Dementia is a series of neurodegenerative disorders that affect 1 in 4 people over the age of 80 and can greatly reduce the quality of life of those afflicted. Alzheimer’s disease (AD) is the most common variation, accounting for roughly 60% of cases. The current financial cost of these diseases is an estimated $1.3 trillion per year. While treatments are available to help patients maintain their mental function and slow disease progression, many of those with AD are asymptomatic in the early stages, resulting in late diagnosis. The addition of the routine testing needed for an effective level of early diagnosis would put a costly burden on both patients and healthcare systems. This research proposes a novel framework for the modelling of dementia, designed for deployment in edge hardware. This work extracts a wide variety of thoroughly researched Electroencephalogram (EEG) features, and through extensive feature selection, model testing, tuning, and edge optimization, we propose two novel Long Short-Term Memory (LSTM) neural networks. The first, uses 4 EEG sensors and can classify AD and Frontotemporal Dementia from cognitively normal (CN) subjects. The second, requires 3 EEG sensors and can classify AD from CN subjects. This is achieved with optimisation that reduces the model size by 83×, latency by 3.7×, and performs with an accuracy of 98%. Comparative analysis with existing research shows this performance exceeds current less portable techniques. The deployment of this model in edge hardware could aid in routine testing, providing earlier diagnosis of dementia, reducing the strain on healthcare systems, and increasing the quality of life for those afflicted with the disease.
1 Introduction
Dementia is defined as a series of disorders which progressively affects a person’s cognitive abilities and is present in roughly 25% of those over the age of 80 [1]. The most common form is Alzheimer’s disease (AD), representing roughly 60% of cases. This is followed by Frontotemporal (FTD) representing roughly 10% of cases [2]. The exact cause of FTD is not fully understood however AD is caused by the build of plaques around the posterior region of the brain that interfere with the electrical transmission between neurons [3]. Early stages of the disease can often be asymptomatic, leading to late diagnosis. While dementia is currently incurable, the efficacy of treatment that maintains mental function and slows the progression of the disease is greatly increased with early diagnosis [4].
Fig. 1.
Depicts the proposed Framework for Dementia Classification. Given raw Electroencephalogram (EEG) individually ranked features are extracted before being combined and iteratively selected. A tuned Neural Network is then optimised for EDGE operation using TensorFlow Lite.
×
The current state of the art for dementia diagnosis is a combination of Medical History and cognitive testing [5], Magnetic Resonance Imagery [5], Positron Emission Tomography (PET) scans [6], Cerebral Spinal Fluid [7], and the identification of specific biomarkers found during blood testing [8]. Collectively these diagnostic methods are effective, however, the overall cost and intrusive nature add increased pressure on the patient and the healthcare system.
Anzeige
This highlights the need for an accurate, easily accessible, automated solution for the routine screening of dementia in primary care centres, this would in-theory function as a triage system to refer patients for further testing. Advances in this field have the potential to reduce the load on healthcare systems, aid in early diagnoses of dementia and therefore enhance the quality of life for those afflicted.
This paper presents a model for dementia classification using Electroencephalogram (EEG) data. The framework, depicted in Fig. 1, is our process for feature extraction, dimensionality reduction, and modelling as designed for deployment in an edge device, minimizing electrodes, without sacrificing classification accuracy. The edge device acts as a routine testing platform for early dementia identification, releasing healthcare resources to enhance patients’ quality of life.
The hypothesis, was to examine if a framework can be developed to produce a lightweight machine
learning (ML) model designed for deployment in a minimal edge device with a view to mass routine
patient screening and triage of potential dementia cases, this would be particularly beneficial in cases where a patient is asymptomatic or shows sign of Mild Cognitive Impairment (MCI). Cases could then be referred for further gold-standard screening in the effective diagnosis of dementia. The primary contributions of this research are in the framework for modelling dementia, the required electrode inputs and their features selected, and the Long Short-Term Memory (LSTM) architectures attained.
