What Can We Expect From Machine Learning Predictions in Daily Clinical Neurology?

We have entered the age of machine learning in neurologic research, with thousands of research articles in the last 2 years (openalex.org). So how soon can we expect new machine learning algorithms to reach clinical practice, and what can we expect in the future? It is certainly likely that machine-learning tools and algorithms will benefit our abilities to predict future outcomes.

Knowing the relative risk of serious neurologic events enable more nuanced personalized patient care and engender patient-physician dialogues about desired goals in relation to known risks. Although the future is always somewhat uncertain, machine learning can refine relative risk estimates – much the way a weather report tells us the risk of rain. Neurology is no stranger to the implementation of risk-based predictive tools. The release of the original CHADS-VASc calculator dates back to 2010.¹ This tool, used to stratify patients by risk of stroke and thromboembolism, is easy to use and transparent in its calculations. Similar tools have also emerged recently, such as the SeLECT score² for determining risk of seizure after ischemic stroke. Such tools will become even more accurate and consistent in their risk estimates as machine learning is incorporated into their development.

We recently developed the DAAE-M score³ as one example of how machine learning can be translated into clinically usable tools. The model estimates 5-year risk of disease progression in people with multiple sclerosis (MS) by combining information on disease status and treatment exposure. It predicts outcomes including transition to secondary progressive disease or progression to high disability with progression independent of relapse^4–6—events that are clinically meaningful because they often mark a shift toward more rapid and less treatment-responsive deterioration.^7,8

The DAAE-M score is freely available here: https://tomafuchs.com/daae-score/ and will soon be integrated into Epic.

However, the development of this tool also highlighted a broader lesson: accuracy alone is not sufficient for clinical adoption. From the outset, we prioritized not only accurate predictive performance but also interpretability, transparency, accessibility, usability, and speed. The final tool can be used in under 30 seconds and produces outputs that can be readily understood in a clinical conversation. This design philosophy reflects a more general requirement for machine learning in medicine—tools must fit clinical workflows if they are to be used at all.

In practice, the value of such tools may lie less in the prediction itself and more in how that prediction is used. Risk estimates can serve as a starting point for discussion, helping physicians and patients navigate decisions during key moments. Importantly, these tools are not prescriptive. Like a weather forecast, they provide probabilistic information that may influence decisions differently depending on the individual patient and clinical context. We view tools like the DAAE-M score as a source of free information. The reports themselves are another tool in the toolbelt, to be wielded when helpful.

One of the most striking observations from this work was the disconnect between methodological innovation and real-world usability. Machine learning is widely applied in research settings, but far fewer efforts are directed toward external validation, accessibility, or the end-user (the doctors) experience. As a result, many promising models remain impractical for daily use. Addressing this gap requires a shift in how such tools are developed—moving from a technology-centered approach toward one that is grounded in clinical realities. In our case, this meant involving practicing physicians throughout the development process. Their input shaped not only which outcomes were prioritized, but also how the tool presents information and how quickly it can be applied in practice. This type of collaboration led us to approach the project not just as researchers, but also as designers and software developers, an approach that may become increasingly important as machine learning moves closer to the bedside – and one requiring that physicians have a seat at the table in the design process.

Looking ahead, the integration of predictive tools into clinical care is likely to be iterative. As physicians begin to use more of these tools, expectations will evolve, and new needs will emerge. This feedback loop—between tool use and tool development—has the potential to drive more meaningful innovation than purely methodological advances. In this sense, the future of machine learning in neurology will not be defined solely by algorithmic improvements, but by how well these tools adapt to the needs of clinical practice. In this process, future research will need to proceed along two complementary lines. First, continued validation across diverse populations is essential to ensure that predictions remain reliable and generalizable. Second, there is a need to expand the range of clinically relevant outcomes that can be predicted. In MS, for example, this may include milestones such as high disability levels or significant cognitive dysfunction—outcomes that are relevant to patients but remain difficult to anticipate.

