Predicting the conversion time from normal cognition to mild cognitive impairment based on dual attention convolutional networks

Abstract

Background

Dementia, a progressive neurological disorder, is a leading cause of disability and death globally, often underdiagnosed in its early stages. Early diagnosis, prevention, and treatment are crucial for mitigating its impact on individuals and society.

Objective

This study aimed to predict the exact conversion time from normal cognition (NC) to mild cognitive impairment (MCI), and to provide insights for early diagnosis and treatment of dementia.

Methods

A novel dual attention convolutional network model was proposed to handle high-dimensional features and limited patients’ records in short-sequence time series data. It integrated feature and temporal attention modules to capture dependencies and used a custom loss function to enhance clinical interpretability.

Results

The model significantly reduced mean squared error (MSE) by 9.67% and mean absolute error (MAE) by 26.24%, while increasing the r-square ( $R^{2}$ ) by 16.71% compared to a basic convolutional model. It effectively predicted NC to MCI conversion, offering valuable guidance for early intervention.

Conclusions

The dual attention convolutional network model effectively predicted NC to MCI conversion, providing a valuable tool for early dementia diagnosis and treatment.

Keywords

Alzheimer's disease attention mechanism deep learning loss function mild cognitive impairment

Introduction

Cognitive impairment refers to the deficits in cognitive domains such as memory, learning, language, and behavioral. Currently, World Health Organization (WHO) has concluded that dementia is underdiagnosed worldwide, and even if a diagnosis is made, the patient is typically at a relatively late stage.¹ Both the Alzheimer's Association and the WHO have set the early diagnosis of dementia as one of their principal goals.^1,2 Thus, early diagnosis is crucial for patients to improve awareness and receive timely treatment. Depending on the severity of cognitive dysfunction, cognitive impairment can be differentiated into mild cognitive impairment (MCI) and dementia.

Research in the field of dementia has mainly focused on classification of different stages, including binary and multi-class categorizations.^3,4 The patients of these studies tend to be static datasets, with fewer investigations addressing the progression of the disease in patients, i.e., dynamic data at different stages over time. Much of the research on dementia progression has focused on whether MCI progresses to dementia, which remains a binary classification issu.^5,6 Delving deeper into this question, discussions often extend to the likelihood of progression at a future point in time.^7–9 However, research specifically focusing on the exact timing of this transition is relatively scarce. To our knowledge, only three studies^10–12 have predicted the specific conversion time from MCI to Alzheimer's disease (AD), while no studies have predicted the specific conversion time to MCI in the normal cognition (NC) population. In the early diagnosis of dementia, timely warning of MCI is also important to raise awareness and early testing for etiologies, so we focus on predicting the specific conversion time to MCI in the NC population. Furthermore, the time series data processed in time-to-event prediction-related research are typically long or uniform in length, posing high demands on practical applications. These existing research methods fail to yield satisfactory results for small, irregular time series data, demonstrating a lack of generalization capability in the common real-world scenarios characterized by relatively shorter but irregular time series. To address this, we have innovatively designed an attention module that enhances the model's generalization to time series data of varying lengths.

Biomarkers based on different modalities, such as magnetic resonance imaging (MRI), positron emission tomography, and cerebrospinal fluid, have been extensively studied to predict dementia progression.¹³ Neuropsychological batteries, cognitive scores, and clinical detection measures have also been used to assist research. On the other hand, traditional machine learning algorithms and deep learning algorithms are widely used in MCI transformation research.^14,15 Conventionally, convolutional networks process image data¹⁶ and recurrent neural networks process time-series data.

This study used data from image variables of MRI scanning and clinical indicators, at dynamic time points from normal cognition to MCI. Compared to conventional time series data, our data had a relatively high feature dimension, a relatively shorter but irregular time series, and a relatively large time span unit. To meet these challenges, we used convolutional networks to process time-series data, proposed a one-dimensional convolutional network method based on dual-module attention, and adapted the backpropagation loss function optimization model. Specifically, we designed two parts of attention modules to calculate the weights of different dimensions and different time records. A feature attention module was built based on multiple convolutional layers to compute the correlation of input features at each time step, which could adaptively select the most relevant input features, and explicitly provide clinical diagnostic credibility. The temporal attention module was built based on the fully connected layer, which was used to compute the weight of each data set from different patients, capture the temporal dependence and weight the sum to facilitate the subsequent convolutional network processing. Subsequently, we used a one-dimensional convolutional network to integrate these final attention results and train the model together. Finally, considering the actual clinical significance of the prediction results, we defined a loss function to make the model and results more clinically interpretable.

The main contributions of this work are summarized below:

We proposed a framework for predicting when NC populations would transit into MCI using time-series data. To the best of our knowledge, this was the first study to predict the specific conversion time to MCI in NC populations. Compared with other prediction studies of conversion time, we also improved the accuracy of specific time prediction, providing a new solution for the early diagnosis of dementia.

The proposed framework offered new technologies for handling time-series data. It incorporated two attention modules that adaptively enhanced the extraction of information from important features and critical times. The attention-weighted results were processed using a convolutional network, treating high-dimensional short-range time-series data as image data, thereby effectively leveraging the inherent characteristics of the data, reducing model complexity, and improving prediction accuracy.

We designed a novel light convolutional attention mechanism for high-dimensional and short-range time-series data, which enabled a better understanding of both local and global contextual information and captured long-distance dependencies within the features.

We defined a loss function based on the practical clinical significance of the prediction results, which was primarily designed to reduce the relative error of the predicted value and provide personalized medical guidance for patients with varying degrees of condition severity. Additionally, the function separately addressed overestimation and underestimation scenarios, ensuring that the loss in underestimation cases was relatively lower, thereby encouraging reasonable underestimation of the prediction results and prompting patients to seek early intervention and get treated.

The rest of this paper was organized as follows: a summary of the related works on MCI conversion prediction and time series forecasting; the materials and methods employed, including a description of the datasets, data preprocessing methods, the architecture of the proposed framework, evaluation metrics, and the definition of the loss function; detailed training process and results of the experiments; and a conclusion of the main results of the study.

Related works

This section summarizes related works, focusing on the prediction of conversion from MCI to dementia and convolutional attention mechanism, a technique for time series forecasting.

Prediction of conversion from MCI to dementia

Current research on MCI has focused primarily on classification tasks. For example, Helaly et al.³ employed the principles of transfer learning to fine-tune a VGG19 pre-trained model to classify patients into NC, early MCI (EMCI), late MCI (LMCI), and AD categories. Similarly, Wen et al.¹⁷ introduce an innovative wavelet convolution unit designed for image-oriented neural networks, enhancing fine-grained classifications of AD. In this study, Li explored the application of machine learning algorithms for differentiating MCI from AD, achieving an SVM classification accuracy of 87.92%.¹⁸

In the realm of translational research, classification studies aimed at determining the likelihood of conversion are more common. Rye et al.⁵ used features such as cognitive function, hippocampal volume, and apolipoprotein E (APOE) genetic status to differentiate MCI converters (converted into AD) from stable MCI (keep stable) via machine learning algorithms. Hong et al.⁸ proposed an LSTM model to predict the stage of AD within six months based on biomarker data from MRI images. Although this involved forecasting future outcomes, at its core, it remains a classification task. Similarly, Wu et al.⁹ used a convolutional neural network (CNN) to predict whether the patients with MCI would convert into AD over three years at five different time points.

