1 Introduction
The global prevalence of Alzheimer’s disease (AD) and related dementias is projected to triple by 2050, reaching an estimated 153 million cases worldwide [
1]. This looming public health crisis creates an urgent imperative for scalable, non-invasive diagnostic tools that can detect pathology in its earliest stages. While the preclinical detection of mild cognitive impairment (MCI) offers a critical window for therapeutic intervention, current diagnostic workflows remain resource-intensive and inaccessible to the wider population. Gold-standard biomarkers, such as amyloid-PET imaging and cerebrospinal fluid (CSF) analysis, are invasive and costly, restricting their use to specialized clinical settings. Conversely, neuropsychological assessments like the Everyday Cognition (ECog) scale, specifically the 12-item short form (ECog-12) provide valid measures of functional decline [
2]. However, these subjective reports can be confounded by anosognosia in patients or recall bias in informants. Consequently, there is a critical need to validate objective physiological biomarkers that can corroborate subjective clinical scores with high precision.
Oculomotor behavior offers a unique window into the integrity of distributed neural networks essential for cognition [
3,
4]. The generation of saccades relies on a complex circuitry involving the dorsolateral prefrontal cortex (DLPFC), the frontal eye fields (FEF), and the anterior cingulate cortex, regions that overlap extensively with networks compromised in early-stage neurodegeneration [
5,
6]. Specifically, antisaccade tasks, which require the suppression of a reflexive glance toward a stimulus in favor of a volitional gaze in the opposite direction, tax inhibitory control and executive function. Previous studies have consistently demonstrated that patients with MCI and early AD exhibit distinct oculomotor signatures, including prolonged prosaccade latencies and increased error rates in antisaccade protocols [
7–
10]. Despite this established physiological link, the translation of these subtle motor metrics into granular clinical scores remains a computational challenge due to the non-linear trade-offs between speed (latency) and accuracy (error rate).
Traditional statistical approaches, such as linear regression or support vector machines, often fail to capture the non-linear, high-dimensional dependencies inherent in physiological data [
11,
12]. Furthermore, conventional classifiers such as Random Forest or XGBoost, while effective in many tabular settings, do not explicitly represent the global interdependencies that characterize individual oculomotor-cognitive profiles. Deep learning (DL) has emerged as a superior alternative for modeling such complexities [
13,
14]. While convolutional neural networks (CNNs) have excelled in extracting spatial features from raw signals, recent advancements in attention-based architectures, such as Transformers, have revolutionized the processing of sequential and tabular data [
15]. However, relying on either architecture in isolation has limitations. While Transformer architectures are commonly associated with high-dimensional sequential data, their utility in low-dimensional tabular settings with complex feature interactions has been increasingly demonstrated [
16]. In the present context, the five oculomotor metrics exhibit physiologically non-additive interactions. For example, the relationship between antisaccade error rate and latency is known to be non-monotonic and subject-dependent [
17]. The Transformer encoder’s self-attention mechanism captures these global, context-dependent feature dependencies, which CNN layers alone (operating on local adjacent feature pairs) cannot model across the full feature set. A more robust approach involves a hybrid strategy that utilizes CNNs to extract local feature interactions, such as the immediate relationship between latency and error. These representations are then passed to Transformer encoders to weigh their global importance [
16,
18]. This capability is particularly relevant for classifying clinical status, where the relationship between input biomarkers and cognitive status is often non-monotonic.
In this study, we propose a Hybrid CNN-Transformer framework that classifies five fundamental oculomotor metrics into three standard clinical stages: cognitively unimpaired, mild cognitive impairment, and mild-to-moderate dementia (MMD). By doing so, we aim to establish a highly accurate, objective, and non-invasive screening tool capable of identifying functional cognitive decline before the onset of severe neurodegeneration.
2 Materials and Methods
2.1 Participants
The study analyzed a large-scale retrospective cohort of 4034 participants, recruited to represent a broad spectrum of cognitive health ranging from cognitively normal aging to severe dementia. Demographic and clinical characteristics of the study cohort, including age, gender distribution, and baseline oculomotor metrics, are provided in Table 1. The primary clinical outcome measure was the ECog-12 score, a validated informant-reported assessment of functional cognitive decline. Scores on the ECog-12 scale range from 0 to 48, with higher scores indicating greater functional impairment. To ensure data integrity, the dataset underwent rigorous quality control, where participants with incomplete oculomotor records or missing clinical scores were excluded. The final analytical sample consisted of 4034 subjects.
The study was approved by the Ethics Committee of China Rehabilitation Research Center with Reference Number CRRC-IEC-RF-SQ-004-01. The participants gave informed consent, and the procedures were conducted in accordance with the Declaration of Helsinki.
