How AI is Revolutionizing Medicine: Predicting Risks Before They Happen
Featured paper: A scoping review of artificial intelligence applications in clinical trial risk assessment
Disclaimer: This content was generated by NotebookLM and has been reviewed for accuracy by Dr. Tram.
Have you ever wondered how a new medicine gets from a scientist’s lab to your local pharmacy? The journey is long, incredibly expensive, and unfortunately, usually ends in failure. To make sure a drug is safe for humans, it has to go through clinical trials—rigorous tests that evaluate whether a treatment works and if it causes dangerous side effects. However, the cost of developing a successful drug has been skyrocketing for decades, and many trials fail simply because they are too difficult to manage or because unexpected problems pop up at the last minute.
This is where Artificial Intelligence (AI) enters the picture. A group of researchers led by Douglas Teodoro recently published a major paper titled “A scoping review of artificial intelligence applications in clinical trial risk assessment”. They looked at over a decade of research—142 specific studies published between 2013 and 2024—to see how AI is being used as a high-tech “crystal ball” to predict and prevent risks in medical research.
The Three Big Hurdles: Safety, Efficacy, and Operations
When scientists run a clinical trial, they are essentially trying to clear three giant hurdles. The researchers found that AI is now being used to help jump over all of them.
1. Safety: Will it hurt the patient? The most important rule in medicine is “do no harm.” AI is now being trained to predict Adverse Drug Events (ADEs)—which are basically injuries caused by a drug. Some AI models can even predict which specific organ, like the liver or kidneys, might be harmed by a new compound before it is even given to a human. By analyzing the molecular structure of a drug, AI can warn scientists if a treatment is likely to be toxic.
2. Efficacy: Does it actually work? A drug might be safe, but it’s useless if it doesn’t fix the problem. AI helps here by predicting a “drug response”. For example, instead of giving a drug to everyone and hoping for the best, AI can look at a patient’s unique data—like their genetics—to estimate how well they will respond compared to a “control group” who didn’t get the treatment. This is a big step toward personalized medicine, where treatments are tailored to the individual.
3. Operational Effectiveness: Can we actually finish the test? This is the “behind-the-scenes” risk. Sometimes trials fail because they can’t find enough volunteers or because they run out of money. AI is now being used to predict the “likelihood of approval”—essentially betting on whether the FDA (the government agency that approves drugs) will give the green light based on how the trial is designed. It can even predict if a trial will “finish gracefully” or be forced to stop early.
How the “Brain” of AI Works
You might be wondering: how does a computer program know if a drug will work? The sources describe a three-step process that these AI “brains” follow:
- Representation: First, the AI takes complex info—like the chemical shape of a drug or the text of a medical protocol—and turns it into numbers (vectors) that the computer can understand.
- Learning: The AI then looks at thousands of past examples to find patterns. It uses algorithms like “Random Forest” (which acts like a giant decision tree) or “Deep Learning” (which mimics the human brain) to learn what a “successful” drug looks like versus a “failed” one.
- Prediction: Finally, the AI makes its guess. It might say, “There is a 90% chance this drug will cause a mild headache” or “There is a 20% chance this trial will be canceled because it’s taking too long”.
Recently, there has been a massive surge in Large Language Models (LLMs)—the same kind of technology behind tools like ChatGPT. In 2023 alone, about 20% of the studies reviewed used these models to read through trial documents and find hidden risks that human researchers might miss.
The “Digital Twin” and Virtual Trials
One of the coolest things the researchers discovered is the rise of “virtual trials”. Instead of testing a drug on a real person first, scientists can use AI to create a “Digital Twin”—a computer-simulated version of a patient.
These simulations allow researchers to run a “trial” thousands of times in a computer lab to see how different dosages might affect someone. While these aren’t meant to replace real human trials yet, they provide a much safer and faster way to narrow down which medicines are worth testing in real life.
The Catch: Why AI Isn’t Perfect (Yet)
While AI is incredibly powerful—some models boast an accuracy of 96% in predicting side effects—there are still big challenges.
The biggest problem is “Selection Bias”. If an AI is only trained on data from one group of people (for example, only men or only people from one country), it might make mistakes when trying to predict what will happen to everyone else. There is also the “Garbage In, Garbage Out” rule: if the data scientists give the AI is messy or incomplete, the AI’s predictions will be wrong.
Another issue is that most AI models today are “siloed”. This means a model might be a genius at predicting liver safety but have no idea if the drug actually cures the disease. Scientists are now working on “multi-task” AI that can look at safety, efficacy, and cost all at the same time to give a “holistic” view of the risk.
The Road Ahead
Even with these challenges, the researchers are very optimistic. Between 2013 and 2020, most AI research focused only on safety. But since 2021, there has been an “exponential growth” in AI being used for all parts of a trial.
In the future, we can expect AI to move from just looking at past data (retrospective) to watching trials happen in real-time (prospective). This would allow doctors to adjust a trial instantly if the AI detects a safety risk, potentially saving lives and millions of dollars.
The goal isn’t to let computers take over medicine, but to give human scientists a powerful new set of tools. By using AI to navigate the “intricate landscape” of medical risks, we can hopefully bring safer, more effective medicines to the people who need them much faster than ever before.