The rest of the paper is laid out as follows: Sect. 2 reports on the related work in the areas of diagnosis of dementia, feature selection, and Machine Learning (ML) models of dementia is covered. Section 3 details the research methodology used to build and evaluate this model, and its implementation. The results are analysed and discussed in the Sect. 4 and Sect. 5, respectively. Before concluding this work in Sect. 6.
Anzeige
2 Related Work
There is a wide variety of screening techniques used to assist in the diagnosis of dementia. Ranging from non-intrusive to intrusive methods, the journey to an accurate diagnosis usually begins in primary care centres with psychometric tests to detect a pattern of loss of skills and function. More recently, blood test biomarkers have shown to be objectively measurable characteristic used to indicate a pathogenic process to improve diagnostic accuracy [9]. However, in the absence of addition intrusive measures measure such as Cerebrospinal (CSF) fluid examination, and non-intrusive Positron Emission Tomography (PET), Computed Tomography (CT) and Magnetic Resonance Imaging (MRI), these techniques lack the holistic approach needed for an accurate and early diagnosis [10]. This section discusses the relevant related work reviewed in order to devise a novel approach for creating the proposed dementia model.
2.1 Psychometric Tests
Cognitive screening tools can provide a time-efficient and objective initial evaluation of cognitive function. For example, the Mini-Mental State Examination (MMSE) is a standardised clinical test used to assess a patient’s cognitive impairment. It consists of eleven questions measuring the patient’s abilities in orientation, concentration, and attention. The precise ranges are debated as factors such as age and education influence the cutoff. Cognitively normal (CN) generally receives a score between 27 and 30. The cutoff for MCI is normally in the range of 21 to 23, and the middle to late stage of cognitive decline receives scores lower than the MCI cutoff [11].
An alternative to MMSE, is the Addenbrooke’s Cognitive Examination (ACE) [12]. The ACE assess six cognitive domains using 29 questions with a total 100-point test battery. The cutoff for MCI is 88, and the assessment technique is high reliability with a 93% sensitivity. Questions featured to Candidates include verbal analogies, arithmetic calculations, spatial relations number series puzzles, comprehension, and reading comprehension. Regardless of how well psychometric tests preform they should be considered as a component of a thorough assessment [13].
2.2 Medical Imaging
For a more holistic approach to diagnosing dementia, psychometric tests should be used in conjunction with Magnetic Resonance Imaging (MRI), Computed Tomography (CT), and Positron Emission Tomography (PET) scans. Authors in [14] researched the benefits of MRI and CT in diagnosing dementia. Their research showed CT imaging excelled in the identification cerebrovascular damage to the white matter of the brain, and MRI generated structural images of the brain can be used to study blood flow used in differentiating frontotemporal dementia.
In a similar vein researcher in [15] report on routine clinical use of hybrid PET/MR in patients with suspected dementia of patients in a memory clinic. The study revealed that a condensed hybrid PET/MRI protocol offers valuable supplementary information, significantly impacting clinical diagnosis and the management of patients in a considerable portion of cases, as compared to individual PET and CT scans. Evidently MRI, PET and CT scans are useful in detecting precursors to dementia, Alzheimer’s disease, Frontotemporal cognitive decline [6].
2.3 Biochemical Markers
The financial implications and potential for creating congestion within an overburdened health system, should medical imaging be used as a primary tool for the diagnosis dementia, is great [14, 15]. Biochemical markers such as Cerebral Spinal Fluid, which is a protective liquid flowing around the brain and spinal cord, have proven to be a frugal approach to screen for dementia. Unfortunately, biochemical markers need to be extracted from the body through invasive test, that either involve a spinal tap, lumbar puncture, or blood tests. Analysis of the fluid can determine the presence of certain biomarkers indicating certain dementia diseases with an accuracy in excess of 80% [7].
Taking a different approach author in [16] combined data from four distinct cohorts to assess the positive and negative predictive values of an Alzheimer’s disease (AD) blood test when applied in primary care settings. Utilising blood samples from 1329 participants, Random Forests analyses were employed to develop a blood screening tool. The tool demonstrated positive predictive values (PPV) and negative predictive values (NPV) of 0.81 and 0.95, respectively. For mild cognitive impairment, the PPV and NPV were 0.74 and 0.93, respectively.
