Collaborative design and feasibility assessment of computational nutrient sensing for simulated food-intake tracking in a healthcare environment
MetadataShow full item record
One in four older adults (65 years and over) are living with some form of malnutrition. This increases their odds of hospitalization four-fold and is associated with decreased quality of life and increased mortality. In long-term care (LTC), residents have more complex care needs and the proportion affected is a staggering 54% primarily due to low intake. Tracking intake is important for monitoring whether residents are meeting their nutritional needs however current methods are time-consuming, subjective, and prone to large margins of error. This reduces the utility of tracked data and makes it challenging to identify individuals at-risk in a timely fashion. While technologies exist for tracking food-intake, they have not been designed for use within the LTC context and require a large time burden by the user. Especially in light of the machine learning boom, there is great opportunity to harness learnings from this domain and apply it to the field of nutrition for enhanced food-intake tracking. Additionally, current approaches to monitoring food-intake tracking are limited by the nutritional database to which they are linked making generalizability a challenge. Drawing inspiration from current methods, the desires of end-users (primary users: personal support workers, registered staff, dietitians), and machine learning approaches suitable for this context in which there is limited data available, we investigated novel methods for assessing needs in this environment and imagine an alternative approach. We leveraged image processing and machine learning to remove subjectivity while increasing accuracy and precision to support higher-quality food-intake tracking. This thesis presents the ideation, design, development and evaluation of a collaboratively designed, and feasibility assessment, of computational nutrient sensing for simulated food-intake tracking in the LTC environment. We sought to remove potential barriers to uptake through collaborative design and ongoing end user engagement for developing solution concepts for a novel Automated Food Imaging and Nutrient Intake Tracking (AFINI-T) system while implementing the technology in parallel. More specifically, we demonstrated the effectiveness of applying a modified participatory iterative design process modeled from the Google Sprint framework in the LTC context which identified priority areas and established functional criteria for usability and feasibility. Concurrently, we developed the novel AFINI-T system through the co-integration of image processing and machine learning and guided by the application of food-intake tracking in LTC to address three questions: (1) where is there food? (i.e., food segmentation), (2) how much food was consumed? (i.e., volume estimation) using a fully automatic imaging system for quantifying food-intake. We proposed a novel deep convolutional encoder-decoder food network with depth-refinement (EDFN-D) using an RGB-D camera for quantifying a plate’s remaining food volume relative to reference portions in whole and modified texture foods. To determine (3) what foods are present (i.e., feature extraction and classification), we developed a convolutional autoencoder to learn meaningful food-specific features and developed classifiers which leverage a priori information about when certain foods would be offered and the level of texture modification prescribed to apply real-world constraints of LTC. We sought to address real-world complexity by assessing a wide variety of food items through the construction of a simulated food-intake dataset emulating various degrees of food-intake and modified textures (regular, minced, puréed). To ensure feasibility-related barriers to uptake were mitigated, we employed a feasibility assessment using the collaboratively designed prototype. Finally, this thesis explores the feasibility of applying biophotonic principles to food as a first step to enhancing food database estimates. Motivated by a theoretical optical dilution model, a novel deep neural network (DNN) was evaluated for estimating relative nutrient density of commercially prepared purées. For deeper analysis we describe the link between color and two optically active nutrients, vitamin A, and anthocyanins, and suggest it may be feasible to utilize optical properties of foods to enhance nutritional estimation. This research demonstrates a transdisciplinary approach to designing and implementing a novel food-intake tracking system which addresses several shortcomings of the current method. Upon translation, this system may provide additional insights for supporting more timely nutritional interventions through enhanced monitoring of nutritional intake status among LTC residents.
Cite this version of the work
Kaylen Pfisterer (2021). Collaborative design and feasibility assessment of computational nutrient sensing for simulated food-intake tracking in a healthcare environment. UWSpace. http://hdl.handle.net/10012/17712