How stroke outcomes can be predicted

Healthcare providers must quickly determine who needs which treatment

Josline Otieno, a PhD student at Umeå University, has studied how different methods can be used to predict stroke outcomes. She compared machine learning with logistic regression, a common statistical method used to estimate risk. The study is based on large datasets from national stroke registers in Sweden and the United Kingdom.

When predicting the risk of death within 30 days after stroke, the difference between machine learning and logistic regression was small. Both methods produced stable results, even when comparing data from different countries.

However, when it came to patients’ functional ability three months after stroke, clearer differences emerged. Here, advanced machine learning models were better at identifying patients who would become dependent on assistance group that is important for planning rehabilitation, according to Otieno.

– Stroke is a complex condition, and healthcare providers must quickly determine who needs different treatments, who is likely to recover well, and who may require long-term support, says Josline Otieno.

The choice of method determines what works best

The study also includes analyses of survival over time and situations where multiple possible outcomes compete with each other, such as recurrent stroke or death.

The results show that the choice of method should be adapted to the structure of the data and the clinically relevant time horizon. Cox regression, a common statistical method for analysing survival over time, works well when its assumptions are met. However, when relationships are more complex or when the data contain a high degree of uncertainty, machine learning often performs better.

– In situations with competing risks, performance changes over time. No model was consistently best at all evaluation time points, says Josline Otieno.

According to Otieno, machine learning often performed better in the short term, when many events occur. Over longer follow-up periods, traditional statistical models were more reliable.

Supporting decision-making in healthcare

The study combines a simulation study—where researchers test methods on computer-generated data—with analyses of real-world data. The simulation study examines how factors such as sample size, censoring, model assumptions, and uncertainty affect the results. The conclusion is that the choice of method should depend on the context and that models should be evaluated using multiple measures.

– More reliable assessments can improve communication between healthcare professionals and patients and provide better support for treatment decisions, especially when models are used at clinically meaningful time points, concludes Josline Otieno.