Similarly, authors in [17] focused on the accumulation peroxidation of major phospholipids (e.g., phosphatidylcholine (PtdCho)) and degradation of antioxidative phospholipids (e.g., ethanolamine plasmalogen (PlsEtn)) in the brain. It was found that individuals with AD exhibited reduced levels of PlsEtn species in their plasma, particularly those containing the docosahexaenoic acid (DHA) component. Furthermore, patients with AD demonstrated lower PlsEtn levels and elevated PtdCho hydroperoxide (PCOOH) levels in their red blood cells (RBCs). In both AD and control blood samples, the levels of RBC PCOOH tend to align with plasma Aβ40 levels, and distinct correlations were observed between each PlsEtn species and plasma Aβ.
In consonant with the aforementioned findings’ authors in [8] reviewed 50 studies comparing concentrations of Aβ40, Aβ42, t-tau, and YKL-40 in 7303 patients to established, that blood testing provides highly accurate results (93%) in detecting dementia, with β-amyloid and t-tau being key biomarkers. Despite their intrusive nature, biochemical markers as a screening tool for dementia have their benefits. Although screening tools are assessed based on their effectiveness in accurately discerning individuals with dementia from those without, the psychological distress caused by their intrusive nature is often overlooked [18]. An overview of the intrusive nature, cost, and portability, of dementia screening techniques using Biochemical Markers and Medical Imaging is shown in Table 1.
Table 1.
An Overview of Diagnostic Methods.
Method
Invasive
Portable
Cost
Acc (%)
Spinal Fluid Test
Yes
No
Medium
80
MRI/CT/PET
No
No
High
89
Blood Test
Yes
No
Medium
93–95
2.4 Dementia Features and Modelling
Less intrusive, substantial more economical and more portable than medical imaging equipment is
the EEG. An EEG is a measurement device that detects electrical signals generated from brain activity and is a popular choice for a variety of applications including brain monitoring, task automation via human-machine interfaces, and emotion detection. Standardised methods for electrode placement are set by the International Federation of Clinical Neurophysiology and include the 10-20 and 10-10 systems with 19 and 63 sensory electrodes, respectively.
Several time and frequency domain features are thoroughly researched in academic literature. Research from Staudinger and Polikar, and Puri et al. [19, 20] showed the efficacy of classifying dementia with signal complexity, entropy, and Hjorth Parameters. When analysing EEG data in the frequency domain, Spiegel and Renna [21] noted that cognitive impairment due to various mental diseases can be “characterized by decreased power and coherence in the alpha/beta frequency band while increased power and coherence in the delta/theta band”.
Further research from Claudio Babilonia et al. [22] observed a particular correlation between AD patients and the posterior electrodes in the alpha frequency band. The performance of extracting power spectrum density features was also observed by Tavares et al. [23] and Alessandrini et al. [24]. Tavares et al. performed a backward wrapper method for feature selection to reduce dimensionality and testing with the best-performing models Linear Regression (LR) and Support Vector Machine (SVM). Alessandrini et al. applied Principal Component Analysis (PCA) to EEG data to reduce dimensionality and leveraged an LSTM to classify AD with an accuracy of 98%.
The use of PCA in research where the goal is the reduce the physical feature space would not be suitable as it would obfuscate the underlying sensor information however, the author’s proposed architecture was used as a base template in the work with multiple variations tested for each network type. An overview of key comparative ML methods or models can be observed in Table 2.
In summary, the current gold standard method for diagnosing dementia involves access to healthcare staff and laboratory resources for analysing the blood samples, which in the current context of scarce and overloaded healthcare systems is not viable or functional. Therefore, the use of EEG signals in combination with an LSTM model for screening of dementia promises a suitable solution for this problem.
3 Methodology
This section centres on describing the dataset employed by this work and the series of steps involved in the creation of the framework used to attain the proposed dementia model.
