MultiDiseasePred: An Interpretable Multi-Disease Risk Prediction Framework and Its Application for Early Risk Stratification in the Emergency Department

In the chaotic, high-stakes environment of a hospital emergency room, the first few minutes are often the most critical. When a patient arrives, they are not just a single problem; they are a complex mix of symptoms that could signal anything from a heart attack to a severe infection or a blocked lung artery. Traditionally, doctors have relied on separate checklists and scoring systems to assess each of these possibilities one by one. A nurse might use one tool to check for sepsis, another for heart trouble, and a third for breathing issues. While these tools are helpful, they are designed to look at only one disease at a time. This creates a fragmented picture, forcing clinicians to mentally stitch together different risk scores while racing against the clock, often without the benefit of lab results or imaging scans that arrive later. The challenge is to see the whole patient immediately, understanding how different risks overlap and compete for attention before the full medical story is known.

To address this, a team of researchers from the University of Minnesota and other institutions developed a new approach called MultiDiseasePred. Instead of building separate models for each disease, they created a single, unified system designed to look at a patient and simultaneously estimate the risk of seven different acute conditions and two critical outcomes, such as whether the patient will need to be admitted to the hospital. The system focuses on the seven conditions that often present with confusingly similar early symptoms: sepsis, acute respiratory distress syndrome, exacerbations of chronic lung diseases like asthma or COPD, acute coronary syndrome (heart issues), viral pneumonia, pulmonary embolism (a blood clot in the lung), and acute kidney injury. It also predicts the likelihood of hospitalization and a composite "critical outcome" involving intensive care admission or death. The goal was not just to predict these risks, but to do so using only the information available the moment a patient walks through the door, such as vital signs, age, and chief complaints, without waiting for blood tests or X-rays.

The researchers trained this system on a massive dataset comprising nearly 1.7 million emergency room visits from a large health system in Minnesota, covering the years 2019 through 2025. They then tested how well the system worked on completely different groups of patients from two other major medical centers: one in Boston and one in California. This external testing was crucial to ensure the system wasn't just memorizing the local data but could actually understand the universal signs of illness. The system learned to identify patterns across these millions of visits, discovering that while each disease is unique, they often share underlying physiological signals. For instance, a drop in blood pressure or a change in breathing rate might be a warning sign for several different conditions at once. By analyzing these shared signals alongside specific clues, the model learned to weigh the evidence for each condition separately while still seeing the connections between them.

A key feature of this new framework is its ability to remain simple and practical for real-world use. The researchers found that they did not need a long list of dozens of variables to make accurate predictions. Through a careful selection process, they narrowed the inputs down to just twelve variables that are routinely recorded by triage nurses, such as heart rate, blood pressure, oxygen levels, age, and specific past medical history. This "parsimonious" set of inputs means the system can be used immediately without requiring complex data entry or waiting for lab results. The model also includes a built-in explanation system. When it flags a patient as being at high risk for a specific condition, it does not just give a number; it highlights exactly which factors contributed to that decision. For example, if the system predicts a high risk of a pulmonary embolism, it might show that the patient's history of pulmonary hypertension was the primary driver of that prediction, allowing the clinician to see the logic behind the alert.

The results of the study were promising. In the internal testing, the system achieved a high level of accuracy in distinguishing between patients who would develop these conditions and those who would not, performing as well as or better than existing single-disease machine learning models. More importantly, when tested on the external data from Boston and California, the system maintained strong performance without needing to be retrained or adjusted for the new locations. This suggests that the patterns the model learned are robust and can travel across different hospital settings. For the seven acute conditions, the system correctly identified risks with a high degree of reliability, and for the two disposition endpoints, it also performed well. The researchers noted that while the system was excellent at ranking patients by risk, its ability to predict very rare events with perfect precision was slightly lower, which is a common challenge in medical prediction. However, the system successfully provided a coordinated, multi-condition risk profile that clinicians could interpret immediately.

The study also examined how the system handles the reality of emergency care, where data is often incomplete or missing. The researchers designed the system to handle missing information by using standard, pre-defined values based on the training data, ensuring that the model behaves consistently regardless of whether a specific vital sign was recorded. They also ensured that the system did not rely on the Emergency Severity Index, a score often assigned by nurses, to avoid circular reasoning where the model simply repeats the nurse's initial assessment. Instead, it builds its own independent assessment from the raw clinical data. The researchers emphasized that this tool is intended to be a decision aid, not a replacement for clinical judgment. It is designed to help clinicians organize their thoughts and prioritize which conditions to investigate first, potentially speeding up the time to critical treatments like antibiotics for sepsis or blood thinners for clots.

Despite the strong performance, the authors are careful to note that this is a retrospective study, meaning it looked at past data rather than testing the system in a live, real-time environment. They acknowledge that before such a tool can be widely adopted, it must be tested prospectively to see how it actually affects patient outcomes, workflow efficiency, and whether it helps or hinders the decision-making process in a busy emergency department. There are also limitations regarding the specific data sources used; for example, the system relies on structured data and does not yet incorporate unstructured notes or imaging results. Furthermore, the study did not include a detailed analysis of how the system performs across different racial or ethnic groups, as that data was not consistently available across all the hospitals studied. These are important areas for future research to ensure the tool is safe and effective for everyone.

Ultimately, this work represents a shift in how emergency risk assessment might be approached. By moving away from a collection of isolated, single-disease tools toward a unified, interpretable framework, the researchers have demonstrated that it is possible to see the full spectrum of a patient's potential risks at the very start of their care. The system provides a clear, calibrated probability for multiple conditions at once, accompanied by an explanation of why those risks were flagged. This approach respects the complexity of the emergency department, where patients often have overlapping symptoms and multiple potential diagnoses. While the system is not yet a finished product ready for immediate deployment in every hospital, it offers a compelling proof of concept that a single, intelligent framework can help clinicians navigate the uncertainty of the first few minutes of an emergency encounter, potentially saving valuable time and improving the care of patients with life-threatening conditions.