Predicting malnutrition in colorectal cancer: a systematic review of risk prediction models

Colorectal cancer, a disease that affects the large intestine and rectum, remains one of the most common and serious health challenges worldwide. For patients navigating this illness, a hidden but critical companion often emerges: malnutrition. This condition is not merely about being underweight; it is a complex state where the body lacks the essential nutrients needed to repair tissue, fight infection, and tolerate the rigors of treatment. When a patient with colorectal cancer becomes malnourished, the consequences are severe. They face higher risks of infection after surgery, longer stays in the hospital, and a reduced ability to withstand chemotherapy. Because of this, medical teams, particularly nurses who are often the first to notice a patient's decline, need a reliable way to spot those at risk before it is too late. The goal is to identify the vulnerable early, so that help can be given before the body's reserves are exhausted.

For years, researchers have tried to build mathematical tools to predict which patients will fall into this dangerous state. These tools, known as prediction models, are like checklists that weigh various factors—such as age, tumor size, blood test results, and physical strength—to calculate a risk score. The hope has been that a nurse could use such a tool to instantly flag a patient who needs immediate nutritional support. However, the landscape of these tools has been messy. Some models predicted different things, like survival rates or specific blood markers, rather than the actual state of malnutrition. Others used different definitions for what counts as "malnourished," making it impossible to compare them fairly. To cut through this confusion, a team of researchers from universities and hospitals in China and the United States set out to gather every available study on this topic and examine them with a critical eye.

The researchers conducted a comprehensive search, looking through eight major scientific databases for studies published up to May 1, 2026. They cast a wide net, seeking any study that built a model to predict malnutrition or nutritional risk specifically in adults with colorectal cancer. They were strict about what they included: the models had to focus on the clinical reality of malnutrition, defined by recognized medical standards, rather than just a single blood test result or a guess at how long a patient might live. After sifting through thousands of records and removing duplicates or irrelevant papers, they found thirteen studies that met their criteria. These studies involved a total of thousands of patients, ranging from those just diagnosed to those recovering from surgery or undergoing chemotherapy.

When the team analyzed these thirteen models, they found that on paper, they looked quite promising. The models were able to distinguish between patients who were at risk and those who were not with a high degree of accuracy. In the language of statistics, their ability to separate the two groups ranged from good to excellent. The researchers noted that some models correctly identified the at-risk patients with an AUC of 0.88, while others were even more precise in their initial tests. The tools used a variety of methods to make these predictions, from simple equations to more complex computer algorithms. They relied on common clues, such as a patient's age, their body mass index, the stage of their cancer, their hemoglobin levels, and how well they could perform daily activities. Some models also looked at specific nutritional screening scores or factors like income and protein intake.

However, the story takes a turn when the researchers looked deeper into how these models were built and tested. Despite the impressive numbers on the surface, the team found that every single one of the thirteen models suffered from significant flaws in their design and reporting. The primary issue was that the models had not been tested on new, different groups of people. Most were built and tested on the same group of patients, a bit like a student taking a practice exam and then immediately taking the same test again. This approach often leads to an overly optimistic view of how well the tool works. The researchers found that the studies often failed to explain how they handled missing data, how they chose which factors to include, or whether the model was prone to overfitting, a technical term for memorizing the specific details of the training data rather than learning the general rules. Because of these issues, the team concluded that the models were not yet ready to be used in real-world hospitals to make decisions about patient care.

The review also highlighted a lack of diversity in the research. Almost all the studies were conducted in China, with only one from the United States. This means the tools were built using data from specific populations with particular dietary habits, healthcare systems, and genetic backgrounds. A model that works well in one region might not work at all in another. Furthermore, the definitions of malnutrition varied slightly between studies, with some using one set of criteria and others using a different set. While the researchers managed to narrow the focus to clinically defined malnutrition, this variation still made it difficult to combine the results into a single, unified picture. The team noted that while machine learning and artificial intelligence were being used in some of the newer models, these advanced techniques did not fix the fundamental problems of poor study design or lack of external testing.

The final message from this review is one of cautious optimism mixed with a call for rigor. The existing models show that it is possible to predict nutritional risk in colorectal cancer patients, and the initial results are encouraging. Yet, the path to a tool that a nurse can confidently use at a patient's bedside is not yet complete. The researchers emphasize that these models should be viewed as preliminary ideas rather than finished products. Before they can be trusted to guide clinical decisions, they need to be tested in large, diverse groups of patients across different hospitals and countries. Future studies must be designed with greater care, ensuring that the models are tested on new people to prove they work in the real world. Until that happens, the best approach remains for healthcare teams to use their professional judgment and established screening methods, keeping these new tools in mind as potential aids for the future rather than immediate solutions. The work is far from over, but this systematic review has provided a clear map of where the field stands and what steps must be taken next to turn these promising calculations into life-saving care.