3.1 Dataset
A dataset containing EEG recordings from 88 subjects was sourced from [25]. Of these, 36 were diagnosed with AD, 23 with FTD, and 29 were said to be CN subjects. Information on the results of a Mini-Mental State Exam was included for each subject, providing a score ranging from 30 to 0 relating to their individual cognitive decline. The original recordings were sourced from the Department of Neurology in the general hospital of Thessaloniki, Greece. The EEG used was a Nihon Kohden 2100 with 19 scalp electrodes (Fp1, Fp2, F7, F3, Fz, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, and O2) and 2 reference electrodes (A1 and A2) according to the 10-20 international system. Clinical protocols were followed during the recording with each subject in the sitting position and their eyes closed. The average record time per subject was 13.5 min at a sampling frequency of 500 Hz, leading to a dataset with 35.6 million rows of information. Preprocessing steps for the dataset include a Butterworth band-pass filter of 0.5 Hz–45 Hz and were referenced to the A1 and A2 electrodes. Artifact removals were performed at the data source and included Artifact Subspace Reconstruction routine (ASR) and the RunICA algorithm for classifying eye and jaw artifacts via the automatic classification routine ICLabel.
3.2 Data Preprocessing
Data is segmented into training, validation, and test groups with a 60:20:20 random split based on subject identification numbers. This aids in building a more robust solution by ensuring none of the readings from a subject assigned to the training set are given to the testing model for inference. The dataset incorporates details such as age, gender, and severity. All efforts were made to ensure a fair distribution among test groups. Due to the limited sample size of 88 subjects, there is a potential for bias to permeate the framework, the authors implemented thorough measures to mitigate this risk. Keeping this in consideration, two binary classification models were developed using the dataset. The first will classify both FTD and AD from CN, and the second will classify AD from CN. A test set comprised of 19 subjects is reserved for the AD-FTD model, and a set of 12 subjects is reserved for AD only model.
Initial research suggested AD patients are more likely to be asymptomatic in the early stages of disease progression. The decision to build a separate model for AD and CN classification is made under the estimation that it would require less features, and therefore sensors, leading to a lower-cost platform that is more accessible.
3.3 Feature Extraction
A wide range of time and frequency domain features were extracted from the EEG data. The following is a summary of the 5 feature sets extracted at 2 and 4-s sampling windows.
Statistics extracted included minimum, maximum, mean, range, standard deviation, energy, and autocorrelation. This set totals 133 features.
The metric used to capture the tighter grouping of signal pairs is Mean Absolute Difference (MAD). This was extracted for each unique pair of sensors. This set totals 171 features.
Hjorth Parameters extracted include Activity, Mobility and Complexity. This set totals 57 features.
In calculating power spectrum densities, total, mean, and relative power (the sum of a frequency band relative to the sum of the power spectrum) are extracted. This is done using the Welsch method of deriving the mean of a series of Fast Fourier Transforms (FFT) to leverage the computational gains of the FFT when dealing with nonstationary signals [26]. This set totals 255 features.
Cross Coherence is a statistical measure that measures the linear relationship between signal pairs at a specified frequency band [27]. This is calculated for each unique pair of sensors, at each frequency band. This set totals 855 features.
3.4 Initial Feature Selection
The goal of this research is to use the least number of features and therefore sensory electrodes. The most accurate method of selecting features would be to evaluate each feature pair, triple, and potentially higher-order combination until no more performance gains can be made. The total number of features extracted is 1,471, generating 1.1 million unique pairs for testing. As this is an inefficient use of time and computing, the least number of features for each category set was reduced by first ranking importance with Mutual Information (MI) [28], and iteratively testing and adding new features with benchmark models Support Vector Classifier (SVC) and K-Nearest Neighbour (KNN) [20] until no more accuracy gains can be made.
Features selected from the previous stage are then grouped and ranked with both MI and ANOVA. The process of testing and iteratively adding and retesting features until no more performance gains can be made is repeated. This provides a core number of features and the process of building the neural network models can begin.
3.5 Modelling
In building a performant Neural Network (NN) for the classification of dementia, a series of artificial neural networks were tested. This includes a simple sequential network known for its ability to model complex nonlinear relationships, a recurrent network which excels at capturing temporal information, and a more advanced type of recurrent network, the Long-Short Term Memory (LSTM) network, with the ability to maintain information over extended sequences. Once a performant model was chosen, model tunning with a large set of hyperparameters was performed.
