Early Mortality Prediction in Upper Gastrointestinal Bleeding Using Explainable Machine Learning: Comparison with Established Clinical Risk Scores

Imagine the Emergency Department (ED) as a bustling train station where thousands of passengers (patients) arrive every day, some just catching a quick connection, others in desperate need of immediate rescue. Among these, a specific group arrives with "Upper Gastrointestinal Bleeding" (UGIB)—a fancy term for a serious bleed in the upper part of the stomach or esophagus. The doctors' biggest challenge is acting like a super-smart conductor: they need to instantly spot which passengers are in danger of not making it to their destination within 30 days, so they can rush them to the VIP rescue team (Intensive Care) instead of sending them to the waiting room.

For years, the conductors have used a set of old, printed rulebooks called "clinical risk scores" (like the Glasgow-Blatchford Score, Rockall, AIMS65, and the newer ABC score). These rulebooks are like simple checklists: "If the patient is over 65, add a point. If their blood pressure is low, add another." They work, but they are a bit rigid, like a robot that only understands straight lines and can't see the messy, tangled web of how a human body actually reacts to stress.

Enter the new challenger: a team of Machine Learning (ML) detectives, specifically a model called XGBoost. Think of this model not as a checklist, but as a super-observant detective who can spot hidden patterns in a crowd that a human eye might miss. This detective doesn't just look at one thing; it looks at 63 different clues available the moment a patient walks through the ED door—like their age, how fast they are breathing, their blood sugar, and even how many white blood cells they have.

The Big Showdown
The researchers gathered data from 719 patients who came to the ED between January 1, 2015, and January 1, 2026. Out of these, 53 patients (about 7.4%) sadly passed away within 30 days. The goal was to see if the new ML detective could predict these deaths better than the old rulebooks.

The results were a bit like a surprise race. The old rulebooks tried their best, with the newest one, the ABC score, being the fastest of the bunch, achieving a "score" (AUROC) of 0.742. But the ED XGBoost model, using only the information available right when the patient arrived, ran a slightly better race with a score of 0.789.

Here is the twist: The paper suggests that the difference between the ML model and the best old rulebook (ABC) wasn't a massive, undeniable victory. The math says the gap was small enough that it might just be luck in the numbers (a p-value of 0.264). However, the ML model did something the old books couldn't do as well: it was much better at "calibration." Imagine a weather forecaster; the old books might say "there's a 50% chance of rain" when it's actually going to be a downpour. The ML model's predictions were much closer to the real weather, with a "Brier score" (a measure of prediction error) of 0.061 compared to 0.142 for the ABC score. In plain English, the ML model didn't just guess who would get sick; it guessed how sick they would get with much more accuracy.

The "Magic" 15 Clues
The researchers wondered: "Do we need all 63 clues, or can we just use the top 15?" They tried to shrink the detective's toolkit. In their initial experiments, a tiny model with just 15 variables (like lymphocyte count, lactate, and albumin) seemed to run even faster, hitting a score of 0.814.

But the paper is very careful here. When they tested this tiny model with a stricter, more cautious method (to make sure they weren't just getting lucky with the data), the score dropped a bit to 0.765. The authors suggest that the initial "super-high" score was partly due to the model getting a little too comfortable with the specific data it was trained on. Even with this drop, the tiny model still performed well and was better than the ABC score.

What the Detective Found
Using a special tool called SHAP (which acts like a magnifying glass to see which clues matter most), the model revealed the most important signs of danger. The top suspects were:

• Lactate: High levels (like 2.50 mmol/L in those who died vs. 1.70 mmol/L in survivors) were a huge red flag. This is like a car engine overheating; it means the body isn't getting enough oxygen.
• Lymphocyte count: Low levels (like 0.95 vs 1.55) suggested the immune system was struggling.
• C-reactive protein: High levels (like 12.0 mg/dL vs 3.46 mg/dL) pointed to serious inflammation.
• Albumin: Low levels (like 3.30 g/dL vs 3.59 g/dL) hinted at poor nutrition or long-term stress.

What the Paper Rules Out
The study explicitly argues against the idea that you need to wait for an endoscopy (a camera test inside the stomach) or transfusion records to predict who will die. The "Full Model" (which included these later details) performed almost exactly the same as the "ED Model" (which only used initial data). The authors suggest that the patient's overall physical state and how their body is reacting to the bleed are more important than the specific details of the bleed itself.

The Final Verdict
The paper concludes that these new Machine Learning models are a promising, practical tool that could help doctors make better decisions right when a patient walks in. They are more accurate and better calibrated than the old rulebooks. However, the authors are honest: this was a single study at one hospital with a relatively small number of deaths (53). They suggest that before we hand these models to doctors to use in real life, they need to be tested in other hospitals to make sure they work everywhere. It's a very strong hint that the future of emergency care might look like a super-smart detective, but we aren't quite there yet.