Prediction at Hospital Admission of In-Hospital Mortality and High-Cost Hospitalization Using Machine Learning in Patients With Severe Community-Acquired Pneumonia: An Exploratory Dual-Risk Stratification Study
This study developed and temporally validated separate machine learning models using admission data to predict both in-hospital mortality and high-cost hospitalization in patients with severe community-acquired pneumonia, demonstrating superior performance over traditional scores and revealing distinct clinical and resource-use patterns across combined risk strata.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Every year, millions of adults are hospitalized with severe pneumonia, a life-threatening infection that strikes the lungs. For doctors, the immediate challenge is twofold: they must decide which patients are in the most danger of dying and which patients will require the most expensive, resource-intensive care. Traditionally, hospitals have relied on simple checklists to gauge how sick a patient is. These tools look at basic signs like age, confusion, breathing rate, and blood pressure to estimate the risk of death. However, these same checklists are not designed to predict financial costs, and they often miss the complex biological signals that indicate a patient will need a long, expensive stay in the hospital. As healthcare systems face growing pressure to manage resources wisely, there is a critical need to understand not just who might die, but who will consume the most medical supplies and staff time, all based on the information available the moment a patient walks through the door.
A team of researchers at Nanjing First Hospital and China Pharmaceutical University set out to solve this problem by building a new kind of prediction system. They gathered data from 1,543 adults admitted with severe community-acquired pneumonia between 2020 and 2025. Instead of relying on a single checklist, they used machine learning, a type of computer program that learns patterns from vast amounts of data, to create two separate prediction models. The first model was trained to predict the likelihood of a patient dying during their hospital stay. The second model was trained to predict whether a patient's total hospital bill would be exceptionally high, defined in this study as costing more than 38,969 Chinese yuan. Crucially, both models were built using only the information doctors have at the very moment of admission, such as vital signs, blood test results, and the patient's medical history, ignoring any data that only becomes available after treatment has begun.
The researchers found that their machine learning models were significantly better at making these predictions than the standard tools currently in use. For predicting death, their best model achieved an AUC of 0.73, a notable improvement over the standard scoring system, which was correct only about 54 percent of the time. For predicting high costs, their model also outperformed a simple baseline that looked only at age and the number of existing health conditions. The study revealed that the factors driving a high risk of death were not exactly the same as those driving a high cost. While both models looked at similar data points like oxygen levels and heart function, the computer learned that different combinations of these signs pointed toward different outcomes.
To make sense of these findings, the researchers grouped patients into four categories based on the combination of risks their models predicted. They discovered two particularly distinct groups that traditional methods often miss. One group consisted of patients who were predicted to have a very high risk of dying but a low risk of generating high costs. These patients were typically older and showed severe signs of physical distress, such as very high breathing rates and high levels of a protein called lactate, which indicates the body is struggling to get enough oxygen. Because their condition was so critical, they often required intensive care and mechanical ventilation, but their hospital stays were frequently cut short by death, resulting in lower total costs.
In contrast, a second group was predicted to have a low risk of dying but a very high risk of generating high costs. These patients were generally more stable physically but required a wide array of expensive resources. Their high costs came from a cumulative use of medical supplies, nursing care, laboratory tests, and medications over a longer period. The study showed that these two groups had very different clinical profiles, suggesting that a single risk score is not enough to guide hospital management. By looking at both risks simultaneously, doctors could potentially identify patients who need immediate life-saving intervention versus those who need careful, resource-heavy management to prevent complications.
The researchers emphasized that while their models performed well in testing, they are still in an exploratory phase. The study was conducted at a single hospital, and the specific costs predicted are tied to local pricing and healthcare policies. The team noted that their work does not prove that changing treatment based on these predictions will improve outcomes, but it does provide a new way to visualize patient needs. They suggest that in the future, such dual-risk assessments could help hospitals allocate their limited resources more effectively, ensuring that the sickest patients get urgent care while those likely to have long, complex stays receive the attention they need without overwhelming the system. The study concludes that using the same admission data to predict both death and cost offers a more complete picture of the challenges facing severe pneumonia patients than current methods allow.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.