The classification of different stages of cognitive dysfunction from MCI to dementia, is to some extent well investigated. However, the regression issue of the exact timing of conversion remains under-researched. Thung et al.¹⁰ employed multimodal neuroimaging data and clinical information to predict the conversion rate and time for patients with MCI, asserting that their study is the first to simultaneously forecast the diagnosis of progressive MCI (pMCI) and its conversion time. El-Sappagh et al.¹¹ introduced a novel two-stage deep learning framework for the detection of AD progression. In the regression phase, multimodal time-series data were employed, and a Long Short-Term Memory (LSTM) model was used to forecast the precise conversion time for patients with pMCI. The authors declared that their study was the first to apply machine learning for regression prediction of the specific conversion time in MCI patients. Zhang et al.¹² developed interval-censored models to predict the conversion time from MCI to dementia. They used a semi-parametric Cox model and a random forest model, identifying sensitive measures through variable selection. The study utilized NACC and ADNI datasets for model building and validation, showing improved prediction accuracy for MCI-to-dementia conversion.

Convolutional attention mechanism

Attention mechanisms have garnered significant attention in research and have been widely integrated into deep neural networks to enhance their performance.¹³ However, their utilization in mobile networks, which typically have limited model size, have not kept pace with their application in larger networks.

Considering the restricted computation capacity of mobile networks, there are mainly three types of attention mechanisms proposed like the channel attention, the spatial attention and both. Being the focal point of attention, the Squeeze-and-Excitation (SE)¹⁹ method actively captures cross-dimensional interactions to extract channel-specific attention. By leveraging global average pooling and fully connected layers, it effectively calculates channel attention, resulting in significant performance enhancements at a remarkably low computational expense. Drawing from the principles of SE mechanism, Zhou et al.²⁰ have developed a Channel Attention Block with the express purpose of capturing the intricate relationships that exist between different channels. Nonetheless, the SE attention predominantly focuses on inter-channel data encoding, overlooking the significance of positional information crucial for delineating object structures in visual tasks.²¹

The Convolutional Block Attention Module (CBAM)²² introduces a method to capture cross-channel and cross-spatial information by considering the semantic inter-dependencies between spatial and channel dimensions within feature maps. It aims to incorporate positional information by reducing the channel dimension of the input tensor and subsequently computing spatial attention through convolutions. As a result, CBAM has demonstrated significant promise in integrating cross-dimensional attention weights into input features. However, convolutions are limited to capturing local relationships and struggle to model long-range dependencies crucial for vision-related tasks.²³ Subsequent advancements, like Attention Augmentation (AA)²⁴ and Convolutional Triplet Attention Module (CTAM),²⁵ expand on the concept of compressing individual 2D feature maps to effectively establish inter-dependencies among channels. This is achieved through the adoption of various spatial attention mechanisms or the creation of advanced attention blocks.

To effectively handle model complexity, utilizing convolution with channel dimensionality reduction is considered one of the most efficient approaches.²⁶ In contrast to SE attention, Coordinate Attention (CA)²⁷ integrates direction-specific details into channel attention along spatial dimensions and determines a suitable reduction ratio for channel dimensionality, thereby achieving performance levels comparable to SE attention. Nevertheless, the constrained receptive field of a 1 × 1 kernel convolution poses challenges in adequately capturing local cross-channel interactions and leveraging contextual information.

Utilizing the grouping structure, Efficient Multi-scale Attention (EMA²⁸ modifies the sequential processing approach of CA while maintaining dimensionality. The parallel subnetworks block is instrumental in capturing the interaction across dimensions and in forging inter-dimensional connections. For the aggregation of multi-scale spatial structure information, a 3 × 3 kernel is deployed alongside a 1 × 1 branch to facilitate rapid responses.

The attention mechanism has achieved excellent results in practical applications. Global Self-Attention Memristive Neural Network (GSA-MNN) proposed by Zhang et al.²⁹ employs Global Spatial Attention Module (GSAM) and Global Channel Attention Module (GCAM) to extract both global and local information from images, applicable flexibly to image restoration tasks of varying resolutions. Su et al.³⁰ proposed an end-to-end spatiotemporal attention network to detect auditory spatial attention from EEG. Soyak et al.³¹ overcame previous limitations by introducing an innovative neural network architecture, which integrates a channel-wise attention module with a fully convolutional network. Nie et al.³² introduced a temporal-spatial correlation attention network (TSCAN) aimed at tackling a variety of clinical characteristic prediction challenges.

Methods

In this section, we described the employed dataset, the data preprocessing techniques, the proposed prediction model of conversion time, the evaluation metrics, and the customized loss function for deep learning model training.

Dataset

The data came from a longitudinal retrospective study involving patients admitted to Dalian Central Municipal Hospital with dynamic cognitive function measurement and MRI examination of at least 3years apart between 1 January 2008 and 1 January 2022. Inclusion criteria: (1) age over 18 years old with a normal baseline cognitive function; (2) patients with dynamic brain image data and detailed clinical data over at least 3years before a diagnosis of MCI. Exclusion criteria: (1) MCI caused by a special unpreventable factor such as brain tumor, brain trauma, surgery, acute or chronic brain infection, poisoning, paraneoplastic syndrome, drug-related, etc.; (2) dementia caused by acute large vessel infarction or hemorrhage; (3) Subjective cognitive impairment.

Patients were categorized into two groups based on their final clinical diagnoses: normal cognitive function and MCI. A total of 463 patients were included in this study, comprising 232 patients with a final diagnosis of NC and 231 patients with a final diagnosis of MCI. The mean age at baseline was 61.77 ± 8.09 years, and the mean age at the last visit was 68.31 ± 7.83 years. There were 237 males (51.19%) and 264 individuals with higher education (57.02%). The median interval between baseline and the last visit was 6 years (interquartile range: 4–9 years). The demographic information of NC and MCI at baseline is shown in Table 1.

Table 1.

The demographics of NC and MCI at baseline.

