I recently attended a Wellcome workshop on bridging the gap between computational models in neuroscience and clinical practice. It was really insightful to hear views from experts from a range of different backgrounds on what the barriers are to implementing their research into clinical settings and the various ways that this could be resolved. The discussions centered around several key barriers, each shedding a light on the multifaceted nature of this critical issue. In this post, I will talk a bit about each of these barriers which were discussed and their implications for the field.
Reproducibility of Findings:
One of the biggest challenges highlighted during the workshop was the issue of reproducibility. Most research studies in neuroscience often have relatively small sample sizes and are carried out on very specific population cohorts. This can create problems when trying to translate these findings into practice as the findings in these specific populations don't match that of the real world population. It is crucial that research is carried out and validated on large enough collections which are representative of the population so that they are robust and can be confidently applied. But it wasn't only this issue with reproducibility; the lack of transparent and well documented methodologies was also a concern as this is imperative to the replicability of research. For research in AI and computational models, publishing code can be really beneficial to reproducibility and testing the robustness of these models in different populations. This should really be encouraged more in the field so that research is transparent and we can take those methods and test them.
Representation of Diverse Patient Populations:
Recognising the diversity of patients, not only different ethnicities, but also the heterogeneity of diseases, is an essential aspect of implementing computational models in clinical settings. One point that kept on coming up was the fact that most of these studies are done on W.E.I.R.D patients (Western Educated Industrialised Rich and Democratic). In fact it was pointed out that a study in 2010 found that 96% of studies represent just 12% of the world population, so how can we assert that a particular behaviour is universal when the studies were based on the sampling of one specific subpopulation. Now we have moved to more diverse data since then but this is still an issue that we face. This was an important point in the workshop as in order for these models to be implemented into clinical practice, we need to ensure that they are not biased and therefore don't have the risk of misdiagnosing a patient just because of their ethnicity. When I attended the House of Lords on discussions around unlocking the potential of AI in health, this was also one of the top concerns among members of parliament. There are laws in place currently which ensure that AI which is implemented is not discriminatory in anyway so having diverse data is going to be really important moving forward for being able to implement these models.
Real World Data:
This point, very similar to the last, was about making sure that the models we develop are generalisable and aren't just specific to the population they are developed on. Moving beyond population-specific studies, the workshop called for a shift towards replicating results in real-world data. This approach provides a more comprehensive and accurate representation of how computational models perform in actual clinical scenarios. This shift in focus aligns with the practical demands of clinical practice, ultimately enhancing the reliability and relevance of these models. There are studies like the QMIN-MC (Quantitative MRI In NHS Memory Clinics) study which aim to collect this real-world memory clinic data for the purpose of being able to validate these AI models on dementia patients, but we need to be seeing more studies like this in other fields other than dementia.
Patient Perspectives and Involvement:
The workshop highlighted the significance of involving patients in the research process. Engaging patients in choosing research that directly benefits them not only empowers them but also ensures that the research conducted is aligned with their needs and priorities. This patient-centric approach fosters a more meaningful and impactful integration of computational models in clinical practice. The workshop also emphasised the need for transparent communication with patients about the use of computational models and their impact on diagnosis and treatment. Understanding and incorporating patient perspectives on AI in healthcare is crucial for building trust and ensuring the ethical application of these powerful tools. Giving patients a basic understanding of what these computational models are and how they are beneficial is also important as we found out in our PPI workshop that if people don't understand what AI is, then they are going to be extremely reluctant for them to be implemented in to practice.
Addressing Disease Heterogeneity:
To develop more informed and effective models, it is imperative to understand the inherent heterogeneity of diseases. The workshop emphasised the need for comprehensive data sets that capture the dynamic nature of diseases over time. Longitudinal data collection was advocated as a means to unravel temporal dynamics and enhance the accuracy of computational models. Stratification of patients to unravel this heterogeneity can be beneficial for specialised treatment in mental health to improve patient outcomes.
Fostering Data Accessibility and Fairness:
Building computational models requires large amounts of data which is why data sharing was an important aspect that people considered when doing this kind of research. Data should be Findable, Accessible, Interoperable, and Inclusive (FAIR) to facilitate collaboration and accelerate progress in clinical neuroscience. Platforms like the Dementias Platform UK Data Portal exemplify this commitment to fostering an environment for encouraging studies to share their data in a secure way which allows other researchers to utilise that data and build these models.
Integration of Multimodal Data
Lastly, the workshop stressed the importance of integrating data from multiple modalities. By combining genomics with imaging, blood tests, and other sources of information, a more holistic and comprehensive understanding of neurological disorders can be achieved. This integrative approach holds great promise for advancing our ability to predict, diagnose, and treat complex neurological conditions.
Conclusion
The workshop on "Bridging the Gap" provided a thought-provoking platform to explore the complexities of implementing computational models in clinical settings. By addressing these key barriers - from reproducibility and diversity to ethics and data integration - we will be in the position to implement effective and personalised healthcare in neurodegenerative diseases and psychiatric disorders. My research focuses on addressing these points so that we can come closer to this reality.
Kommentare