2.2 Oculomotor protocol
Five quantitative oculomotor metrics were selected as model inputs based on their established sensitivity to executive dysfunction and inhibitory control deficits. These metrics included (1) fixation intrusive saccades, quantified as the total number of involuntary saccadic eye movements recorded during a sustained fixation task (N), which reflect deficits in baseline gaze stability; (2) pursuit catch-up saccades, quantified as the total number of corrective saccades superimposed on smooth pursuit tracking (N); (3) prosaccade latency, which reflects basic processing speed and reflex integrity; (4) antisaccade latency, which measures the time required to suppress a reflexive saccade and initiate a volitional eye movement; and (5) antisaccade error rate, which serves as a marker of inhibitory control failure. Prior to model ingestion, all input features were normalized to zero mean and unit variance (Z-score) to facilitate stable gradient descent.
2.3 Dataset partitioning
The normalized dataset was partitioned into training and testing sets using a stratified hold-out validation scheme with an 80/20 split. This resulted in a training set of approximately 3227 samples and a held-out independent test set of 807 samples. To address the inherent class imbalance where cognitively normal subjects significantly outnumbered those with impairment, we applied a “Gentle Class Balancing” strategy to the training set. Minority classes (mild and moderate CI) were oversampled to achieve a target prevalence of approximately 20% relative to the normal class. This approach ensured that the model remained sensitive to pathological cases without overfitting to synthetic data, while the test set remained unmodified to reflect real-world prevalence.
2.4 Model architecture
We developed a Hybrid CNN-Transformer architecture designed to map the oculomotor feature space directly to discrete clinical categories (Fig. 1). The network accepts an input sequence of dimension and proceeds through three primary stages. First, for local feature extraction, the input sequence is processed by a 1D Convolutional Layer (32 filters, kernel size 2) followed by Batch Normalization and a Rectified Linear Unit (ReLU) activation. Unlike standard dense layers, this convolutional block explicitly models the local relationships and trade-offs between adjacent metrics across the expanded five-feature sequence. Second, a global attention mechanism utilizing a custom Transformer Encoder block is employed to capture complex, non-linear dependencies. This layer utilizes multi-head self-attention mechanisms (16 hidden units, 2 heads) to dynamically weigh the relative importance of specific features based on the global context of the patient’s profile. Finally, the encoder output is passed through a global average pooling layer into a discrete classification head. This final stage utilizes a dropout layer (rate 0.20) and a fully connected dense layer with three units. Crucially, the network concludes with a Softmax activation function rather than a continuous Sigmoid regression function. This structural shift completely eliminates continuous scaling issues, allowing the model to directly output definitive probability distributions for the cognitively unimpaired, mild cognitive impairment, and MMD categories.
2.5 Training configuration
The model was implemented in MATLAB using the Deep Learning Toolbox. Training was executed using the Adam optimizer with a mini-batch size of 64 and an initial learning rate of . The training protocol spanned 70 epochs, with data shuffling performed at the start of every epoch. A gradient clipping threshold of 1.0 was applied to ensure stable convergence. Training was optimized using categorical cross-entropy loss.
2.6 Statistical analysis
Model output probabilities were assigned to one of three clinically distinct categories using prespecified, study-defined ECog-12 thresholds. The cohorts were defined as follows: (1) cognitively unimpaired (ECog-12 score ≤ 10), (2) mild cognitive impairment (ECog-12 score 11–17), and (3) mild-to-moderate dementia (ECog-12 score ≥ 18). The MMD category (ECog-12 ≥ 18) constitutes the most severe tier in this three-class framework. No higher-severity class was defined or included in the analysis.
3 Results
The primary objective of the hybrid deep learning model was to classify cognitive status based on the five input oculomotor metrics. On the held-out test set, the model achieved an overall classification accuracy of 85.13% (Fig. 2). Given the class imbalance inherent in this dataset (CU: 86.4% of test set), per-class performance metrics are reported: CU sensitivity 86.8% (precision 97.9%); MCI sensitivity 23.1% (precision 3.9%); MMD sensitivity 100% (n = 4; interpret with caution). The limited MCI sensitivity reflects the challenge of identifying a severely underrepresented class and is discussed as a key limitation.
As shown in Fig. 2, the model achieved an overall accuracy of 85.13%, predominantly driven by accurate CU classification (sensitivity 86.8%, precision 97.9%). MCI detection was limited, with a sensitivity of 23.1% and precision of 3.9%, reflecting the challenge of identifying a class comprising fewer than 2% of the test set. MMD achieved 100% sensitivity based on only 4 test samples and must be interpreted with caution. The high diagonal density indicates that the model rarely misclassified subjects between non-adjacent categories (e.g., misidentifying a normal subject as having MMD). The few misclassifications observed were primarily concentrated at the boundaries between adjacent severity levels, reflecting the inherent continuity of cognitive decline rather than model failure.
