Nursing Assessment–Derived Risk Profiles and VALENF Instrument–Based Classification in Adult Inpatients: A Two-Stage Machine Learning Study

Hospitals are places where the complexity of human illness often hides behind simple labels. A patient arrives with a diagnosis, but that single medical term rarely tells the whole story of what they need. Two people might have the same disease, yet one is strong enough to walk and eat independently, while the other is frail, confused, and at high risk of falling or developing bedsores. For decades, hospital records have focused heavily on medical diagnoses and administrative codes, which are useful for billing and tracking diseases but often miss the daily reality of a patient's physical condition and safety needs. Nurses, however, are trained to see this missing picture. Every day, they assess how well a patient can move, how likely they are to slip, and whether their skin is at risk of breaking down. These observations create a rich, detailed map of a patient's vulnerability that standard medical records often overlook.

The challenge has always been how to turn these detailed, human observations into a system that can help doctors and nurses make faster, better decisions for thousands of patients. If a nurse can spot a pattern of risk in one patient, can a computer learn to spot that same pattern in the next? This is the question a team of researchers in Spain set out to answer. They wanted to see if they could use the specific, structured notes nurses write down when a patient arrives to group patients into clear categories of risk. Their goal was not just to sort patients, but to build a digital tool that could automatically assign a new patient to the right risk group based on a short, routine nursing check, ensuring that the most vulnerable people get the attention they need without delay.

The researchers worked with data from a hospital in Spain, looking at thousands of hospital stays. They started by gathering a massive collection of records from over 2,600 patients. For each of these patients, they had access to the full, detailed nursing assessments taken when they were admitted. These assessments included scores for how well a patient could perform daily tasks like eating and walking, how likely they were to develop pressure sores, and their risk of falling. They also included the patient's age, sex, and a score that counted how many other health problems they had. Using a computer method that looks for hidden patterns without being told what to look for, the team let the data speak for itself. They asked the computer to find natural groups among the patients based on these combined factors.

The computer found five distinct groups, or profiles, that made sense to human experts. The first group consisted of younger patients who were very healthy and independent, with almost no risk of falling or getting bedsores. The second group was similar in terms of physical independence but was made up of older patients with more chronic health problems. The third group represented patients with moderate difficulties; they could still do some things for themselves but were clearly struggling. The fourth group included patients who were significantly dependent on others for care and faced high risks. Finally, the fifth group was the most vulnerable: very elderly patients with many health issues who were completely dependent on others for movement and had the highest risks of falls and skin damage. These five categories created a clear ladder of risk, moving from those who needed little help to those who needed constant, intensive care.

Once these five groups were established, the researchers faced a new task. They needed a way to assign new patients to these groups without having to perform the full, time-consuming detailed assessment every single time. They turned to a shorter tool called the VALENF instrument, which is a seven-item checklist that nurses use routinely. This checklist asks simple questions that estimate the same risks as the longer, more detailed tests. The team trained a computer program to look at the answers to these seven questions and predict which of the five risk groups a patient belonged to. They tested several different types of computer learning algorithms, trying to find the one that was most accurate and, crucially, the one that made the safest mistakes.

In the world of risk assessment, a mistake can go two ways: the computer might think a patient is safe when they are actually at risk, or it might think a patient is at risk when they are actually safe. The researchers found that one specific type of algorithm, known as a random forest, was the best choice. While another method was slightly more accurate overall, the random forest was better at being cautious. It was more likely to assign a patient to a higher-risk group if there was any doubt, rather than accidentally putting a vulnerable person into a low-risk category. This "safety-first" approach is vital in a hospital setting, where missing a risk can have serious consequences. The chosen computer model correctly assigned patients to their risk groups about 82 percent of the time during testing, a level of performance that the researchers considered strong enough to move forward.

To see if this system would work in the real world, the team tested it on a completely new set of patients, involving more than 8,000 hospital stays recorded a few years later. They did not change the computer program at all; they simply fed it the new data to see what happened. The results were encouraging. The system successfully sorted the new patients into the five risk groups, and the overall pattern of risk remained consistent. The patients in the high-risk groups were indeed older, had more health problems, and were less mobile than those in the low-risk groups. The computer's assignments matched the clinical reality of the patients' conditions in the vast majority of cases.

However, the study also revealed some important nuances. The system worked best for the patients in the middle and high-risk categories. For the two groups of patients who were generally healthy, the computer sometimes shifted them between the "low risk" and "low risk with health problems" categories. This happened because the new group of patients was slightly older and had slightly more chronic health issues than the original group the computer was trained on. While the computer handled the complex, high-risk cases with great stability, the subtle differences in the healthier population showed that the system is sensitive to changes in the patient mix. The researchers noted that while the tool is technically sound and clinically logical, it is not yet a finished product ready for every hospital.

The study concludes that it is possible to translate the complex, multidimensional picture of a patient's health into a simple, automated risk profile using routine nursing data. The five groups identified provide a clear, ordered way to understand patient vulnerability, moving beyond simple medical diagnoses to capture the reality of daily function and safety. The computer tool developed to assign these profiles works well, particularly for identifying those who need the most care. Yet, the researchers are careful to state that this is a starting point, not a final solution. Before such a system can be used to make decisions about patient care in hospitals, it needs to be tested in many different hospitals and watched over time to ensure it helps improve outcomes and does not create new problems. The work proves the idea is feasible and the data is there, but the journey from a successful computer experiment to a trusted part of daily nursing care requires further proof and careful planning.