|
Getting your Trinity Audio player ready...
|
Understanding the intricacies of the brain is essential for advancing treatments and improving patient outcomes in the rapidly evolving field of mental health and ageing-related disorders. This pursuit has taken a significant leap forward through innovative research at the Yong Loo Lin School of Medicine, National University of Singapore (NUS Medicine).
A dedicated team at NUS Medicine has developed a groundbreaking method that leverages artificial intelligence (AI) and brain activity data to reconstruct visual experiences directly from brain signals. Their approach uses AI models and unsupervised learning to decode images, videos, texts, and audio from brain recordings, enhancing the interpretation of fMRI data.
OpenGov Asia conferred its Recognition of Excellence Award to NUS for its groundbreaking initiative, which has the transformative potential to revolutionise early brain disease detection, personalise treatments, and enhance learning programmes. These advancements are poised to significantly boost patient independence and improve therapeutic outcomes.
The study focuses on enhancing the adaptability and interpretability of findings, crucial for broader clinical applications. It represents a paradigm shift in addressing mental health and ageing-related disorders, demonstrating the potential for significant advancements in medical technology through advanced AI models and interdisciplinary collaboration.
Led by Dr Helen Zhou Juan, an Associate Professor from the Centre for Sleep and Cognition and Director of the Centre for Translational Magnetic Resonance Research at the Yong Loo Lin School of Medicine, National University of Singapore, the team includes Dr Zijiao Chen, Research Fellow from the Centre for Sleep and Cognition at the Yong Loo Lin School of Medicine, National University of Singapore and Dr Li Ruilin, Research Fellow from the Centre for Sleep and Cognition at the Yong Loo Lin School of Medicine, National University of Singapore. Comprising members from diverse disciplines, this formidable multidisciplinary group addresses complex challenges with a collaborative approach.
In an exclusive interview with OpenGov Asia, the team shared details of their development process, highlighting innovative approaches, key breakthroughs, and significant challenges. Their insights offered a glimpse into the collaborative efforts, technical advancements, and personal experiences that have defined their groundbreaking work in brain research.
Articulating their vision, Dr Helen Zhou Juan shares that their overarching “goal is to decode brain activity to advance neuroscience and medicine, assist individuals with sensory or motor impairments, and promote precision medicine.”
During the study, participants viewed videos featuring moving objects, animals, and humans while their brain activity was monitored using functional MRI (fMRI). The collected brain activity data was meticulously processed using an advanced AI model called Stable Diffusion. This model decoded neural signals and translated them into reconstructed videos lasting approximately two to three seconds, accurately mirroring participants’ visual experiences.
While brain research itself is not new, their unique approach focuses on decoding brain activity through visual and auditory signal patterns. Initially reliant on public datasets, they are now positioned to collect local data in Singapore, advancing their work and tailoring models to better suit local populations.
“Over the past few years, we have made significant strides, building on previous efforts and other initiatives,” says Helen, underscoring the significant progress they have made. “We have published our first paper and have made substantial advances, including ongoing work in analysing and decoding existing figures as well as obtaining our own primary data.”
They differ from traditional methods by using self-supervised learning on large-scale fMRI recordings. This technique enables the team to work with vast amounts of data without the need for precisely paired datasets, which are difficult to obtain. Helen compares it to Generative AI models, “It is like a mini GPT, but trained with brain recordings instead of natural languages.”
A key goal is to enhance both understanding and inclusivity by improving the quality of training data for clinical applications with limited patient data. To address the high costs and limitations of MRI, the team is investigating portable devices like EEG as complements to fMRI, aiming to make research more practical, scalable, and cost-effective by increasing sample sizes.
Additionally, a significant challenge in brain research is the limitation of training data, which affects model performance and raises concerns about bias. The predominance of Caucasian data often restricts the generalisation ability of findings. To counter this, the team is incorporating data from Singapore’s three major ethnicities and considering various factors to enhance the inclusivity and applicability of their research.
Dr Zijiao says current models rely on data from a finite number of categories – such as images and videos from approximately a thousand categories. This limitation means that the models may not perform optimally when encountering data outside these categories, potentially leading to biased outcomes.
Moreover, the data collected is often limited to volunteers who may not represent the broader population. This inherent bias can affect the validity of the research outcomes, as laboratory studies are typically constrained by the diversity of participants and the specific conditions under which the data is collected.
