Assessing real-life food consumption in hospital with an automatic image recognition device: A pilot study

Accurate dietary intake assessment is essential for nutritional care in hospitals, yet it is time-consuming for caregivers and therefore not routinely performed. Recent advancements in artificial intelligence (AI) offer promising opportunities to streamline this process. This study aimed to evaluate the feasibility of using an AI-based image recognition prototype, developed through machine learning algorithms, to automate dietary intake assessment within the hospital catering context.

Data were collected from inpatient meals in a hospital ward. The study was divided in two phases: the first one focused on data annotation and algorithm's development, while the second one was dedicated to algorithm's evaluation. Six different dishes were analyzed with their components grouped into three categories: cereals and starchy food, meat and fish, and vegetables. Manual weighing (MAN) was used as the reference method, while the AI-based prototype (PRO) automatically estimated component weights. Lin's concordance correlation coefficients (CCC) were calculated to assess agreement between PRO and MAN. Linear regression models were applied to estimate measurement differences between PRO and MAN for each category and their associated 95 % confidence intervals (CI).

A total of 246 components were used for data annotation and 368 for testing. CCC values between PRO and MAN were: Cereals and starchy food (n = 219; CCC = 0.957, 95 % CI: 0.945-0.965), meat and fish (n = 114; CCC = 0.845, 95 % CI: 0.787-0.888), and vegetables (n = 35; CCC = 0.767, 95 % CI: 0.604-0.868). Mean differences between PRO and MAN measurements were estimated at -12.01g (CI 95 % -15.3, -8,7) for cereals and starchy food (reference category), 1.19 g (CI 95 % -3.2, 5.6) for meat and fish, and -14.85 (CI 95 % -22.1, -7.58) for vegetables.

This pilot study demonstrates an AI-based system to assess food types and portions in a hospital setting. Further improvements are necessary before the system can be reliably used in direct patient care.