This study is actually a sequel. Kind of.

Back in 2020, Dr. Blomberg (co-author of this paper) submitted his dissertation on the use of machine learning for recognition of cardiac arrest in emergency calls. The thesis was based on a collaboration with Corti, a Danish tech company specialising in machine learning and IT solutions for the health care sector. The results were promising, and this led to the idea of taking the next step – Investigate another acute illness, one that is less binary in its presentation, but where time is also critical: Stroke.

Stroke remains a leading cause of death, disability, and cognitive impairment, with the American Heart Association projecting that 4% of US adults will suffer from a stroke by 2030, incurring an annual cost of $240 billion​​. 

The critical aspect of stroke management lies in its time-sensitive nature, with the best results achieved when treatment starts within the first 4.5 hours after the stroke begins​​. However, the current prehospital triage process by call-takers is not sufficiently accurate, with as much as half of stroke patients not receiving the correct initial assessment​​. The authors of this paper who worked with patients clinically in daily practice have long been aware of this challenge and agreed wholly with the consensus that addressing this gap in early and precise stroke symptom recognition during emergency calls is crucial for enhancing patient outcomes and healthcare efficiency.

The challenge in accurately recognizing strokes through traditional methods is rooted in the complexity and variability of stroke symptoms. Despite efforts to improve healthcare professional training and introduce more specific assessment tools, the subjective nature of human analysis limits stroke identification accuracy​​. The integration of artificial intelligence (AI) into this process could introduce a significant paradigm shift, requiring technological advancement and a reconfiguration of operational frameworks and data management in emergency medical services.

In our study, we trained a machine learning framework which significantly outperformed human call-takers in identifying stroke cases in prehospital medical helpline calls​​. By transcribing speech and analyzing language patterns during nurse and emergency calls, the machine learning model can detect subtle signs of stroke, leading to higher sensitivity and predictive accuracy.  This could not only enhance the call-taker's ability to triage effectively, but also ensure faster and more accurate medical response, potentially saving critical treatment time and improving patient outcomes.

The study presented a set of challenges for the research group. For all researchers, getting reliable and valid data requires hard work, and can sometimes be a struggle. In Scandinavia, well-maintained mandatory registries are in place, which enabled us to obtain reliable data on patients with stroke. A major challenge in this study came in matching the patients with stroke to the medical helpline calls. Without the region's commitment to healthcare excellence, digital infrastructure, and openness to technological advancements this may have proved insurmountable. Thankfully, broad cooperation and commitment enabled us to complete the project.

Another challenge, which doubled as a strength, was the interdisciplinary nature of the author group. This study brought together neurologists, data scientists, machine learning experts, and emergency medical professionals, and it could be surprising how styles and conventions differ between disciplines. This applied to everything from different words describing the same concepts, to the style of presentation, even which section of the paper different information should be presented in. A study such as ours could not be completed without interdisciplinary cooperation and understanding. Meeting these challenges was not only necessary but allowed us to get an outside look at our own respective fields, better enabling us to explain why we do what we do. In the end, it was our experience that it also works wonders for the ability of the researchers to explain what your study is about, and how the study was carried out to people outside the author group. Most researchers who have attended at least one scientific conference will recognize the necessity of developing this skill.

The process of writing the paper presented its own challenges. We aimed to make technical results accessible to clinicians and decision-makers, emphasizing the framework as a supportive tool for call operators, not a replacement. Finding the best way of doing this required many discussions, compromises, and explanations back and forth, and we are proud of the results.

It would be easy to neglect the challenges in implementation of novel solutions, such as the one presented in our paper. We tried to avoid this by making the results as transparent as possible and including in the supplementary as much information as possible on the performance across a variety of metrics. We also highlighted that our proposed machine learning framework is a supporting tool; without the human-to-human interaction in the call, the framework cannot function. It is a tool for aiding telephone operators in the difficult task of stroke recognition, not a tool that can replace them.

While we recognize that this study is only the first step, we believe that the development of this machine learning framework for stroke recognition is a prime example of the power of interdisciplinary collaboration where the team has utilized their varied expertise to develop a system that improves the quality of prehospital care. This collaborative endeavor represents not just the fusion of medical knowledge with advanced technology but also a collective dedication to advancing healthcare and saving lives.