To further validate the reliability of the model for clinical application, we performed a Bland-Altman analysis (Fig. 3). The Bland-Altman plot reveals a mean bias of approximately −2.5 ECog-12 points. Notably, a substantial number of data points fall outside the 95% limits of agreement (approximately +5.0 to −9.5), concentrated among subjects with intermediate mean scores corresponding to the CU/MCI boundary region. This indicates that individual-level prediction reliability is lowest precisely where clinical discrimination is most consequential, and the model should be applied with caution for individual-level clinical decision-making.
Finally, we evaluated the model’s capability as a binary screening tool for detecting “Any Impairment” (ECog-12 > 10). The Receiver Operating Characteristic (ROC) curve (Fig. 4) yielded an AUC of 0.60, indicating modest but above-chance discriminatory ability for distinguishing any cognitive impairment (ECog-12 > 10) from normal cognition in this imbalanced sample.
4 Discussion
This study demonstrates that a hybrid deep learning model utilizing five oculomotor metrics can classify functional cognitive decline with high accuracy. Unlike traditional neuropsychological assessments that can be confounded by informant recall bias, anosognosia, and practice effects [
2,
19], our approach leverages the physiological integrity of the oculomotor system to estimate functional status. These results suggest that eye movement features contain sufficient information to distinguish between distinct stages of cognitive impairment, from normal aging to MMD.
A key contribution of this work is the application of a Hybrid CNN-Transformer architecture to tabular physiological data. While conventional models often struggle with the non-linear relationships between latency and error rates [
17], our dual-stage approach addresses this limitation. The initial 1D-Convolutional Layers capture local trade-offs such as the inverse relationship between speed (latency) and accuracy (error) while the subsequent self-attention mechanism allows for the dynamic weighting of these features within the global patient profile. A formal ablation study comparing the Hybrid CNN-Transformer architecture against simpler baselines including standalone CNNs, Random Forest, and multilayer perceptrons would provide empirical quantification of each component’s contribution and is an important direction for future work.
The development of an accurate, objective, and non-invasive screening tool has profound implications for the early detection of neurodegenerative disease. The ROC analysis yielded an AUC of 0.60. While above chance, this modest value underscores the challenge of binary impairment detection in a heavily imbalanced dataset and is acknowledged as a limitation requiring improvement in future work. Current gold-standard biomarkers, such as PET scans and CSF analysis [
20,
21], are invasive, expensive, and limited to specialized centers. In contrast, oculomotor assessment is rapid, cost-effective, and scalable.
Despite the promising results, this study has limitations. The analysis was performed on retrospective data, and while the cohort was large, external validation on independent datasets is necessary to ensure generalizability across different demographics and recording devices. It is important to acknowledge that the present model is trained on cross-sectional ECog-12 data and is therefore best characterized as a tool for objective assessment of current cognitive status at a single time point, rather than a predictor of longitudinal cognitive decline in the strict sense. Prospective longitudinal validation, tracking oculomotor changes over time relative to clinical cognitive outcomes, is an important next step and would be required before the framework can be positioned as a predictor of future cognitive decline. Additionally, while the ECog-12 is a validated measure of functional decline, future work should correlate oculomotor predictions with specific biological biomarkers, such as amyloid-beta or phosphorylated tau levels, to establish a direct link to underlying neuropathology [
20,
21]. Future iterations of the model could also incorporate additional eye movement features, such as smooth pursuit or fixation stability, which have also been shown to degrade in early Alzheimer’s disease [
3], to further refine predictive accuracy. The overall classification accuracy of 85.13% must be interpreted in the context of the severe class imbalance in this dataset. The limited MCI sensitivity (23.1%) reflects the inherent difficulty of detecting a minority class comprising fewer than 2% of the test set. From a clinical standpoint, the model’s limited MCI sensitivity (23.1%) indicates that the majority of MCI cases would be missed in a real-world screening scenario a limitation with direct implications for early intervention that must be addressed before clinical deployment. Furthermore, the Bland-Altman analysis further reveals that a considerable number of individuals fall outside the 95% limits of agreement, particularly near the CU/MCI diagnostic boundary, underscoring the model’s current limitations as a standalone diagnostic tool for individual patients.
5 Conclusions
In conclusion, we present a robust hybrid deep learning framework that effectively translates oculomotor behavior into clinically meaningful cognitive categories. By combining the physiological specificity of eye tracking with the computational power of CNN-Transformer architectures, we provide a viable pathway toward objective, scalable, and precise cognitive screening. These findings reinforce the potential of oculomotor metrics as digital biomarkers for Alzheimer’s disease and related dementias.
PDF
(1532KB)