Group	Age, mean ± SD		Sex, male (%)	Higher education (%)
Group	Baseline	Last Visit	Baseline	Baseline
Total (N = 463)	61.77 ± 8.09	68.31 ± 7.83	237 (51.19%)	264 (57.02%)
NC (N = 232)	60.56 ± 7.91	67.31 ± 7.94	117 (50.43%)	162 (69.83%)
MCI (N = 231)	62.99 ± 8.11	69.32 ± 7.59	120 (51.95%)	102 (44.16%)

Details about data and variables were described in the previous paper.³³ The characteristics corresponding to each patient included the following categories:

Clinical variables: including age, education, gender, medical history, and dynamic control of risk factors [including smoking index (number of years smoked * number of cigarettes smoked per day), alcohol consumption index (daily alcohol consumption/100 mL * years), diabetes mellitus years, mean glycated hemoglobin level (Normal, 0: 4–6%; ideal control, 1: 6–7%; under control, 2: 7–8%; poor control, 3: 8–9%; very poor control, 4: ≥9%; 5. unknown), years of hypertension, baseline mean systolic blood pressure (SBP), baseline percentage of SBP over 140 mmHg (%), baseline mean diastolic blood pressure (DBP), baseline percentage of DBP over 90 mmHg (%)];

Image variables based on MRI (Philips Achieva 3.0T magnetic resonance system, Philips Healthcare, United States, 5 mm thick slices): this group of features was classified into continuous features and discrete features.

(2.1) Continuous features were mainly used to quantitative analyze the degree of brain atrophy, defined as: the inner diameter of the forehead angle, the maximum width of the anterior longitudinal fissure, the width of lateral ventricular anterior horn, the index of the lateral ventricular anterior horn, the width of lateral ventricular posterior horn, index of lateral ventricular posterior angle, the width of the third ventricular, index of caudate nucleus, index of lateral ventricular body, mean width of the sulcus, distance of bilateral hippocampal uncinate gyrus, mean width of the hippocampus, mean distance from the temporal lobe to the anterior orbit, minimum width of the middle temporal lobe, mean width of the lateral fissure, maximum transverse width of the midbrain, maximum longitudinal diameter of the midbrain, maximum transverse width of the pons, and maximum longitudinal diameter of the pons. The units for the above features are all mm.

(2.2) Discrete features included the graded brain atrophy of the total brain, degree of total WMH [evaluated through Fazakas score (12)], and WMH regional distribution, which was reported to be independently associated with cognitive dysfunction (6, 13). Atrophy grade was defined as: 0, no atrophy; 1, mild cerebral atrophy, widening and deepening of cerebral sulcus and brain fissure; 2, moderate cerebral atrophy, decreased gyrus volume; 3, severe cerebral atrophy, blade-like gyrus. Paraventricular white matter hyperintensity Fazekas score was defined as 0, no abnormalities; 1, cap-shaped or thin pencil-like; 2, smooth halo; 3, irregular extension to deep white matter. Deep white matter hyperintensity Fazekas score was defined as 0, no abnormality; 1, punctate lesions; 2, the lesion tends to fuse; 3, large-scale fusion of lesions. Total Fazekas score was defined as 0–6 points, scoring the white matter in the paraventricular and deep parts separately, and then adding the scores of the two parts to calculate the total score.

The clinical and image variables were classified into 3 datasets. Data set 1 was defined as the clinical variables, including demographic characters, medical history, and previous control of risk factors. Data set 2 was defined as image variables, including the degree and distribution of brain atrophy and WMH. Data set 3 was defined as a combination of clinical and image variables.

Data preprocessing

Multi-source data integration

The dataset was derived from three Excel files: cross-sectional demographic and medical history variables, dynamic risk factor control variables, and MRI imaging variables. Demographic and medical history data were from the patients’ first visit, while dynamic risk factor data included multiple measurements during hospitalizations. Imaging data comprised numerical indicators of dynamic MRI from the initial examination onwards. We integrated these data sources by matching features based on medical record numbers and visit years to enhance the robustness of the deep learning model.

Handling of outliers and missing values

To ensure model reliability, outliers were removed from the dataset. Inconsistencies and missing values for patient visit year, age, and diagnosis year were manually corrected. Data distribution was confined within the interquartile range (IQR) to enhance data validity. The IQR is defined as:

\begin{matrix} IQR = Q_{3} - Q_{1} \end{matrix}

(1)

where

Q_{3}

and

Q_{1}

are the third and first quartiles, respectively. The upper and lower limits for outlier detection are calculated as:

\begin{matrix} W_{u} = Q_{3} + 1.5 \times IQR \end{matrix}

(2)

\begin{matrix} W_{l} = Q_{1} - 1.5 \times I Q R \end{matrix}

(3)

Missing values were handled by grouping data based on patient identification numbers. Within each group, missing values were filled using forward or backward filling methods to maintain data integrity.

Feature selection

Feature selection involves identifying the most informative and predictive subset of features from the original set to improve model performance and generalization. This process also enhances model training and inference speed. In this study, we employed a wrapper method using Recursive Feature Elimination (RFE) with a Random Forest (RF) estimator.

RFE is an iterative process that repeatedly builds models and eliminates the least important features until the optimal subset is identified. We used 5-fold cross-validation and the $R^{2}$ score to evaluate the models.

To validate the RFE results, we used the XGBoost model to calculate the relative importance of features. The selected features from RFE were compared with their importance scores from XGBoost to ensure consistency and further validate the feature selection process.

Data augmentation

To predict diagnosis time, we utilized feature data from each patient's examinations. Since the diagnosis time is influenced by both current and previous examination data, we grouped the dataset by patient ID and performed data clipping and padding at the group level.

For each patient, we clipped sequences from different starting points to generate multiple training samples. For a group with n records, this yielded $n (n + 1) / 2$ groups. Each new group included examination data from a specific year or consecutive years, allowing the model to predict the diagnosis time based on more historical records.

To ensure consistent sequence dimensions, we padded the sequences to a maximum length ( $maxlen$ ) using data from the earliest year. This process is illustrated in Figure 1.

Figure 1.

The process of data clipping and padding.

Records 1 to n are sorted in descending order by time. For group 1 to n, the records within the group start from record 1 and are clipped from different starting points, ending with records 1, 2, and n respectively; for groups n + 1 to 2n − 1, the records within the group start from record 2 and are clipped from different starting points, ending with records 2, 3, and n respectively; the remaining groups follow the same pattern. Finally, the last record within each group, which is the earliest record when sorted in descending order by time, is used to fill in the gaps.

Dual attention convolutional model

The feature matrix is composed of the measured M features at T time points. We first split the matrix into $M$ vectors, where $x_{t} = (x_{t}^{1}, x_{t}^{2}, \dots, x_{t}^{m}, \dots, x_{t}^{M})$ represents a vector of $M$ feature items at time point t, and $x^{m} = (x_{1}^{m}, x_{2}^{m}, \dots, x_{t}^{m}, \dots, x_{T}^{m})$ presents the $m$ -th feature series time of length $T$ in the history.

For the $t$ -th record of patient data $x_{t} = (x_{t}^{1}, x_{t}^{2}, \dots, x_{t}^{m}, \dots, x_{t}^{M}), 0 \leq t \leq T$ , the total number of detections T varies among different patients. We utilize a temporal attention module to identify the importance of each record as the number of detections varies, allowing for adaptive allocation of different weights to the records of different patients. The temporal attention module is inspired by the SE attention,¹⁹ constructed by a two-layer perceptron. The output of this module, which represents the weights for each record, is generated through a softmax function. For each record, the formula for calculating its temporal attention score is as follows:

\begin{matrix} α_{t} = s o f t m a x (W_{f} x_{t} + b_{f}) \end{matrix}

(4)

where

W_{f}

b_{f}

are weights and bias of the two-layer perceptron,

s o f t m a x

refers to the softmax activation function.

