Global study shows longer brain scans lower research costs, provide more accurate predictions

Artificial intelligence (AI) models trained on large datasets are increasingly seen as the key to unlocking personalized treatments for brain disorders. An important bottleneck for scaling AI is the cost of data collection. This raises a fundamental dilemma: is it more cost-effective to scan more people for a short time, or fewer people for longer?

A study, published in the journal Nature, led by Associate Professor Thomas Yeo from the Center for Sleep and Cognition, Yong Loo Lin School of Medicine, National University of Singapore (NUS Medicine), now offers a clear answer: 30-minute functional MRI (fMRI) scans deliver up to 22% in cost savings while still retaining or even improving prediction accuracy.

Traditional thinking in neuroscience emphasizes collecting massive datasets by scanning thousands of people for brief durations, usually around 10 minutes for fMRI. AI models can then be trained to use the brain scans to make predictions of individual-level traits or outcomes. These traits and outcomes might include cognitive abilities (e.g. memory, executive function), mental health indicators and clinical outcomes (e.g. risk of Alzheimer’s disease).

Yet as participant numbers climb, so do the costs: even a brief scan can turn expensive once the hidden costs of recruiting, scheduling, and administratively tracking those volunteers are factored in. Short scans also may not capture enough high-quality information to make reliable personalized predictions.

The team posed a practical question: what if we focused on scanning fewer individuals, but for longer periods? Working with collaborators around the world, including Professor Thomas Nichols from the University of Oxford and Professor Nico Dosenbach from Washington University in St. Louis, the researchers developed a mathematical model that predicts how changes in scan time and number of participants affect the performance of brain-based AI models.

They validated their model using nine international imaging datasets encompassing thousands of individuals of varying ages, ethnicities, and health statuses. They found that their model can be used to customize study design to maximize prediction accuracy and minimize cost. Scanning each person for 30 minutes provides a sweet spot to maximize prediction accuracy and minimizes research costs.

“For years, the mantra has been ‘bigger is better.’ We’ve chased ever-larger cohorts without asking how long each person should be scanned. We show that in brain imaging, ‘bigger’ doesn’t have to mean larger cohorts. It can also mean more data per person,” said A/Prof Yeo. “In essence, we can get the best of both worlds—better prediction at a lower cost.”

This finding could reshape how researchers design neuroscience and mental health studies, especially for hard-to-recruit populations, such as patients with rare neurological conditions.

The team is now refining their model using real-world clinical data and emerging brain imaging technology. Their goal: make it even easier for researchers and health systems worldwide to design smarter, more cost-effective brain studies.

By helping studies collect better data for less money, the work could shape future research in neurology and psychiatry—and guide national and global efforts to deliver more personalized, affordable health care.

Professor Nico Dosenbach, a neurologist from Washington University in St. Louis, a co-author of the study, added, “This is a game-changer for the field. It gives research teams a rigorous, quantitative way to design smarter studies, especially critical as we move toward precision neuroscience. Longer scans mean better estimates of brain connectivity, which translates into more reliable links to cognition and clinical symptoms.”

The study was jointly first authored by Dr. Leon Ooi, Dr. Csaba Orban, Dr. Shaoshi Zhang, research fellows in the laboratory of Associate Professor Thomas Yeo, who is the senior and corresponding author of the study.

Artificial intelligence (AI) models trained on large datasets are increasingly seen as the key to unlocking personalized treatments for brain disorders. An important bottleneck for scaling AI is the cost of data collection. This raises a fundamental dilemma: is it more cost-effective to scan more people for a short time, or fewer people for longer?

A study, published in the journal Nature, led by Associate Professor Thomas Yeo from the Center for Sleep and Cognition, Yong Loo Lin School of Medicine, National University of Singapore (NUS Medicine), now offers a clear answer: 30-minute functional MRI (fMRI) scans deliver up to 22% in cost savings while still retaining or even improving prediction accuracy.

Traditional thinking in neuroscience emphasizes collecting massive datasets by scanning thousands of people for brief durations, usually around 10 minutes for fMRI. AI models can then be trained to use the brain scans to make predictions of individual-level traits or outcomes. These traits and outcomes might include cognitive abilities (e.g. memory, executive function), mental health indicators and clinical outcomes (e.g. risk of Alzheimer’s disease).

Yet as participant numbers climb, so do the costs: even a brief scan can turn expensive once the hidden costs of recruiting, scheduling, and administratively tracking those volunteers are factored in. Short scans also may not capture enough high-quality information to make reliable personalized predictions.

The team posed a practical question: what if we focused on scanning fewer individuals, but for longer periods? Working with collaborators around the world, including Professor Thomas Nichols from the University of Oxford and Professor Nico Dosenbach from Washington University in St. Louis, the researchers developed a mathematical model that predicts how changes in scan time and number of participants affect the performance of brain-based AI models.

They validated their model using nine international imaging datasets encompassing thousands of individuals of varying ages, ethnicities, and health statuses. They found that their model can be used to customize study design to maximize prediction accuracy and minimize cost. Scanning each person for 30 minutes provides a sweet spot to maximize prediction accuracy and minimizes research costs.

“For years, the mantra has been ‘bigger is better.’ We’ve chased ever-larger cohorts without asking how long each person should be scanned. We show that in brain imaging, ‘bigger’ doesn’t have to mean larger cohorts. It can also mean more data per person,” said A/Prof Yeo. “In essence, we can get the best of both worlds—better prediction at a lower cost.”

This finding could reshape how researchers design neuroscience and mental health studies, especially for hard-to-recruit populations, such as patients with rare neurological conditions.

The team is now refining their model using real-world clinical data and emerging brain imaging technology. Their goal: make it even easier for researchers and health systems worldwide to design smarter, more cost-effective brain studies.

By helping studies collect better data for less money, the work could shape future research in neurology and psychiatry—and guide national and global efforts to deliver more personalized, affordable health care.

Professor Nico Dosenbach, a neurologist from Washington University in St. Louis, a co-author of the study, added, “This is a game-changer for the field. It gives research teams a rigorous, quantitative way to design smarter studies, especially critical as we move toward precision neuroscience. Longer scans mean better estimates of brain connectivity, which translates into more reliable links to cognition and clinical symptoms.”

The study was jointly first authored by Dr. Leon Ooi, Dr. Csaba Orban, Dr. Shaoshi Zhang, research fellows in the laboratory of Associate Professor Thomas Yeo, who is the senior and corresponding author of the study.