Healthcare turns to AI to track patient nutrition

Hospital malnutrition remains one of the most under-addressed challenges in clinical care. A new clinical pilot study suggests that AI-powered image analysis could significantly improve how hospitals monitor patient food intake, a longstanding weak point in healthcare systems that contributes to widespread malnutrition, prolonged hospital stays, and rising costs.

The study, titled "Enhancing Hospital Nutrition Assessment Through Artificial Intelligence: A Prospective Tray-Level Pilot Study," was published in Nutrients. Conducted in an Italian hospital setting, the research evaluates whether AI-based image analysis can match or improve upon traditional methods such as manual weighing and nurse-recorded dietary diaries in assessing patient food consumption.

AI targets a hidden healthcare crisis: hospital malnutrition and poor monitoring

The study highlights that between 30% and 50% of hospitalized patients are affected by disease-related malnutrition, a condition closely linked to increased morbidity, longer hospital stays, and higher healthcare costs. In many cases, patients experience declining nutritional status during their hospital stay, even if they were well-nourished at admission.

A major problem lies in how food intake is monitored. Current hospital practices rely heavily on nursing dietary diaries, which typically record intake using simple categories such as none, half, or full consumption. While widely used, these methods are subjective, time-consuming, and often inaccurate due to workload pressures and inconsistent documentation practices. Nurses, despite playing a critical role in patient care, frequently overestimate food intake, particularly when patients eat slowly or irregularly.

These limitations are not trivial. Inaccurate intake records can delay the detection of malnutrition, reduce the effectiveness of nutritional interventions, and ultimately compromise patient outcomes. Traditional screening tools provide only a static snapshot of nutritional status and fail to capture daily fluctuations, which are often early indicators of clinical decline.

Against this backdrop, AI offers a potential solution. By analyzing images of meal trays before and after consumption, AI systems can automatically identify food items, estimate portion sizes, and calculate intake with greater consistency. This approach aims to transform nutritional monitoring from a subjective task into a data-driven process.

Pilot study reveals mixed but promising results for AI-based intake tracking

The study was conducted at the General Medicine Unit of St. Antonio Hospital in Padua, Italy, over a two-month period in 2025. Researchers analyzed 362 meals from 67 adult patients, comparing three methods of dietary assessment: gold-standard manual weighing, nursing dietary diaries, and an AI-based image analysis system.

The AI system relied on a custom-built imaging platform equipped with multiple high-resolution cameras capturing trays from different angles. As shown in the system setup on page 4, the device used a four-camera configuration to generate multi-view images, enabling more accurate volume estimation of food items. These images were processed using a deep learning model based on Mask R-CNN architecture, which segmented food items and converted visual data into weight estimates.

Results showed that traditional nursing documentation achieved only 60.8% concordance with gold-standard weighing, confirming widespread inaccuracies in routine monitoring. Errors were particularly pronounced when meals were recorded as fully consumed, with data often revealing partial intake instead. Variability in nurse-recorded estimates was also high, with discrepancies frequently exceeding 100 grams.

The AI system demonstrated more consistent performance at the tray level, with a mean absolute error of approximately 40 grams, equivalent to about 10% of the average tray weight. This suggests that AI-based estimates may offer greater precision than routine documentation in certain contexts. However, the system also showed a systematic tendency to underestimate intake, indicating that further refinement is needed.

Importantly, the study emphasizes that these findings are exploratory. The AI model was validated on a limited subset of real-world data, and the analysis was conducted at the tray level rather than the patient level. As a result, the researchers caution that the system's superiority over traditional methods cannot yet be established.

Performance also varied depending on the type of food. Items with clear shapes and textures, such as bread or packaged foods, were identified more accurately, while mixed or texture-modified meals posed significant challenges. These limitations are particularly relevant in hospital settings, where patients often require specialized diets, including pureed or soft foods.

High food waste and workflow barriers highlight broader system challenges

The study sheds light on another critical issue: hospital food waste. Across the 362 meals analyzed, more than 30% of food served was discarded, amounting to over 72 kilograms of waste during the study period. The distribution of waste, as illustrated in the chart on page 10, reveals that both common and therapeutic diets contribute significantly, with texture-modified meals showing the highest waste intensity per tray.

This level of waste reflects not only inefficiencies in hospital food systems but also unmet nutritional needs. Patients who do not consume enough food are at greater risk of malnutrition, creating a cycle of poor outcomes and increased healthcare costs. The findings suggest that improving intake monitoring could also support more targeted interventions to reduce waste and enhance patient nutrition.

However, integrating AI into hospital workflows presents practical challenges. The current system required trays to be transported to a dedicated imaging station before and after meals, adding to staff workload and disrupting routine processes. The study notes that future implementations should aim to embed imaging technology directly into meal delivery systems, such as carts or distribution stations, to enable seamless data collection.

The researchers also highlight organizational and human factors as key barriers to adoption. Successful implementation will depend on staff training, digital literacy, and acceptance of new technologies within clinical environments. Concerns about workflow disruption and professional autonomy may also influence uptake.

Despite these challenges, the study points to significant long-term potential. AI-based systems could eventually integrate with electronic health records, enabling real-time monitoring and automated alerts for patients at risk of malnutrition. This would mark a shift toward more proactive and personalized nutritional care.

From pilot evidence to clinical transformation: what comes next

While the study stops short of declaring AI a replacement for traditional methods, it provides compelling evidence that technology can play a supportive role in improving nutritional assessment. The findings align with broader trends in healthcare, where AI is increasingly used to enhance efficiency, reduce human error, and support decision-making.

The authors call for larger, multi-center studies to validate their results and address current limitations. Future research will need to include patient-level analysis, improve algorithm performance for complex meals, and evaluate the clinical impact of AI-assisted monitoring on outcomes such as recovery rates, hospital stay duration, and readmission.

There is also a need to explore cost-effectiveness and scalability. While AI systems may reduce workload and improve accuracy, their implementation requires investment in infrastructure, training, and integration with existing systems. Understanding the balance between costs and benefits will be critical for widespread adoption.

Overall, AI is a promising but still evolving tool in hospital nutrition care. Its ability to provide standardized, objective, and timely data could help address long-standing gaps in monitoring, particularly in high-pressure clinical environments where traditional methods fall short.