AI Combines Brain Scans and Medical Records to Predict Second Strokes

When a patient survives a stroke, one of the scariest questions is simple: will it happen again? Doctors currently weigh risk using a mix of scans and medical history, but it's an imperfect guess. A team of researchers in Austria and Germany wanted to see whether AI could do better by looking at both kinds of information at the same time — detailed 3D pictures of the brain's blood vessels (called CTA scans) plus facts about the patient, such as age, gender, and whether they have coronary heart disease, a common heart condition.

The problem they ran into is common in medical AI. When you feed a model two different types of data — pictures and numbers — it tends to pick a favorite and mostly ignore the other. The researchers call this "modality imbalance," and it means the AI misses half the story. Their fix was clever and simple: before teaching the model about strokes, they let it study thousands of scans on its own, a process called self-supervised pre-training. Think of it as letting someone flip through photo albums before asking them to diagnose a patient. The AI learns what healthy and unhealthy vessels look like without needing anyone to label every image by hand.

After that warm-up, they fine-tuned the model on the actual stroke data using two different strategies, one where most of the AI stays frozen and only a small part adjusts. The results were clear: every pre-trained model beat both the older baseline and versions trained from scratch. The best one, built on a Vision Transformer (an AI architecture that breaks images into patches and studies how they relate), finally used both the scans and the patient data in a balanced way instead of collapsing onto just one. The team also found meaningful interactions between the imaging data and two specific factors: gender and coronary heart disease.

Here's the honest catch. This is an academic paper presented at a medical imaging conference, not a product you can ask your doctor about. The abstract doesn't report accuracy numbers, patient counts, or whether the model was tested on people from different hospitals — all of which matter enormously before anything touches real care. The code is public, which is good for science. But turning a promising lab result into a tool that changes treatment would take years of testing on diverse patients. Still, the direction is genuinely encouraging: getting AI to use all the clues, not just the easy ones, is exactly how medical AI gets trustworthy.