A Multimodal Machine Learning Model for Predicting Prolonged Hospital Stay After Total Hip and Knee Arthroplasty：Development, Internal Validation, and External Validation in MIMIC-IV

Hospitals are vast, complex ecosystems where the goal is always to heal, but the path to recovery is rarely the same for two different people. When a patient undergoes major joint replacement surgery, such as swapping out a worn hip or knee, the medical team faces a critical question: how long will this person need to stay in the hospital? A stay that stretches too long is not just a financial burden on the healthcare system; it increases the patient's risk of infections and other complications. For decades, doctors have tried to predict these extended stays using a checklist of standard facts: the patient's age, their blood test results, and the severity of their other medical conditions. These numbers tell a story, but they often leave out the most detailed parts of a patient's life and history, which are frequently buried in the written notes and diagnostic codes that accompany their medical records.

A team of researchers has now built a new kind of digital tool designed to read both the numbers and the words to make a better prediction. By teaching a computer to look at standard medical data alongside the unstructured text found in patient records, they created a system that can spot the signs of a difficult recovery much earlier than traditional methods. This approach does not rely on guessing or intuition; instead, it learns from thousands of past cases to find patterns that human eyes might miss. The result is a clearer picture of who is likely to face a long hospital stay, allowing doctors to prepare better and potentially shorten the time a patient spends in the hospital.

The researchers began their work by gathering data from a large database of patients who had undergone primary total hip or knee arthroplasty at a major hospital in South Korea. They focused on 2,647 adults who had their first-time joint replacement surgery. To define what a "prolonged" stay looked like, they set a specific benchmark: any hospitalization lasting seventeen days or more. This threshold represented the top 25 percent of the longest stays in their group, a point where the clinical and economic costs of hospitalization become significant. The team then fed their computer model two distinct types of information. The first type consisted of structured data, which are the neat, organized facts doctors routinely collect, such as age, gender, body mass index, and specific blood test results like albumin and hemoglobin levels. The second type was unstructured text, derived from the ICD-10 diagnostic codes. These codes are the standard shorthand doctors use to list every condition a patient has, from arthritis to high blood pressure. The researchers treated these codes as a narrative, converting the list of diagnoses into a format the computer could analyze for meaning.

To build their prediction engine, the team trained nine different machine learning algorithms, which are computer programs designed to learn from data rather than follow rigid, pre-written rules. They tested these models to see which one could best distinguish between patients who would leave the hospital quickly and those who would stay for seventeen days or longer. The most successful model was a sophisticated system called LightGBM, which proved to be exceptionally good at finding the subtle connections between a patient's history and their recovery time. When tested on a portion of the data the model had never seen before, it achieved a high level of accuracy, correctly identifying the risk of a long stay with an AUC of 0.9160. This performance was significantly better than models that relied only on the standard blood tests and demographics, proving that the extra layer of information from the diagnostic text added real value.

One of the most revealing aspects of the study was discovering exactly what the computer found most important. When the researchers asked the model to explain its reasoning, they found that six of the top twenty most influential factors came directly from the diagnostic text rather than the blood work. While factors like body mass index and low levels of albumin in the blood were indeed strong predictors, the text analysis highlighted specific conditions that traditional checklists often overlook. For instance, the presence of knee osteoarthritis, unspecified pain, and other forms of joint disease emerged as powerful signals that a patient might need more time to recover. This suggests that the specific combination of conditions a patient carries, as written in their record, tells a more complete story than the numbers alone. The model also confirmed that the duration of the surgery itself was a key factor, reflecting the complexity of the procedure.

To ensure their findings were not just a lucky match for one specific group of patients, the researchers took their model to a completely different database from the United States, known as MIMIC-IV. This database contains records from a large intensive care unit in Boston and represents a different healthcare system, a different time period, and a different population. Initially, the model's performance dipped when applied to the entire pool of patients in this new database, which included many people who had not undergone joint replacement surgery. However, when the researchers filtered the data to look only at the patients who had actually received hip or knee replacements, the model's accuracy soared. In this specific group, the model performed even better than it had in the original study, achieving an AUC of 0.9396. This dramatic improvement in an independent, external setting suggests that the tool is robust and can adapt to different environments, provided it is used on the right group of people. It is important to note that for this external validation, certain variables like body mass index were missing from the database and had to be estimated by the researchers, yet the model still demonstrated exceptional performance.

The study also examined how much data was needed to make the model reliable. By testing the system with increasing amounts of patient records, the researchers found that the model learned quickly and stabilized once it had reviewed about two-thirds of the available training data. This indicates that the tool does not require an impossible amount of information to function effectively and that it is not simply memorizing the data but learning genuine patterns. The ability of the model to generalize so well to a new, independent database is a strong sign that it has captured fundamental truths about recovery after joint surgery, rather than just the quirks of a single hospital's records.

Despite these successes, the researchers are careful to note the boundaries of their work. The model was built using data from a single institution in South Korea and validated in a single intensive care unit in the United States, meaning it has not yet been tested across a wide variety of global healthcare systems. Additionally, some important information, such as the exact body mass index or the specific classification of a patient's overall health, was missing from the external database and had to be estimated, which could introduce some uncertainty. The researchers also point out that their definition of a "long stay" was based on the specific distribution of their initial group, which might not apply perfectly to every hospital in the world. They emphasize that while the tool is powerful, it is a starting point for future research rather than a final solution.

The ultimate goal of this work is to move beyond simple prediction and toward action. By identifying patients who are at high risk for a prolonged stay before they even enter the operating room, doctors can intervene earlier. They might adjust a patient's medication, arrange for additional physical therapy, or coordinate with social workers to ensure a smooth transition home. The study demonstrates that the information needed to make these decisions is already sitting in patient records, waiting to be read correctly. By combining the hard numbers of blood tests with the rich context of diagnostic notes, this new approach offers a way to make hospital care more efficient and personalized. It suggests that the future of medical prediction lies not in choosing between data and text, but in weaving them together to see the full picture of a patient's health.