For the $m$ -th feature x of a patient $x^{m} = (x_{1}^{m}, x_{2}^{m}, \dots, x_{t}^{m}, \dots, x_{T}^{m})$ , different features have varying impacts on the prediction outcome. We employ a feature attention module to compute the relative importance of each feature in predicting the target variable. Feature attention module consists of global average pooling, multi-scale convolution layer, batch normalization and sigmoid activation function.

The global average pooling is frequently employed in channel attention mechanisms to encode spatial information globally and compress global spatial details into channel descriptors. Our block decomposes the data into the 1D feature encoding vectors to capture cross-channel dependencies along with spatial positional information. The 1D global average-pooling for encoding the global information along the vertical dimension direction at width m can be denoted by

\begin{matrix} Z_{m} = \frac{1}{T} \sum_{0 \leq t \leq T} x_{t}^{m} \end{matrix}

(5)

Through these encoding processes, our block excels at capturing long-range dependencies along the vertical dimension while retaining precise positional information along the horizontal dimension.

Utilizing various kernel sizes allows CNNs to gather multi-scale spatial information within the same processing stage. Inception³⁴ introduced a multi-branch structure to enrich the feature space, where each branch has varying local receptive fields. This aggregation approach empowers CNNs to collect and combine multi-scale information from different branches effectively. Our multi-scale convolution layer consists of 3 convolutional kernel sizes: 1 × 1 kernel, 1 × 3 kernel, 1 × 7 kernel. Given the aggregated feature map Z produced by Equation (5), we send it to multi-scale convolutional transformation function, batch normalization and sigmoid activation function, yielding

\begin{matrix} f = δ (F_{1} (Z) + F_{3} (Z) + F_{7} (Z)) \end{matrix}

(6)

where

F_{1}, F_{3}, F_{7}

denote 1 × 1 kernel, 1 × 3 kernel, 1 × 7 kernel convolutional transformation function respectively,

δ

is a sigmoid activation function and

f \in R^{(1 \times M)}

is the feature weight vector. For our high-dimensional data, large kernel size effectively establishes short and long-range dependencies for better performance.

According to time attention weights $α \in R^{(T \times 1)}$ and feature attention weights $f \in R^{(1 \times M)}$ produced by time attention block and feature attention block, the original data can be reweighted and the output of our attention block X can be calculated as

\begin{matrix} X = \sum_{0 \leq t \leq T} x_{t} \times f \times α_{t} \end{matrix}

(7)

Finally, X is fed into a two-layer one-dimensional convolutional network, and the final prediction is obtained through a fully connected layer. The forecasting framework of the dual attention convolutional model is presented in Figure 2.

Figure 2.

The framework of the dual attention convolutional model.

Quantile mean square and relative error loss function

The prediction results indicated the exact conversion time of a potential MCI patient with NC to a diagnosis with MCI. To enhance the acceptability of the prediction results among clinicians, we had referred to MSE loss and MRE loss. Firstly, for time prediction, we preferred the model to produce underestimated results. That was, if a potential patient would be diagnosed with MCI in 5 years, a model prediction of 4.9 years would be preferred over 5.1 years. Given the same absolute error, the underestimation would guide the patient to start preventive treatment earlier. By introducing a coefficient ( $γ$ ), we aimed to reduce the loss value in cases of underestimation. When ( $γ < 0.5$ ), the loss associated with overestimation was greater than that of underestimation, thus encouraging the model to favor underestimation.

Additionally, most loss functions scarcely consider the relative error of predictions. However, for predicting the conversion time to MCI, we also required a relatively small relative error. Therefore, we incorporated a relative error penalty term into the loss function. For non-zero true values, we calculated their relative error. The final formula for our Quantile Mean Square and Relative Error (QMSRE) Loss function is as follows:

\begin{matrix} L o s s = \frac{1}{N} \sum_{i = 1}^{N} I_{{\hat{y}}_{i} \geq y_{i}} (1 - γ) M S R E + I_{{\hat{y}}_{i} \leq y_{i}} γ M S R E \end{matrix}

(8)

in which

\begin{matrix} M S R E = 0.8 \times {(y_{i} - {\hat{y}}_{i})}^{2} + 1.1 \times \frac{| y_{i} - {\hat{y}}_{i} |}{| y_{i} |}, y_{i} \neq 0 \end{matrix}

(9)

where

y_{i}

denotes true value, and

{\hat{y}}_{i}

denotes predicted value

Evaluation metrics

The MSE, mean absolute percent error (MAPE), MAE and $R^{2}$ are applied to evaluate the prediction performance of the presented model. They are defined by:

\begin{matrix} M S E = \frac{1}{n} \sum_{i = 1}^{n} {({\hat{y}}_{i} - y_{i})}^{2} \end{matrix}

(10)

\begin{matrix} M A P E = \frac{100 %}{n} \sum_{i = 1}^{n} | \frac{{\hat{y}}_{i} - y_{i}}{y_{i}} | \end{matrix}

(11)

\begin{matrix} M A E = \frac{1}{n} \sum_{i = 1}^{n} | y_{i} - {\hat{y}}_{i} | \end{matrix}

(12)

\begin{matrix} R^{2} = 1 - \frac{\sum_{i = 1}^{n} {(y_{i} - {\hat{y}}_{i})}^{2}}{\sum_{i = 1}^{n} {({\bar{y}}_{i} - {\hat{y}}_{i})}^{2}} \end{matrix}

(13)

where n represents the number of predicted points;

y_{i}

and

{\hat{y}}_{i}

represent the real value and the predictive value of the

i

-th point, respectively.

{\bar{y}}_{i}

is the mean value of

y_{i}

Results

The conversion time from NC to MCI was defined as the number of years from the first visit with NC to the year of MCI diagnosis, so that we could predict it based on the patient's past and present data.

We attempted to predict it using Random Forest, XGBoost, Linear Regressor, AdaBoost, LightGBM, one-dimensional convolution model, GRU, GRU with attention model, and our dual attention convolutional model. This paper conducted three experiments based on different data and models, discussing the results according to the evaluation metrics of MSE, MAPE, MAE, and $R^{2}$ , and presenting them in tables.

Experiment based on machine learning models under dataset D1

In the first experiment, we used only dynamic MRI and demographic data, named as dataset D1, to make predictions based on machine learning models. We removed outliers beyond the quartiles from the data, imputed missing values, performed feature selection based on RFE, and validated the feature importance.

Figure 3 presented the distribution histograms and kernel density estimation curve of key features and labels, including the specific conversion time from NC to MCI, age, and educational level.

Figure 3.

The distribution histograms and kernel density estimation curves of key features and label. (a) MCI year. (b) education. (c) age (year).