3.6 Final Feature Selection and Optimisation
The tuned model is chosen as the testing metric for the remaining features. Features are scaled by removing the mean and scaling to unit variance. The process begins with testing each feature in isolation, then each feature pair, and each potentially higher-order combination until each model no longer classifies subjects incorrectly. The resulting set is the least number of features and therefore sensory electrodes required by the model to correctly classify the target label. Google’s TensorFlow Lite (TFLite) library is leveraged to reduce model size and increase inference speed for edge deployment.
4 Results
This section presents results - graphs and tables - attained from executing each step depicted in the LSTM Framework, which was designed and followed as part of this work to create a model to classify dementia that will be deployed in an edge device. The scores and measurements attained help authors to make informed decisions at each step.
4.1 Feature Selection and Benchmark Results
The resulting reduction in dimensionality of the initial feature selection stage can be seen in Table 3. The modelling was sensitive to small changes, as such metrics Error Count and Mean Error were derived to produce more clarity into the underlying performance. Error Count refers to the number of test subjects incorrectly classified, and Mean Error refers to the mean unrounded prediction error.
Table 3.
Set Stage Feature Results.
Data
Method
Features Selected
Error Count
Mean Error
Acc (%)
Statistics
SVC
6
6
0.31
66
MAD
SVC
5
4
0.21
73
Hjorth
SVC
13
5
0.26
75
PSD
SVC
25
4
0.26
75
Coherence
KNN
24
5
0.26
71
Grouping previously selected features and ranking with ANOVA and MI before iteratively testing with SVC highlighted 57 features to be the least amount with the best performance. A confusion matrix of this result can be seen Fig. 2. The diagonal cells indicate true positives that are correctly classified. The off-diagonal cells indicate false positives that are incorrectly classified. The overall accuracy for the classifier was 0.93. The Precision and Recall were 0.94 and 0.84, respectively.
Fig. 2.
The Confusion Matrix for the proposed framework trained and tested on the OpenNeuro dataset of 88 EEG recordings from: Alzheimer’s disease, Frontotemporal dementia and Healthy subjects.
×
4.2 Initial Modelling Results
The results from building a simple sequential model showed a high mean error and a drop in accuracy compared to the SVC model. The recurrent model architecture of the base RNN and the LSTM showed a notable improvement in mean error and classification accuracy, an overview of which is provided in Table 4. A selection of hyperparameters tested is shown in Fig. 3, with time step and batch size both showing the best results from a value of 16.
Table 4.
Initial Neural Network Model Results.
Method
Error Count
Mean Error
Acc (%)
SEQ
0.0
0.41
83
RNN
2.0
0.24
84
LSTM
0.0
0.09
91
LSTM Tuned
0.0
0.09
95
Fig. 3.
Plot of Hyperparameter Results, showing the Regularization Accuracy, Epoch Accuracy, Time Step Accuracy, Batch Accuracy, Regularization Type Accuracy and Model Accuracy.
Fig. 4.
Box Plot Value Distribution of Selected AD and AD_FTD Features After Removing the Mean and Scaling to Unit Variance.
×
×
4.3 Final Feature Selection
The success criteria for testing sets in higher order combination with the tuned LSTM was for the model to exceed 95% accuracy and no longer incorrectly classify any of the subjects. This was achieved for the AD-only models with 2 features and for the AD-FTD model with 3 features, Table 5 and Table 6 highlight the ranking of the feature sets. These findings mirror research from [21] and [22], with particular importance from posterior electrodes on the alpha frequency band. Figure 4 shows the distribution of AD Model Feature Value Range and AD_FTD Model Feature Value Range, respectively.
Table 5.
Alzheimer’s Disease Model Feature Sets.
Features
Count
Error
Acc (%)
O2_alpha_rel_power, coh_O1vsT6_theta
0.0
0.305
98.1
coh_O1vsT6_theta, coh_O2vsF8_alpha
0.0
0.25
97.3
coh_O1vsT6_theta, coh_O2vsF8_beta
0.0
0.225
97
coh_C3vsT5_alpha, coh_O1vsT6_theta
0.0
0.245
95.4
Pz_alpha_rel_power, coh_O1vsT6_theta
0.0
0.3
95.2
Table 6.
