The development of an automated machine learning pipeline for the detection of Alzheimer’s disease
In this proof-of-concept study, we demonstrate that quantitative analysis of brief (5 min) resting-state EEGs in the frequency domain using a portable low-density (14-channel) montage reveals significant differences between AD and HC patients. Additionally, a transparent and explainable machine learning approach guided by conventional statistical methods to identify relevant data features in specific channels and frequency intervals based on empirically significant values results in classifier models that can distinguish subjects in the HC or AD categories with high accuracy.
Alzheimer’s disease is the most common cause of dementia in the elderly but lacks treatments capable of slowing the progression of the disease1. The lack of reliable endpoints and/or biomarkers contributes in part to the lack of effective therapies12. Functional imaging studies could provide information, however, further assessment of brain activity at the rate of neuronal activity is needed. Moreover, a qualitative, definitive and more complete diagnosis of AD, especially at an early stage before neural cell death, would open up more possibilities for targeted therapeutic interventions aimed at neuroprotection, thus potentially delaying the progression of AD. before major deficiencies appear.seven.
Several EEG studies have been conducted to detect brain function abnormalities in patients with AD13,15,35especially in the early stages7,16,17. While most studies have focused on the analysis of related evoked potentials (ERPs) in the EEG as subjects are engaged in various cognitive tasks to identify disturbances in specific cognitive processes16, the experimental constraints imposed by these paradigms could exceed the tolerance and capacity of elderly subjects. Conversely, resting state protocols (as in this study) are simpler, shorter, and easier to implement. According to various opinions7,16,18, the most commonly reported resting-state EEG findings are the generalized slowing of brain activity in the frequency domain in patients with AD. Specifically, progression to AD is characterized by an increase in low-frequency power (delta and theta bands), accompanied by a decrease in high-frequency power (alpha, beta, and gamma). Our results are largely consistent with these changes; while no significant change was noted between HC and AD patients in delta bands, we observed the expected increase in theta bands and decrease in alpha and beta bands in AD patients.
It is important to note that most AD research involving EEG relies heavily on qualitative examination of raw traces to remove artifacts as a first step, a subjective procedure that undermines rigor. Therefore, automated and quantitative EEG analysis is essential for objectivity and reproducibility in the evaluation of EEG data. Using an analytical pipeline developed by our team based on ML, our first step in EEG analysis is quantitative, automated and efficient. Additionally, the development of ML techniques has enabled more sophisticated analysis of EEG in frequency domains, thus enabling classification methods, such as decision trees, support vector machine, nearest neighbors K and linear discriminant analysis to more accurately identify patients with AD.27,28,29 and distinguish between AD and healthy subjects30,31,32., with some classifiers likely achieving sensitivity and specificity as high as 90%35.36.
Nevertheless, several issues hamper the clinical use of resting EEG for AD screening, as discussed in our abstract. In this study, we present a fully automated framework that overcomes these problems simultaneously for the first time:
a lack of automation and unbiased artifact removal, overcome by implementing automatic artifact removal through SVMs.
a reliance on a high level of expertise in data preprocessing and ML for non-automated processes – our analytics pipeline negates the need for such a high level of expertise as depicted in Fig. 3.
the need for very large sample sizes and precise EEG source localization using high density systems – we demonstrated good results with a sample of 41 patients using only 14 channels.
and a reliance on ML black box approaches such as deep neural networks with inexplicable feature selection – we used statistics-guided PSDs for feature selection for increased interpretability, which was input into a model of logistic regression that offers greater interpretability than many more complex ML models.
Despite promising results, this study still has several limitations. Our sample size was relatively small; Large-scale multicenter data are needed to further assess the generalizability of our model. Second, integrating multimodal data into ML models maximizes the chances of discovering meaningful biomarkers.37. The inclusion of vital signs, genetic data and comorbidities, as well as EEG, can lead to more accurate biomarkers. Third, although the average age of our study groups was different, the two populations overlapped in terms of age (healthy controls = 65.5 ± 6.8 years and AD = 75.7 ± 7.5 years). ); moreover, this small age difference is unlikely to be sufficient to explain the significant differences in the EEGs of these patients. Finally, our pipeline has not been validated on an external and independent dataset, which would increase the generalizability of our results.
Although our machine learning approach can be framed as conventional, the novelty of our approach is twofold: (1) “transparent” machine learning techniques as opposed to black-box deep learning methods, and ( 2) the preprocessing of EEG signals in an automated way to remove artifacts so that our results are reproducible, rigorous and scalable. These two new aspects allowed us to obtain proof-of-concept data in a relatively small sample.
In summary, we explored the development of a fully automated discrimination process for AD based on brief resting-state EEG epochs using low-density channel editing, an automated end-to-end analysis pipeline. end-to-end for data pre-processing and statistically-guided feature extraction, leading to explainable ML classification with high accuracy. Therefore, this study presents a proof of concept for a scalable technology that could potentially be used to diagnose AD in the clinical setting as an adjunct to conventional neuropsychological testing, thereby improving the efficiency, reproducibility, and convenience of AD diagnosis. MY. Further evaluation and testing in larger datasets is needed to further validate our results.