Figure 3(a) depicts the distribution of the specific conversion time from NC to MCI. The histogram bars represent the frequency of occurrence across different time periods, while the red line denotes the kernel density estimate (KDE), providing a smoothed representation of the data distribution. A distinct peak near zero indicates that much of the data is concentrated at the baseline when diagnosed with MCI. The KDE curve further highlights this central tendency, revealing a slightly skewed distribution towards positive values, suggesting that although most conversions happen quickly, there are instances where the conversion takes longer.

Figure 3(b) examines the distribution of educational levels among the patients. The histogram reveals a multimodal distribution with peaks at certain levels, possibly corresponding to common educational milestones such as high school graduation or college completion. The KDE curve smoothes these discrete categories into a continuous probability density function, illustrating the relative frequencies of different educational backgrounds within the cohort. This information is valuable for assessing the impact of education on cognitive health, as higher educational attainment has been linked to better cognitive resilience against neurodegenerative diseases.^35,36

Figure 3(c) focuses on the age distribution of the patients. The histogram shows a unimodal distribution with a peak around the middle-aged category, indicating that the majority of patients fall within this age range. The KDE curve reinforces this observation, displaying a bell-shaped curve centered around the same region. This age distribution is crucial for understanding the typical age at which cognitive transitions occur, potentially reflecting the natural progression of cognitive decline associated with aging.

The distributions of the specific conversion time from NC to MCI, age, and educational level provide essential insights into the demographic profile of the study population and the variability in key variables related to cognitive health transitions. Understanding these distributions is fundamental for interpreting subsequent analyses and modeling efforts aimed at predicting the onset of MCI. Moreover, the observed patterns suggest potential areas for further investigation, such as examining the interplay between age and educational level in modulating the risk of cognitive decline.

Subsequently, the data were grouped by education level, and age was divided into n groups. The conversion time for each age interval was distributed across the IQR from low to high. The IQR was obtained and the upper and lower limit values were calculated.

As illustrated in Figure 4, the boxplots offer a vivid depiction of the conversion time distributions across different educational levels and age brackets. The boxes encapsulate the IQR, with the median marked by a horizontal line within each box, providing a clear indicator of the central tendency. Whiskers extend to capture the extent of the non-outlier data, while individual outlier points are explicitly highlighted, offering a comprehensive view of the data dispersion.

Figure 4.

Box diagram of quartile.

This systematic approach not only underscores the heterogeneity in conversion times but also invites speculation on the underlying mechanisms that might drive these variations. By juxtaposing educational attainment with chronological age, the analysis hints at complex interactions between sociocultural influences and biological aging processes, underscoring the multifaceted nature of cognitive decline.

Figure 5 illustrates the cross-validation score variation as a function of the number of selected features in dataset D1. The evaluation metric used is R², which measures the proportion of variance explained by the model. The goal is to determine an optimal number of features that balances model performance and computational efficiency.

Figure 5.

Cross-validation score variation along with numbers of selected features in dataset D1.

To assess the impact of feature selection on model performance, we varied the number of features from 1 to 56, incrementing by steps of 2. We employed 5-fold cross-validation to ensure robustness and generalizability of our results. The cross-validation scores for each configuration of selected features are plotted in the figure.

Initially, the scores rise sharply, indicating significant improvements in model fit as more relevant features are included. However, beyond a certain point, the gains become marginal, and the scores plateau. This behavior suggests diminishing returns in terms of predictive power with additional features.

Considering both model performance and computational time, the optimal number of features for the current task was determined to be 13. At this threshold, the model achieves a satisfactory balance between accuracy and efficiency. Adding more features beyond this point does not significantly enhance the R² score, yet it would increase computational complexity and potentially introduce overfitting risks.

The names of these 13 features are as follows: ‘age’, ‘CT pontine maximum transverse diameter’, ‘Anterior angle left width of the ventricle’, ‘The widest distance of the forehead angle’, ‘Atrophy grade’, ‘Ventricular index’, ‘Lateral ventricle G/L, defined as the maximum distance of anterior horn of the bilateral ventricle/ maximum diameter of the cranial cavity at the same level’, ‘CT pontine maximum length diameter’, ‘Maximum outer diameter of the head’, ‘CT midbrain maximum length diameter’, ‘Frontal angle index’, ‘Right width of the occipital angle’, ‘CVSD total load score’.

By identifying an optimal feature count of 13, we can construct a more efficient and effective predictive model without sacrificing predictive power. Future work could explore alternative feature selection techniques or investigate the impact of different evaluation metrics on the optimal feature set.

To validate the results of the RFE process, we computed the relative importance of each feature using the XGBoost algorithm and sorted them in descending order. The scores of relative feature importance are depicted in the Figure 6, where the features are ranked based on their contribution to the model's predictive power.

Figure 6.

Cross-validation score variation along with numbers of selected features in dataset D1.

The alignment of the feature rankings derived from XGBoost with those obtained from the RFE process confirms the reliability of our feature selection strategy. This consistency indicates that the most influential features identified by RFE are indeed crucial for the model's performance, as corroborated by their high importance scores in the XGBoost analysis.

Furthermore, the gradual decrease in importance scores as we move down the list suggests a natural hierarchy among the features, with some exhibiting significantly higher relevance than others. This observation supports the notion that a subset of key features can effectively capture the essential information needed for accurate predictions, thus justifying the use of feature selection techniques like RFE.

In conclusion, the feature importance analysis conducted using XGBoost not only validates the outcomes of the RFE process but also provides valuable insights into the relative contributions of individual features towards the overall model performance. This dual validation approach enhances our confidence in the selected feature set and lays a solid foundation for subsequent modeling efforts.

Subsequently, predictions were made based on each machine learning model, and evaluation metrics such as MSE, MAE, and $R^{2}$ were calculated. The results were presented in Table 2.

Table 2.

The performance of machine learning models under dataset D1.

Model	MSE	MAE	$R^{2}$
AdaBoost	6.8170	1.9412	0.4194
LightGBM	7.1378	2.1167	0.4348
SVM (‘rbf’)	8.4792	2.2905	0.2698
Random Forest	8.0361	2.2181	0.3193
XGBoost	6.9414	2.0134	0.4488

The MAE of each model including AdaBoost, LightGBM, SVM (kernel = 'rbf’), Random Forest, and XGBoost were 1.9412, 2.1167, 2.2905, 2.2181, and 2.0134, respectively. The AdaBoost model achieved the best result, of 1.9412 for MAE. Clearly, none of the machine learning models achieved satisfactory results, which could be attributed to the insufficiency of data features or quantity. Consequently, we decided to merge multi-source data, incorporating additional relevant features.

Experiment based on machine learning models under dataset D2

In the second experiment, we integrated additional relevant medical history data, such as history of hypertension,³⁷ smoking,³⁸ diabetes,³⁹ alcohol,⁴⁰ etc., into our dataset, referred to as dataset D2. We then imputed missing values, performed feature selection based on RFE, and validated the feature selection using feature importance.

