3.2.1. AI in Clinical Nutrients Intake
Among the identified studies on the application of AI in clinical practice, there is a need to distinguish those that aimed to develop systems that monitor, support and modulate the nutrition of chronically ill people. Lu et al. presented a novel system based on AI to accurately estimate nutrient intake, by simply processing RGB depth image pairs captured before and after meal consumption [43]. Oka et al. compared AI-supported nutrition therapy with a mobile application (n = 50) versus human nutrition therapy (n = 50) in a randomized controlled trial [44]. An interesting technological solution in the AI area was used by Vasiloglou et al. in relation to the clinical problem of controlling carbohydrate intake in patients with type 1 diabetes. These authors used GoCARB as a computer vision-based smartphone system in determining plated meals’ carbohydrate content. In this study, the estimation of carbohydrate content in 54 plated meals made by GoCARB was compared to the estimation made by six experienced dietitians. It was found that GoCARB estimated the carbohydrate content with the same accuracy as professional nutritionists (p = 0.93) [45].
Chin et al. tested the Automated Self-Administered 24-Hour Dietary Assessment Tool (ASA24) on the example of lactose with regard to the Nutrition Data System for Research (NDSR) [46]. ASA24, also known as food diaries, is a web-based tool that enables multiple, automatically coded, self-administered 24-h diet recalls. NDSR is a dietary analysis software application widely used for the collection and coding of 24-h dietary recalls and the analysis of menus. Nine machine learning models have been developed based on the nutrients common to ASA24 and the NCC database. The results obtained in this study suggest that computational methods can successfully estimate an NCC-exclusive nutrient for foods reported in ASA24.
In order to monitor eating behaviors, a rapid automatic bite detection algorithm (RABID) that extracts and processes skeletal features from videos was constructed. Konstantinidis et al. used it to analyze the eating behaviors of n = 59 patients (three types of dishes, 45 meals), the results of which showed an agreement between algorithmic and human annotations (Cohen’s kappa κ = 0.894; F1-score: 0.948) [47].
Chi et al. proposed a knowledge-based system (KBS) for patients with chronic kidney disease using the Web Ontology Language (OWL) and the Semantic Web Rule Language (SWRL) [48]. In order to evaluate the designed system in recommending appropriate food serving amounts from different food groups, information was collected from n = 84 patients. It was found that the OWL-based KBS can achieve accurate problem solving and reasoning questions while maintaining the ability to share and extend the knowledge base.
AI techniques can also be useful in diagnosing mild dehydration. Posada-Quintero et al., using machine learning, investigated the possibility of detecting mild dehydration with autonomic responses to cognitive stress (n = 17) [49]. Taking into account the autonomic control indicators based on electrodermal activity (EDA) and pulse rate variability (PRV) in the Stroop test, they obtained 91.2% overall accuracy of mild dehydration detection.
In the area of AI applications in the improvement of dietary solutions, two articles describing prototype solutions should be mentioned. Khan and Hoffmann proposed a menu construction using an incremental knowledge acquisition system (MIKAS) [50]. This system asks the expert to provide an explanation for each of their actions, in order to include the explanation in its knowledge base, so MIKAS could in the future automatically perform them.
Fuzzy arithmetic has been used to create “Nutri-Educ”—software for proper balancing of meals, according to the energy needs of the patient. Heuristic search algorithms are used to find a set of actions, acceptable from a nutritional point of view, that will transform the initial meal into a well-balanced one [51].
Baek et al. applied the hybrid clustering-based food recommendation method that uses chronic disease-based clustering and a nutrition knowledge base [52]. Food products are grouped using the k-means algorithm and food and nutrient data system. Based on the created clusters and data on food preferences, a knowledge base on diet and nutrition is generated.
Mezgec and Koroušić Seljak introduced a new “NutriNet” tool for food image recognition based on a deep convolutional neural network architecture [53]. It was tested on a collection of 225,953 images (512 × 512 pixels) of 520 different foods and beverages. This tool with an implemented training component is used in practice as a part of a mobile app for the dietary assessment of Parkinson’s disease patients.