Dr Ruilin notes that this is a common issue in research laboratories, where it is impractical to account for every possible variation in real-life scenarios. The bias inherent in the training data is a significant challenge, particularly when testing models with out-of-distribution data. This limitation reflects the broader issue of generalising findings from controlled research settings to applications.
Helen emphasises the crucial role of rigorous testing and validation of hardware and algorithms to mitigate biases. Harmonisation techniques and continuous validation are essential for ensuring the accuracy and interpretability of research results. Despite these efforts, a persistent challenge remains in ensuring that models and findings apply to diverse real-world contexts.
When discussing the clinical applicability of their models, Helen was candid about their current limitations. Despite the team’s significant progress, she acknowledged that the models are not yet ready for clinical use. “I do not think we are ready yet. Practical implementation is still a way out, depending on ongoing research advancements.”
They have filed for several patents and expect significant progress in the next six months to a year. However, acknowledging that practical outcomes will require collaboration and granting that achieving clinical application will require broader participation, they are committed to sharing their models for research.
“We welcome conversations and collaborations, both locally and internationally,” Helen emphasises, appreciating that sharing information can lead to valuable connections and developments. “This is a significant endeavour that will require collaborations with academics, funders, and the private sector.”
“We are always open to international collaborations. Sharing new datasets or algorithms helps us advance our research,” Dr Zijiao and Dr Ruilin agree. “We remain open to more international partnerships, recognising the value of diverse perspectives and expertise in advancing this undertaking.”
The team, which includes neuroscientists, psychologists, engineers, and machine learning experts, collaborates effectively towards a unified goal. Their wide approach addresses various aspects of brain research and applications, including projects on dementia, stroke, and psychosis that complement their brain-decoding efforts.
Leading a large, multidisciplinary team, especially in the field of brain research, where precision and coordination are crucial, comes with its own set of challenges. Helen shares valuable strategies for overcoming these obstacles, focusing on keeping the team both motivated and effective.
One critical method is aligning projects with team members’ interests, which helps sustain enthusiasm and efficiency. By bringing together skilled and dedicated individuals, she ensures that everyone remains committed to the team’s shared goals.
Given the complexity and detail of their project, it is essential to prevent working in siloes. With each member focusing on their specific area, there is a risk of losing sight of the bigger picture.
Helen utilises digital communication tools with specific channels for projects and as a common channel for broader discussions, streamlining team interactions and supporting continuous learning. This setup allows team members to efficiently share updates, ask questions, and engage in conversations despite their busy schedules.
Structured meetings and dedicated research channels further enhance collaboration, facilitating the exchange of ideas and advice. This approach helps prevent isolation and nurtures a culture of learning and innovation.
“We use platforms like Slack for both project-specific and general communication,” Helen explains, “Plus weekly in-person gatherings and research meetings ensure that everyone stays informed and can contribute feedback.”
Helen reflects on the thrilling “aha moments” that drive their research, often sparked by unexpected breakthroughs. These revelations can come from reading papers or pondering ideas in everyday life, offering surprising insights. She notes, “When you adjust a model and it unexpectedly delivers the exact result you hoped for, the sense of achievement is truly gratifying.”
Dr Zijiao also shared her “Aha!” moment while working on a particular iteration. She was initially concerned that their results would not match those of previous studies but was thrilled to find that they surpassed expectations by around 50%.
As they refine their models and explore practical applications, the potential for clinical impact remains promising. Helen anticipates expanding her team to include expertise for practical applications and industry collaborations, balancing research with real-world development, noting, “As we grow and explore new areas, we plan to integrate new members and disciplines to enhance our research and impact.”
The team anticipates breakthroughs with more advanced models and techniques as technology evolves, integrating precise digital variables with physiological signals to advance brain research and its real-world applications. Helen envisions a future driven by iterative progress and cross-disciplinary collaboration, with cutting-edge generative models and representation learning shaping the next generation of research.
Her approach includes seeking additional funding and engaging with industry partners to support these goals. By aligning team expansion with strategic objectives and external collaborations, she aims to enhance the impact and applicability of their research.
Insights from Helen Zhou Juan, Dr Zijiao Chen, and Dr Li Ruilin highlight both the ongoing challenges and future possibilities in brain research. Their work at NUS demonstrates the transformative potential of combining AI with neuroscience and medicine, positioning the team to drive significant progress and future advancements in brain decoding.