We set the number of features for selection from 1 to 65 with a step size of 2, and the number of cross-validation folds to 5. The cross-validation scores as a function of the number of features selected are shown in the Figure 7. The plot shows a clear trend in the cross-validation scores as the number of features increases. Initially, the scores rise sharply, indicating significant improvements in model fit as more relevant features are included. However, beyond a certain point, the gains become marginal, and the scores plateau. This behavior suggests diminishing returns in terms of predictive power with additional features.

Figure 7.

Cross-validation score variation along with numbers of selected features in dataset D2.

Considering both model accuracy and computational time, the optimal number of features for the current task was determined to be 15. The names of these 15 features were as follows: ‘age’, ‘atrophy grade’, ‘frontal angle index’, ‘ventricular index’, ‘CT pontine maximum transverse diameter’, ‘the widest distance of the forehead angle ‘, ‘caudate nuclear index’, ‘right width of the occipital angle’, ‘maximum outer diameter of the head’, ‘anterior angle left width of the ventricle ‘, ‘hippocampal uncinate distance index’, ‘CT pontine maximum length diameter’, ‘CT midbrain maximum length diameter’, ‘CT midbrain maximum transverse’, ‘right hippocampus width, defined as the maximum width of the right hippocampus’.

The inclusion of additional medical history data in dataset D2 led to a slight increase in the number of selected features, which suggests that these new variables contribute valuable information to the model. The optimal number of features for dataset D2 (15) is higher than that for dataset D1 (13), indicating that the additional data improves the model's performance without introducing unnecessary complexity. This finding is consistent with the principle of parsimony, where the model should be as simple as possible while still capturing the essential patterns in the data.

Despite the differences in the number of selected features, there is a significant overlap between the features chosen from datasets D1 and D2. Both datasets include ‘age’, ‘atrophy grade’, ‘frontal angle index’, ‘ventricular index’, ‘CT pontine maximum transverse diameter’, ‘the widest distance of the forehead angle’, ‘right width of the occipital angle’, ‘maximum outer diameter of the head’, ‘CT pontine maximum length diameter’, and ‘CT midbrain maximum length diameter’. This consistency highlights the robustness of these features in predicting the conversion from NC to MCI. The repeated selection of these features across different datasets underscores their importance and reliability in the context of cognitive decline prediction.

To validate the results of the RFE process, we computed the relative importance of each feature using XGBoost and sorted them in descending order. The scores of relative feature importance were shown in the Figure 8, and their ranking aligns with the findings from RFE, thereby confirming our feature selection.

Figure 8.

The importance of top features in dataset D2.

The specific features highlighted in Figure 8 provide valuable insights into the underlying factors driving the outcome of interest. For instance, if age consistently ranks among the highest, it suggests that chronological aging plays a pivotal role in the phenomenon under study. Similarly, other highly ranked features might correspond to critical biological markers, environmental exposures, or lifestyle choices that warrant further investigation.

Subsequently, predictions were made based on each machine learning model, and evaluation metrics such as MSE, MAE, and $R^{2}$ were calculated. The results were presented in Table 3.

Table 3.

The performance of machine learning models under dataset D2.

Model	MSE	MAE	$R^{2}$
AdaBoost	2.0417	0.8901	0.8430
LightGBM	2.1153	1.0087	0.8373
SVM (‘rbf’)	3.7160	1.2689	0.6909
Random Forest	2.0402	0.9873	0.8431
XGBoost	2.0574	0.9354	0.8418

The MAE of each model including AdaBoost, LightGBM, SVM (kernel = 'rbf’), Random Forest, and XGBoost were 0.8901, 1.0087, 1.2689, 0.9873, and 0.9354, respectively. The AdaBoost model achieved the best result, of 0.8901 for MAE. Compared to the experiment on dataset I, the performance of all machine learning models had been significantly improved, indicating that medical history data was highly beneficial for time prediction. However, our approach only utilized data from the current examination. Incorporating historical data from previous examinations could potentially enhance prediction accuracy. Consequently, we considered the application of deep learning models for prediction.

Experiment based on deep learning models under dataset D2

In the third experiment, we considered the learning effectiveness of deep learning models on the data. Initially, we performed regression prediction based on one-dimensional convolution model using the current examination data and calculated the evaluation metrics MSE, MAE, and $R^{2}$ . Subsequently, considering the temporal relationships among multiple records for each patient, we decided to exploit this aspect using RNN. However, the number of records for each patient varied and was relatively small, typically only a few entries. To make the data suitable for RNNs, we grouped the data by patient ID and applied the data augmentation mentioned in Section 3.2.4, ensuring consistent data dimensions, and increasing the amount of trainable data. Based on the processed data, we performed regression prediction using both GRU and GRU with attention models, and calculated the evaluation metrics MSE, MAE, and $R^{2}$ . The RNN did not yield a significant improvement. Considering the high dimensional features and the limited number of records per patient, for such short-sequence, high-dimensional time-series data, employing convolutional networks might yield better results. Thus, we conceptualized a dual attention convolutional network for prediction. The prediction performance results of the deep learning models based on dataset D2 were presented in Table 4.

Table 4.

The performance of deep learning models under dataset D2.

Model	MSE	MAE	$R^{2}$	MAPE
Conv1d	0.7186	0.6394	0.8109	90.70%
GRU	0.4224	0.5129	0.8006	20.86%
GRU + attention	0.3313	0.5229	0.8254	18.33%
Conv1d + feature weight	0.2391	0.4130	0.8317	35.50%
Conv1d + time weight	0.1963	0.3117	0.9155	20.26%
Conv1d + feature weight + time weight	0.0695	0.1678	0.9464	11.67%
Conv1d + feature weight + time weight + loss	0.0682	0.1202	0.9603	08.44%

From the perspective of the MSE, MAE, and $R^{2}$ metrics, the performance of the one-dimensional convolution model, GRU, and GRU with attention model was relatively similar. However, RNN increased the consumption of training resources. The proposed dual attention convolutional network achieved the best predictive performance, with MSE, MAE, and $R^{2}$ metrics of 0.0695, 0.1678, and 0.9464, respectively. Compared to the one-dimensional convolution model, the MSE and MAE were reduced to 9.67% and 26.24%, while the $R^{2}$ score increased by 16.71%. The performance of the one-dimensional convolution model was also enhanced by the individual attention modules: The model using only feature attention reduced the MSE and MAE to 33.27% and 64.59%, respectively, while the $R^{2}$ score increased by 2.57%. Similarly, the model using only temporal attention reduced the MSE and MAE to 27.32% and 48.75%, respectively, with the $R^{2}$ score improving by 12.90%. From the above comparisons, it was evident that the dual-attention module provided greater performance improvements than the single-attention module. This demonstrated that both individual attention modules and their combined effect contributed significantly to enhancing performance.

Finally, by replacing the MSE loss function with our custom QMSRE loss function during the training of the dual attention convolutional network, and by removing the data with label 0 from the training process, the performance metrics MSE and MAE were lightly reduced to 98.13% and 71.63%, respectively, while the $R^{2}$ score improved by 1.47%. These three metrics saw slight enhancements. However, in terms of the MAPE evaluation metric, there was a significant improvement of 27.68%, demonstrating that the custom QMSRE loss function plays a substantial role in reducing relative error, thereby enhancing the clinical guidance value of the model's predictions.

