Structured-to-Narrative Representation Framework for Next-Day Discharge Prediction Using Structured EHR Data and Clinical Language Models

In the busy rhythm of a hospital, the moment a patient leaves is just as critical as the moment they arrive. For surgeons and nurses, knowing whether a patient who has undergone a planned spine operation will be ready to go home the very next day is a vital piece of information. It allows the medical team to coordinate care, manage bed availability, and ensure that resources are used efficiently. However, predicting this outcome is difficult because every patient is different, and recovery depends on a complex mix of factors: the specific type of surgery performed, the patient's age, their overall health, and how their body responds to the procedure. Traditionally, computers have tried to solve this by looking at rows and columns of data—lists of numbers and codes representing lab results, diagnoses, and procedures. While these methods work well, they often feel like a black box to the doctors using them, offering a prediction without a clear explanation of why that decision was made.

A team of researchers at the University of California, Irvine, has explored a new way to bridge this gap between raw data and human understanding. They asked a simple question: what if a computer could read a patient's medical history not as a spreadsheet of numbers, but as a short, clear story? In their study, published in August 2026, the researchers developed a system that takes the structured data from electronic health records—such as binary indicators for whether a specific lab test was ordered or a specific surgery was performed—and automatically converts them into plain English sentences. They then trained a specialized computer model to read these generated stories and predict whether a patient would be discharged the next day. The goal was not just to get the answer right, but to do so in a way that mirrors how a human doctor thinks, by looking at the narrative of the patient's condition rather than isolated data points.

The researchers analyzed the records of 1,958 adults who had undergone elective spine surgery at a large medical center in the United States. For each patient, they gathered about 600 different pieces of information, including age, sex, laboratory results, and the specific codes for the procedures and diagnoses involved. They carefully excluded personal details like race or zip code to ensure the model focused only on medical factors. The team split this data into two groups to test their ideas. First, they trained standard machine learning models, which are like powerful calculators that look for patterns in numbers, to predict the outcome. Second, they built a different kind of model based on a technology known as a transformer, which is designed to understand language. This second model never saw the raw numbers; instead, it only read the short, generated stories describing each patient's medical situation. These stories were created by a set of rules that turned data points into sentences, such as "The patient is 65 years old and underwent a specific fusion procedure," omitting any information about the final outcome so the model had to learn to predict it on its own.

When the researchers compared the results, the traditional machine learning models proved to be the most accurate at distinguishing between patients who would go home and those who would stay. These models achieved a high level of discrimination, correctly identifying the outcome in about 94 percent of cases when tested across different groups of patients. They were also very good at providing reliable probability estimates, meaning their confidence levels matched reality. The model that relied on the narrative stories, however, performed slightly less well in terms of raw accuracy, achieving a discrimination score of about 84 percent. Yet, this lower number told only part of the story. The narrative-based model showed a different strength: it was much better at catching the patients who would be discharged the next day. While the traditional models sometimes missed these early discharges, the story-reading model successfully identified about 75 percent of them. This is a crucial distinction in a hospital setting, where missing a patient who is ready to leave can delay their care and tie up a bed unnecessarily.

The study also looked closely at how these models made their decisions to ensure they were thinking like doctors. For the traditional models, the researchers used a method called SHAP analysis, which highlights which specific data points were most important for a prediction. They found that the models correctly identified that more complex surgeries, older age, and unstable lab results were the main reasons a patient would stay longer. This matched what medical experts already knew. For the narrative model, the reasoning was built directly into the text it read. When the model predicted a patient would stay, the story it had read likely contained details about a difficult, multi-level surgery or signs of physiological stress. When it predicted a discharge, the story described a simpler procedure with fewer complications. This direct link between the input text and the output decision makes the model's reasoning transparent and easy for a clinician to follow, unlike the hidden calculations of the traditional approach.

One of the most significant findings of the study was that generic, pre-trained language models—those that have read vast amounts of general text but not specific medical data—failed completely at this task. When the researchers tried using these off-the-shelf models without any special training, they could not learn to distinguish between the two groups of patients, often guessing randomly. This confirmed that while language models are powerful, they need to be specifically taught how to read medical data to be useful in a hospital. The researchers showed that by fine-tuning a model specifically on the structured-to-narrative stories derived from their patient data, they could create a system that understood the nuances of clinical recovery. The model learned to recognize that certain combinations of procedures and lab results, when described in a sentence, carried a specific meaning about a patient's readiness to go home.

The researchers concluded that while the traditional machine learning models remain the most accurate tool for pure prediction, the narrative-based approach offers a powerful alternative for real-world use. In a busy hospital, a prediction is only as good as the trust the medical team has in it. A system that explains its reasoning in the same language doctors use every day—describing a patient's condition as a series of events rather than a list of codes—may be more likely to be adopted and trusted. The study suggests that transforming structured data into stories does not just make the output easier to read; it fundamentally changes how the computer processes the information, aligning its logic more closely with human clinical reasoning. This approach provides a path forward for integrating artificial intelligence into daily hospital workflows, where the goal is not just to predict the future, but to do so in a way that helps doctors make better, more informed decisions for their patients.