Millions of people are diagnosed with Alzheimer’s disease each year, comprising 60% to 70% of dementia cases worldwide. While cognitive impairment and structural brain changes are indicative of Alzheimer’s disease progression, the process to accurately predict who will develop Alzheimer’s is time-consuming and requires a variety of techniques, including brain imaging, blood biomarkers, and neurocognitive testing by a neuropsychologist.
While deep learning models have shown some success in predicting Alzheimer’s diagnosis, the few deep learning models that predict cognition require expensive multimodal neuroimaging and longitudinal data. Additionally, the heterogeneity of Alzheimer’s makes early detection difficult, so qualitative measures like cognitive assessment remain critical for better prognosis, disease trajectory tracking, and clinical trial participation. Current neurocognitive batteries done by a neuropsychologist are time-consuming, and patient access to testing can be a challenge.
And while MRI is the most clinically accessible testing measure, on its own it struggles to capture variability in clinical progression and cognitive impairment when used in deep learning frameworks.
In order to achieve accurate prediction of cognitive impairment associated with Alzheimer’s without extensive test measurements, researchers from UC San Franciso proposed a multitask deep learning strategy that leverages specialized domain knowledge, custom-built deep learning models, and large pretrained models to predict cognitive scores using only a baseline MRI and relevant demographics. The key innovation in their strategy is an image model trained to achieve adjacent tasks – namely tissue class segmentation of brain images into gray/white/cerebrospinal fluid – designed to overcome the limitations of off-the-shelf AI models.
In a study publishing May 18 in Nature Aging, this domain knowledge-informed, multitask framework outperformed all existing AI methods, including standard transfer learning (a type of machine learning used with limited data). It produced accurate and field-leading prediction of multiple clinically relevant outcomes, including Alzheimer’s diagnosis, tissue segmentation and both current and future cognitive scores, all from a single baseline scan.
“Unlike previous approaches, our model does not require baseline cognitive assessment, specialized image pipelines, expensive PET scans, genetic analysis, or fluid proteomics, making it a fast, accurate, and easily implementable tool for most clinical settings,” said senior study author Ashish Raj, PhD, UCSF professor of Radiology and Biomedical Imaging. “Our technique provides other users, especially clinicians, the opportunity to benefit from implicit spatial brain representations learned by our proposed models, without requiring expertise in these computational pipelines.”
The researchers used data for training, testing, and validating from the Alzheimer’s Disease Neuroimaging Initiative database, which included demographics, MRI, diagnosis, and cognitive assessment. They incorporated data from adult brain scans from the Human Connectome Project Young Adult cohort in training to expose the model to adult brain scans which typically exhibit minimal or no atrophy. In addition, they collected an external testing dataset from the Dallas Lifespan Brain study, which contained multi-modal data on aging subjects, including MRI and cognitive scores. The researchers found that this strategy improved the generalizability and robustness of the segmentation models, and reduced susceptibility to segmentation errors in downstream tasks.
“We reported meaningful gains in speed and performance over other pipelines, which could prove valuable in developing a quick clinical prediction of cognitive impairment prior to referring the patient to a more advanced imaging lab and/or a full neuroradiology report,” said study first author Daren Ma, MSc, machine learning specialist in the Raj Lab at UCSF. “This circumvents the need to employ highly specialized, time-consuming, and computationally demanding MRI morphometry software and has broad implications for early diagnosis, prognosis, and clinical trial design.”
Predict Parkinson’s, ALS, and Huntington’s
The authors believe their work may help better characterize the link between morphology and cognition beyond Alzheimer’s, such as other neurodegenerative diseases like Parkinson’s, ALS, and Huntington’s disease. Prediction of baseline cognition may also be helpful in community settings where the lack of specialized neurocognitive assessment skills is a huge obstacle. Additionally, the study also provided a way of forecasting longitudinal cognition changes using minimal input data.
“The ability to correctly predict progressors from non-progressors using only baseline data can dramatically reduce sample sizes and cost,” said Raj. “Our model may also have potential as a tool for patient selection and progression tracking in large clinical trials of disease-modifying drugs,”
The authors say a future model could include additional measurements where they are available to further improve clinical utility and aid the prediction of cognition. These include longitudinal MR and PET, genetics, blood and CSF protein biomarkers.
“Clinical utility in practice will depend on specific use cases and environments and will require careful assessment in future studies,” Raj said.