As mentioned, only three studies^10–12 have predicted the specific conversion time from MCI to AD, while no studies have predicted the specific conversion time to MCI in the NC population. Firstly, Thung et al.¹⁰ achieved a maximal time prediction correlation of 0.665 and a maximal $R^{2}$ score of 0.369. Secondly, El-Sappagh et al.¹¹ predicted the specific conversion time with MAE of 0.1452, MSE of 0.0568 and RMSE of 0.2382. Finally, Zhang et al.¹² achieved a IBS of 0.099, $P_{w i t h i n}$ (MCI-to-dementia) of 12% and $P_{b e f o r e}$ of 45%. In comparable scenarios, the result of this study outperforms research¹⁰ and research.¹¹

Discussion

In this study, to evaluate the effectiveness and stability of the dynamic weights in the feature attention module of the model, we calculated the dynamic weight values generated by the model for each feature after each sample input, based on 100 sample data points from the test set. In this way, we obtained 100 weight observations for each feature, thereby constructing the weight distribution for each feature. Subsequently, for the 100 dynamic weight values of each feature, we calculated their mean and variance respectively to quantify the central tendency and fluctuation degree of the feature weights. To further assess the reliability of the feature weight estimation, we assumed that the sampling distribution of the sample mean followed a t-distribution and accordingly calculated the confidence interval for each feature weight at a 95% confidence level. Specifically, we obtained the error range (i.e., the half-width of the confidence interval) by multiplying the quantile of the t-distribution with the standard error, thereby constructing the upper and lower confidence bounds for each feature weight. To identify the most representative and stable features, we adopted a ranking strategy based on the lower limit of the confidence interval, that is, we prioritized features with higher lower limits of the confidence interval, as this indicates more consistent performance across different samples and more reliable weight estimation. Finally, we selected the top ten features with the highest lower limits of the confidence interval as the optimal feature set. To more intuitively present the average weights of these features and their uncertainties, we drew an error bar chart, in which each bar represents the average weight of a feature, and the error bar on it indicates the corresponding 95% confidence interval range, thereby making the weight differences and stability among features clearly visible Figure 9.

Figure 9.

The importance of top 10 features with error bars.

The top 5 features identified by our model—namely total Fazekas score, atrophy grade, temporal lobe atrophy, age, and education. For instance, regional white matter hyperintensity burden has been reported to be strongly correlated with amyloid deposition and cognitive impairment.⁴¹ Regional brain atrophy is a well-established imaging marker for cognitive decline and dementia.⁴² This atrophy has been associated with reduced microstructural integrity in white matter pathways, including the perforant path, where white matter hyperintensities correlate with the severity of entorhinal cortex atrophy.⁴³

Conclusion

This study aimed to predict the conversion time of MCI in potential patients with NC by integrating clinical, imaging parameters, and significant etiological factors. Through numerous experiments on various datasets and models, we had identified features beneficial for conversion prediction. We proposed a novel dual attention convolutional network and the QMSRE loss function, which had significantly enhanced the accuracy of predicting conversion time and increased the clinical guidance value compared to existing research.^10,11

We developed a new feature attention module and a temporal attention module, which found out the correlations between different features and prediction outcomes in scenarios with a high number of features, and discovered the temporal dependencies among different records. The light multi-scale convolutional attention module can adaptively enhance the influence of highly relevant features on the prediction points and reduce the impact of irrelevant variables. The proposed method does not require variable selection and feature engineering, providing guidance for variable selection and aiding in solving the problem of identifying the most important features, making it easier for engineers to use.

The study has two primary limitations. Firstly, the unit of specific conversion time predicted in this study is in years, which is still not precise enough for actual clinical diagnostic guidance. To enhance the accuracy of the model's predictions, additional diagnostic data with labels that are more precise, such as to the month or day, could be incorporated. Secondly, this study shares the common limitation of small deep learning models, which lack the generalizability of pre-trained models. In the future, we plan to supplement the study with more time-series data related to cognitive impairments and undertake pre-training of larger models to enhance the generalizability of our research findings.

Footnotes

Acknowledgements

We would like to thank all the participants for their contributions to this study. We also acknowledge the support from the medical staff at Dalian Central Municipal Hospital for their assistance in data collection.

ORCID iDs

Sibo Liu

Chao Huang

Ethical considerations

The studies involving humans were approved by the Ethics Committee of the Dalian Central Municipal Hospital (Approval Code: 2023–011-01; Approval Date: 2023.2.16). The studies were conducted in accordance with the local legislation and institutional requirements.

Consent to participate

Written informed consent for participation was not required from the participants or the participants’ legal guardians/next of kin since the data came from a longitudinal retrospective study. The requirement for written informed consent was waived by the Ethics Committee of the Dalian Central Municipal Hospital due to the retrospective nature of the study.

Consent for publication

This study utilized anonymized medical testing data from multiple patients. All data were de-identified to ensure patient confidentiality. Informed consent for publication was not required as the data do not contain any individual person's details, images, or videos that could identify participants. The study was approved by the Ethics Committee of the Dalian Central Municipal Hospital (Approval Code: 2023–011-01; Approval Date: 2023.2.16), which waived the requirement for individual consent for publication due to the use of anonymized retrospective data.

Author contribution(s)

Xiawei Zhu: Methodology; Software; Writing – original draft.

Sibo Liu: Visualization; Writing – review & editing.

Long Wang: Formal analysis; Writing – review & editing.

Jinjie Liu: Visualization; Writing – review & editing.

Dean Gu: Methodology; Supervision; Writing – review & editing.

Chao Huang: Conceptualization; Methodology; Supervision; Writing – review & editing.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported in part by the Dalian Medical Key Specialty “Climbing the Peak Plan” under Grants 2022ZZ267 and 2023ZZ012, in part by the Dalian High-Level Talents Innovation Support Plan under Grants 2021RQ029 and 2023RY019, in part by Fundamental Research Funds for the Central Universities under Grant 06500103, in part by the Liaoning Province Science and Technology Joint Project under Grant 2024-MSLH-072, in part by the Guangdong Basic and Applied Basic Research Foundation under Grant 2022A1515240044, and in part by Beijing Natural Science Foundation under Grant 4232040.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

The raw data are protected and are not available due to data privacy laws.

References

World Health Organization . Global action plan on the public health response to dementia 2017–2025. Geneva: World Health Organization, 2017.

Alzheimer's Association . Alzheimer’s disease facts and figures. Alzheimers Dement 2023; 19: 1598–1695.

Helaly

Badawy

Haikal

. Deep learning approach for early detection of Alzheimer’s disease. Cognit Comput 2022; 14: 1711–1727.

Siuly

Alçin

ÖF

Wang

, et al. Exploring rhythms and channels-based EEG biomarkers for early detection of Alzheimer's disease. IEEE Trans Emerg Top Comput Intell 2024; 8: 1609–1623.

