In a recent systematic review published in BMC Neurology, researchers examined recent studies exploring neuroimaging modalities aided by advanced machine learning (ML) algorithms that predicted Alzheimer’s disease (AD) progression from mild cognitive impairment (MCI) to AD dementia.
MCI is the early stage of AD when a patient is symptomatic, but full-fledged AD dementia has not set in.
Additionally, they reviewed ML techniques that helped identify factors contributing to MCI to AD dementia transition.
Study: Neuroimaging and machine learning for studying the pathways from mild cognitive impairment to alzheimer’s disease: a systematic review. Image Credit: LightField Studios/Shutterstock.com
Background
AD is a progressive neurodegenerative disorder that begins decades before clinical diagnosis. AD-related damage to neurons and their connections hampers the daily life activities of the affected individuals, such as swallowing, walking, and talking.
As the number of older adults (aged 65 or above) increases worldwide, AD-related deaths might rise because the risk of death due to AD is directly associated with a patient's age. Currently, treatment for severe AD is unavailable.
Thus, researchers have been pursuing neuroimaging modalities, such as positron emission tomography (PET), single-photon emission computed tomography (SPECT), MRI (magnetic resonance imaging), functional MRI (fMRI), and electroencephalogram (EEG), to detect and predict AD progression.
They are also testing targeted interventions for individuals at high risk of progressing from MCI to AD dementia in clinical trials.
However, combining these modalities to obtain a more comprehensive understanding of structural and functional brain data is challenging, given the complexities of integrating multiple measures associated with the transition from MCI to AD dementia.
ML methods and deep learning algorithms could make it easier to analyze high-dimensional neuroimaging data and identify individuals at risk of developing AD dementia; however, there remains an unmet medical need for generalizable predictive ML-based models that aid the translation of these approaches into 'real-world' healthcare settings.
Another challenge with using these methods for AD prediction is that they require a large sample for training and tuning parameters, and available AD datasets have a small sample size.
About the study
Researchers followed the PRISMA guidelines for the current systematic review, wherein they extensively searched three electronic databases, PubMed, Scopus, and Web of Science, between January 1, 2017, and March 1, 2019, to identify the most recent scientific work focused on the prediction of conversion from MCI to AD.
All included studies used at least one type of neuroimaging modality, preferably using biomedical image processing with ML techniques, described the methodology in detail for enabling replication, and results in a way that facilitated the extraction of the accuracy, sensitivity, and specificity of the method used for comparison purposes.
The researchers deployed adapted search strategies for each database and even searched the reference lists of included studies for further analysis. Moreover, they did not impose any language restrictions.
The team used biomedical image processing and ML techniques for data extraction from all the identified studies.
Study data points included author, publication year, data source, follow-up time, sample size, neuroimaging features, modalities, data analysis techniques, and results. They used narrative synthesis to explore heterogeneity in studies.
Results
The majority of studies screening MCI patients at risk of AD dementia used structural MRI and PET scans. While these neuroimaging modalities returned data with high spatial resolution, these systems are expensive, tedious to maintain, and immobile, which limits the utility of prediction tools using these modalities.
In comparison, EEG, an affordable and easy-to-implement modality in real-world settings, might be more appropriate for screening MCI individuals. Yet, a single study used EEG to study AD progression.
The unavailability of longitudinal EEG datasets of MCI patients could be one of the driving factors of this unbalanced reality. Unfortunately, there is insufficient time to conduct new longitudinal studies using EEG.
The researchers noted that 80.3% of studies used the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database to monitor AD progression. This dataset had high-quality data that did not represent the overall population.
Further, most studies monitored the MCI to AD transition over a short time window of 6-36 months, implying some individuals classified as stable MCI had progressed to AD dementia already.
Furthermore, most studies used simple linear methods for classification, such as Support Vector Machine (SVM) and regression, when more sophisticated classification techniques are available.
In addition, all reviewed studies relied on two cross-validation techniques for accuracy evaluations, i.e., 10-fold cross-validation and leave-one-out cross-validation techniques.
Another observation was that most of the reviewed studies ignored using appropriate modalities and neuropathological features, likely because researchers in the field of computer science were more focused on developing their data analysis algorithms.
Thus, researchers emphasized the significance of a collaborative effort among computer science, neuroscience, and cognitive science researchers to overcome the current challenges in this relatively underdeveloped domain.
Conclusions
To summarize, the authors observed a concerning trend regarding data modality and data analysis methods currently used in studies monitoring AD progression.
They showed that most studies used costly and logistically challenging methods, which would be tedious to put to practical use in healthcare settings. They also noted that these studies used nonrepresentative participants and had shorter clinical follow-up periods.
Overall, there is a need for developing deep learning approaches that analyze brain scans with high precision and resolve issues related to neuroimaging modalities currently used.
It would enable the deployment of predictive models that identify AD progression at the early stage of MCI in real-world healthcare settings, opening new therapeutic avenues for those at the highest risk of AD dementia.
How did machine learning models perform in identifying hepatitis in patients with diabetes?