However, successful implementation of machine learning in neurology will depend heavily on the engagement of clinicians themselves. Physicians are uniquely positioned to identify the barriers to adoption: What makes a tool trustworthy? What would make it usable in a busy clinic? What information is actually needed to support decision-making or dialogues with patients? These questions should not be addressed after a model is built, but at the earliest stages of development. Ensuring that clinicians have a voice in this process will be critical to bridging the gap between innovation and impact. We therefore implore physicians to take part whenever possible. Share your ideas. Write directly to researchers. There are impressive efforts in the MS community, such as MSBase,⁹ the Czech National MS Registry,¹⁰ or the New York State MS consortium,¹¹ for collaborating with clinical data across the world to improve the impact and generalizability of clinical research. If possible, take part or send your thoughts to these organizations about what you believe should be prioritized. There is an exciting future for machine learning in neurology and if we work together as a community, we can create the best tools to fit the needs of our evolving clinical practice.

REFERENCES
1. Lip GYH, Nieuwlaat R, Pisters R, Lane DA, Crijns HJGM. Refining Clinical Risk Stratification for Predicting Stroke and Thromboembolism in Atrial Fibrillation Using a Novel Risk Factor-Based Approach. Chest. 2010;137(2):263-272. doi:10.1378/chest.09-1584
2. Galovic M, Döhler N, Erdélyi-Canavese B, et al. Prediction of late seizures after ischaemic stroke with a novel prognostic model (the SeLECT score): a multivariable prediction model development and validation study. The Lancet Neurology. 2018;17(2):143-152.
3. Fuchs TAN, Schoonheim MM, Strijbis EMM, et al. Predicting disease progression in multiple sclerosis with clinically accessible information and technology. J Neurol. 2026;273(5):281. Published 2026 Apr 19. doi:10.1007/s00415-026-13802-4
4. Fuchs TA, Zivadinov R, Pryshchepova T, et al. Clinical risk stratification: Development and validation of the DAAE score, a tool for estimating patient risk of transition to secondary progressive multiple sclerosis. Multiple Sclerosis and Related Disorders. 2024;89:105755. doi:10.1016/j.msard.2024.105755
5. Jelgerhuis JR, Uher T, Havrdova EK, Horakova D, Schoonheim MM, Fuchs TA. Clinical risk stratification: Prague validation of the DAAE score, a clinical tool for estimating risk of disesase progression in multiple sclerosis. Multiple Sclerosis and Related Disorders. 2025;98:106423. doi:10.1016/j.msard.2025.106423
6. Fuchs TA, Ziccardi S, Jelgerhuis JR, et al. Verona validation of the DAAE Score, a clinical tool for estimating patient risk of transition to secondary progressive multiple sclerosis. Multiple Sclerosis and Related Disorders. Published online June 13, 2025:106586. doi:10.1016/j.msard.2025.106586
7. Fuchs TA, Ziccardi S, Dwyer MG, et al. Response heterogeneity to home-based restorative cognitive rehabilitation in multiple sclerosis: An exploratory study. Multiple Sclerosis and Related Disorders. 2019;34:103-111. doi:https://doi.org/10.1016/j.msard.2019.06.026
8. Kappos L, Wolinsky JS, Giovannoni G, et al. Contribution of relapse-independent progression vs relapse-associated worsening to overall confirmed disability accumulation in typical relapsing multiple sclerosis in a pooled analysis of 2 randomized clinical trials. JAMA neurology. 2020;77(9):1132-1140.
9. Butzkueven H, Chapman J, Cristiano E, et al. MSBase: an international, online registry and platform for collaborative outcomes research in multiple sclerosis. Multiple Sclerosis Journal. 2006;12(6):769-774.
10. Šťastná D, Drahota J, Lauer M, et al. The Czech National MS Registry (ReMuS): Data trends in multiple sclerosis patients whose first disease-modifying therapies were initiated from 2013 to 2021. Biomedical Papers. 2024;168(3).
11. Jacobs LD, Wende KE, Brownscheidle CM, et al. A profile of multiple sclerosis: the New York state multiple sclerosis consortium. Multiple Sclerosis Journal. 1999;5(5):369-376.