Rye

Vik

Kocinski

, et al. Predicting conversion to Alzheimer’s disease in individuals with mild cognitive impairment using clinically transferable features. Sci Rep 2022; 12: 15566.

Mao

Rasmussen

, et al. AD-BERT: using pre-trained language model to predict the progression from mild cognitive impairment to Alzheimer's disease. J Biomed Inform 2023; 144: 104442.

Liu

Dong

Sun

, et al. Risk prediction of Alzheimer’s disease conversion in mild cognitive impaired population based on brain age estimation. IEEE Trans Neural Syst Rehabil Eng 2023; 31: 2468–2476.

Hong

Lin

Yang

, et al. Predicting Alzheimer’s disease using LSTM. IEEE Access 2019; 7: 80893–80901.

Guo

Hong

, et al. Discrimination and conversion prediction of mild cognitive impairment using convolutional neural networks. Quant Imaging Med Surg 2018; 8: 992.

10.

Thung

Yap

Adeli

, et al. Conversion and time-to-conversion predictions of mild cognitive impairment using low-rank affinity pursuit denoising and matrix completion. Med Image Anal 2018; 45: 68–82.

11.

El-Sappagh

Saleh

Ali

, et al. Two-stage deep learning model for Alzheimer’s disease detection and prediction of the mild cognitive impairment time. Neural Comput Appl 2022; 34: 14487–14509.

12.

Zhang

Song

, et al. Predicting conversion time from mild cognitive impairment to dementia with interval-censored models. J Alzheimers Dis 2024; 101: 147–157.

13.

Kishore

Goel

. A review of machine learning techniques for diagnosing Alzheimer’s disease using imaging modalities. Neural Comput Appl 2024; 36: 21957–21984.

14.

Yang

Nasrallah

Shou

, et al. A deep learning framework identifies dimensional representations of Alzheimer’s disease from brain structure. Nat Commun 2021; 12: 7065.

15.

Shen

, et al. White matter hyperintensities and risks of cognitive impairment and dementia: a systematic review and meta-analysis of 36 prospective studies. Neurosci Biobehav Rev 2021; 120: 16–27.

16.

Bravo-Ortiz

Holguin-Garcia

Quiñones-Arredondo

, et al. A systematic review of vision transformers and convolutional neural networks for Alzheimer’s disease classification using 3d mri images. Neural Comput Appl 2024; 36: 21985–22012.

17.

Wen

Fang

, et al. Fine-grained and multiple classification for Alzheimer's disease with wavelet convolution unit network. IEEE Trans Biomed Eng 2023; 70: 2592–2603.

18.

. Identification of mild cognitive impairment by machine learning algorithm. Neural Comput Appl 2023; 35: 25121–25130.

19.

Shen

Sun

Squeeze-and-excitation networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), 2018, pp.7132–7141.

20.

Zhou

Lei

, et al. Rfnet: reverse fusion network with attention mechanism for RGB-D indoor scene understanding. IEEE Trans Emerg Top Comput Intell 2022; 7: 598–603.

21.

Wang

Zhu

Green

, et al. Axial-deeplab: stand-alone axial-attention for panoptic segmentation. In: European conference on computer vision (ECCV). Cham: Springer, 2020, pp.108–126.

22.

Woo

Park

Lee

, et al. CBAM: convolutional block attention module. In: Proceedings of the European conference on computer vision (ECCV), 2018, pp.3–19.

23.

Hou

Zhang

Cheng

, et al. Strip pooling: rethinking spatial pooling for scene parsing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2020, pp.4003–4012.

24.

Bello

Zoph

Vaswani

, et al. Attention augmented convolutional networks. In: Proceedings of the IEEE/CVF international conference on computer vision (ICCV), 2019, pp.3286–3295.

25.

Misra

Nalamada

Arasanipalai

, et al. Rotate to attend: convolutional triplet attention module. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision (WACV), 2021, pp.3139–3148.

26.

Zhang

Zheng

, et al. Shufflenet v2: practical guidelines for efficient CNN architecture design. In: Proceedings of the European conference on computer vision (ECCV), 2018, pp.116–131.

27.

Hou

Zhou

Feng

Coordinate attention for efficient mobile network design. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2021, pp.13713–13722.

28.

Ouyang

Zhang

, et al. Efficient multi-scale attention module with cross-spatial learning. In: ICASSP 2023–2023 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2023, pp.1–5.

29.

Zhang

Xiao

Xie

, et al. A global self-attention memristive neural network for image restoration. IEEE Trans Emerg Top Comput Intell 2024; 8: 2613–2624.

30.

Cai

Xie

, et al. Stanet: a spatiotemporal attention network for decoding auditory spatial attention from eeg. IEEE Trans Biomed Eng 2022; 69: 2233–2242.

31.

Soyak

Navruz

Ersoy

, et al. Channel attention networks for robust mr fingerprint matching. IEEE Trans Biomed Eng 2021; 69: 1398–1405.

32.

Nie

Zhang

, et al. Temporal-spatial correlation attention network for clinical data analysis in intensive care unit. IEEE Trans Biomed Eng 2023; 71: 583–595.

33.

Zhu

Liu

, et al. Machine learning-based prediction of mild cognitive impairment among individuals with normal cognitive function. Front Neurol 2024; 15: 1352423.

34.

Szegedy

Vanhoucke

Ioffe

, et al. Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), 2016, pp.2818–2826.

35.

Zhong

Liu

, et al. The impact of education and occupation on cognitive impairment: a cross-sectional study in China. Front Aging Neurosci 2024; 16: 1435626.

36.

Lövdén

Fratiglioni

Glymour

, et al. Education and cognitive functioning across the life span. Psychol Sci Public Interest 2020; 21: 6–41.

37.

Qin

, et al. Prevalence of mild cognitive impairment in patients with hypertension: a systematic review and meta-analysis. Hypertens Res 2021; 44: 1251–1260.

38.

Crawford Holland

Rabin

Fornazzari

, et al. Smoking gun? Effect of smoking history on cognition in ad and mci. Alzheimers Dement 2022; 18: e68722.

39.

Ding

Yin

Zhang

, et al. Diabetes accelerates Alzheimer's disease progression in the first year post mild cognitive impairment diagnosis. Alzheimers Dement 2024; 20: 4583–4593.

40.

Cui

Zhu

, et al. Alcohol consumption and mild cognitive impairment: a Mendelian randomization study from rural China. Nutrients 2022; 14: 3596.

41.

Uchida

Nishimaki

Soldan

, et al. Acceleration of brain atrophy and progression from normal cognition to mild cognitive impairment. JAMA Netw Open 2024; 7: e2441505.

42.

Uchida

Hou

Gomez Isaza

, et al. Quantification of perforant path fibers for early detection of Alzheimer's disease. Alzheimers Dement 2025; 21: e70142.

43.

Uchida

Onda

Hou

, et al. Microstructural neurodegeneration of the entorhinal-hippocampus pathway along the Alzheimer’s disease continuum. J Alzheimers Dis 2023; 95: 1107–1117.