact Daryl Laware at firstname.lastname@example.org.