By using a machine learning-based model, researchers accurately predicted the person at early stage of Alzheimer’s disease would remain stable or progress over time. The team suggests this prediction tool is easier to scale up and has a strong potential to transition to clinics globally. This would also help clinicians make more accurate and timely decisions on which intervention approaches benefit the patient the most down the road. 

Alzheimer’s is a progressive condition. Disease-related changes in the body can occur at least 10 years before the first cognitive symptom appears. For instance, individuals with preclinical Alzheimer’s show healthy cognitive ability, while the biological markers of the disease already present in their body, such as the brain and blood. 

Recently, the U.S. Food and Drug Administration (FDA) approved two drugs, donanemab and lecanemab, to reduce cognitive ability decline and slow the disease progression. These two drugs target patients with mild cognitive impairment or mild dementia due to Alzheimer’s, which are the early stages of Alzheimer’s. 

All of these mean that identifying at-risk individuals early, accurately, and intervening the condition at an appropriate timing are crucial to gain more time for them to still live a regular life. 

Researchers have long been looking for solutions, but there are challenges. Positron emission tomography and cerebrospinal fluid test with biological markers have been the standard methods to diagnose Alzheimer’s. However, they also created the barriers for patients and clinics, as these two methods are not easy to access, invasive to patients, and expensive. In recent years, scientists have also embraced machine learning technology to develop mathematical models to predict disease patterns based on a large amount of patients’ data. But so far, the machine learning-powered approach also showed limitations. For example, the data researchers used to build the models were from positron emission tomography or cerebrospinal fluid test, or lacked diversity in populations. The data used for training the models were not from longitudinal studies that follow up study participants for a long period of time. These limitations would restrict the use of the technology in real world. 

That is a long list of challenges. Yet, researchers from the UK, Australia, Singapore, and the US developed a machine learning model, called predictive prognostic model, to address some of these challenges.

Building a machine learning model includes training, validating, and testing the model with data before using it in real world. To do so, the researchers used data from three databases. They are the Alzheimer’s Disease Neuroimaging Initiative (ADNI) in the US, Quantitative MRI of Brain Structure and Function in NHS Memory Clinics (QMIN-MC) in the UK, and Memory Ageing & Cognition Centre at the National University of Singapore (MACC). Compared to the ADNI, the researchers stated, the data from QMIN-MC and MACC reflect patients’ diversity in the real world better. 

Specifically, they used data from patients’ first or baseline brain magnetic resonance imaging scans and cognitive tests from ADNI to train the predictive prognostic model. Then they used QMIN-MC and MACC to validate and test the model. To predict future decline of cognitive ability, they created the prognostic index based on the model, which is a machine learning-guided marker. The higher the index number is, the more likely a person will develop future decline of cognitive ability.  

Using the predictive prognostic model, the researchers predicted individuals at early stage of Alzheimer’s remain stable or progress to Alzheimer’s, with 81.66% accuracy, 82.38% sensitivity, 80.94% specificity, and the area under the curve of 0.84. They considered this level of prediction as robust. 

The team also found that the index numbers in individuals who progressed to Alzheimer’s were notably higher than those remained at the early stage. 

The index was also more accurate in predicting progression to Alzheimer’s, compared with using the standard clinical markers, such as grey matter decrease and cognitive score, to diagnose the condition. This could reduce misdiagnosis for patients at early stage of the disease.

The team used patients’ data from non-invasive, more accessible, and lower-cost methods, including magnetic resonance imaging and cognitive tests, to develop this model. They tested the model with data from the real world with diverse patient populations. These make the model easier to scale up and enter the clinic globally.   

To move clinical use of this model forward, researchers would need to train and test it with larger real-world databases from healthcare systems in different countries.   

This study was published in eClinicalMedicine. Image credit: Canva

Got any comments or questions on this article? Please reach out to me on LinkedIn, X (Twitter), or the CONTACT ME page, I would love to hear thoughts from you. 

Share this article: