Reconsidering the case against risk prediction in self-harm: routinely collected health data distinguishes groups at higher and lower risk of adverse outcomes following paracetamol overdose

The Detective's Dilemma: Can Data See the Storm Before It Breaks?

Imagine you are a weather forecaster trying to predict a sudden, dangerous storm. In the world of medicine, specifically when people come to the hospital after taking too much paracetamol (a common painkiller) to hurt themselves, there is a big debate about whether we can predict the "storm" of a future suicide or a return to the hospital. For a long time, the official rule in the UK has been: "Don't try to predict it." The thinking was that the data doctors collect every day—like age, past medical visits, and prescription lists—is just too noisy and useless to tell us who is at risk. It's like trying to guess the weather by looking at a single, blurry photo of a cloud; the experts said the photo just didn't have enough detail to be useful.

But what if that photo actually does have a secret pattern? What if, by using a super-smart computer brain (machine learning) to look at thousands of these blurry photos at once, we could spot a shape that the human eye misses? This is the question scientists have been asking. They want to know if the "routinely collected health data"—the boring, everyday records doctors write down—actually holds a hidden signal that can separate people who will have a bad outcome from those who won't. If the answer is "yes," it doesn't mean we have a crystal ball, but it might mean we shouldn't have thrown the map away just yet. If the answer is "no," then we know for sure that we need to look elsewhere for help.

The Great Paracetamol Puzzle: Finding the Signal in the Noise

In this study, a team of researchers decided to put the "data is useless" theory to the test. They didn't try to build a magic tool to stop suicide; instead, they acted like detectives looking for a single clue. They gathered the medical records of 4,095 adults who had come to emergency departments in NHS Lothian (Scotland) between 2017 and 2023 with a paracetamol overdose.

Think of these records as a giant puzzle box. Inside were 37 different pieces of information that a doctor would already know at the moment the patient walked in: their age, where they lived (which gives a clue about how deprived the area is), what medications they were already taking, and whether they had visited a mental health specialist before. The researchers fed these puzzle pieces into a smart computer model called "elastic-net logistic regression." You can think of this model as a very strict librarian who is trying to sort a massive pile of books into two stacks: "Books about people who will have a bad outcome" and "Books about people who won't."

The researchers looked at three different time periods after the patient left the hospital:

1. The first week (0–7 days): Did they die or get admitted to a mental health ward?
2. The next month (8–30 days): Did the trouble continue?
3. The whole year (31–365 days): Did the storm eventually break?

The Big Discovery
The results were surprising. The computer model did find a pattern. It wasn't perfect, but it was definitely better than flipping a coin.

• In the first week, the model could rank patients correctly about 65% to 82% of the time (depending on how you measure it).
• In the 8-to-30-day window, it was between 63% and 90%.
• Over the full year, it was between 71% and 85%.

In the world of statistics, a score of 50% is just guessing. Anything above that means the model is actually seeing something real. The researchers found that the model was good at spotting who was at higher risk, and it was very good at making sure the numbers it gave matched the real-life outcomes (this is called "calibration").

Where Did the Clues Come From?
You might think the model was just looking at age or gender, like a stereotypical detective. But the researchers found that wasn't the main story. The real "superpowers" of the model came from mental health-related clues. The model got its best guesses from things like:

• Whether the person had been prescribed antipsychotic or anxiety medications.
• If they had a history of seeing a psychiatrist.
• If they had a diagnosis of schizophrenia or alcohol-related liver disease.
• How many different types of medications they were taking (polypharmacy).

Age was still used by the model, but it wasn't the heavy lifter. The real signal was hidden in the complex history of a person's mental health journey.

What This Paper Says (and What It Doesn't)

It is very important to understand what this study is not saying. The authors are not claiming that we can now predict exactly which individual person will attempt suicide next. They are not saying we should start using this computer model in hospitals tomorrow to decide who gets treatment.

In fact, they explicitly warn against that. The model's accuracy, while better than chance, is not high enough to be used as a final judge for a single person. If you used this tool on one person, it might still get it wrong. The paper argues that the current UK rules, which say "no risk prediction tools at all," might be throwing the baby out with the bathwater. The rule was based on the idea that no useful signal exists in the data. This study suggests that a useful signal does exist, but it's just not strong enough yet to be a standalone crystal ball.

The researchers also point out that the "boring" structured data they used (the 37 checkboxes) has a ceiling. It's like trying to understand a movie by only reading the cast list and the running time; you miss the plot, the emotions, and the dialogue. The real, deep clues about a person's risk—like their specific thoughts, their feelings of hopelessness, or their support system—are usually written in the free text of doctor's notes, not in the checkboxes. The paper suggests that if we teach computers to read those notes (using something called Natural Language Processing), we might find an even stronger signal.

The Takeaway

So, what's the verdict? The paper concludes that the idea that "routinely collected health data contains no useful predictive signal" is not true. The data does have a signal; it can distinguish between groups of people at higher and lower risk. However, the signal is currently too faint to be used for making life-or-death decisions about single individuals.

The authors are essentially saying: "Don't close the book on this research just because the current chapter isn't perfect." They are urging scientists to keep looking, perhaps by teaching computers to read the messy, human parts of medical notes, rather than giving up on the idea that data can help us save lives. The door to better prediction is still open; we just need better tools to walk through it.