Alzheimer’s Disease Frontotemporal Model Feature Sets.
Table 7 provides an overview of sensors required and optimisations gained from converting the standard TensorFlow model with TFLite. The lite version reduced inference speed by an average of 3.7×, and model size by 83× while maintaining prediction accuracy.
Fig. 5.
LSTM RNN Model Architecture with Time Distributed Wrapper Applied to Dense Layers. A Time Step of 16 Leverages 16 Rows of Sequential Data per Single Output. Shape is in the Format of Batch, Time Step, Neurons. The Dense output is in the form of a Binary Sigmoid to Produce Either a CN or AD Result.
Table 8.
Final Model Training Parameters
Parameter
AD Model
AD-FTD Model
Learning Rate
0.0005
0.001
Optimiser
Adam
Adam
Epochs
18
15
Shape
[2,4]
[3]
Batch Size
16
32
Activation
Elu
Elu
Times Step
16
32
×
Table 8 provides the final training parameters of both models, and Fig. 5 highlights the architecture of the AD model. This is identical to the AD-FTD model with some adjustments to the parameters listed in Table 7. The Confusion Matrix for the AD model and the AD FTD model are shown in Fig. 6.
Fig. 6.
From left to right, the Confusion Matrix for the Alzheimer’s Disease Model, and the Alzheimer’s Disease Frontotemporal Model, respectively.
×
5 Discussion
The research objectives was to design a performant neural network for the routine testing of dementia. By systematically extracting an extensive set of features, undergoing multiple meticulous feature selection phases, conducting model testing, and fine-tuning hyperparameters, this study introduces two innovative models designed for distinguishing between AD and CN, as well as AD-FTD and CN subjects. This was achieved with 3 and 4 sensory electrodes, respectively. This research also shows that through optimsations with TFLite, the models computational footprint is reduced without a trade-off of performance.
Table 9.
Comparative Analysis
Authors
Label
Electrodes
Method
Acc (%)
Tavares, G. et al. (2019)
AD
13
LR/SVM
97.1
Alessandrini, M. et al. (2022)
AD
16
LSTM
97.9
Puri. et al.(2022)
AD
6
SVM
97.5
This Research
AD
3
LSTM
98.1
This Research
AD-FTD
4
LSTM
98.5
Comparative analysis with existing research [23, 24], and [20] shown in Table 9, highlights the proposed models hold, or evens exceed existing performance accuracy with a significant reduction in sensor requirement. Research from [29] and [30] put forward designs for miniaturized, portable EEG headbands. The electrode placement of these designs may then be adapted to mirror the sensor findings of this paper, which are highlighted in Fig. 7. The resulting device may then function as a lightweight, portable, and accessible platform for the monitoring or routine testing of dementia, reducing the load on healthcare systems, and increasing the quality of life for dementia patients.
Fig. 7.
Electrode Placement shows the required electrodes and their placement with respect to the 10-20 system
×
6 Conclusion
This research analysed the efficacy of a machine learning framework for the early detection of dementia using sensory EEG information. The findings demonstrated through careful and thorough feature selection and modelling, that a lightweight LSTM can predict the presence of AD and FTD from CN subjects with an accuracy of 98%. This can be achieved with optimisation that significantly reduces storage and inference speeds without sacrificing accuracy. Comparative analysis with existing research revealed that these proposed models require significantly fewer electrodes and provide performance accuracy comparable to or in excess of less-portable techniques currently researched. Future work can be devoted to creating a low-cost sensor platform, which will allow us to capture fresh data and re-test this model and approach. The resulting deployment holds the potential for developing a practical and accessible tool for the early detection of dementia, which could significantly impact healthcare systems, and patient outcomes.
Acknowledgments
The authors would like to acknowledge the work by Miltiadous et al. [25] for providing the dataset source that made this research possible.
Disclosure of Interests
The authors have no competing interest to declare that are relevant to the content of this article.
Open Access This chapter is licensed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.
