Assistive assessment of neurological dysfunction via eye movement patterns and ocular biometrics

Image preprocessing module

During the image preprocessing stage, three different resolutions are generated for the input image, with one resolution randomly selected during training. Furthermore, MobileNetV3, a lightweight object detection model, is utilized to segment the input image and identify the facial region. Standard data augmentation techniques, including random cropping, scaling, horizontal flipping, small-angle rotations (± 10°), and color jittering, are then applied to the detected facial region. To further refine the preprocessing stage, a region enhancement module based on an attention mechanism is designed to dynamically adjust the weights assigned to different regions in the image, thereby improving the model’s focus on crucial areas.

The accuracy of canthus detection is crucial for the diagnosis, treatment, and prognostic assessment of facial nerve movement disorders. Therefore, an optimization procedure is designed for ocular regions to enhance ocular feature representation in OFKNet^[6]. Specifically, the periocular region is finely divided into six functional subregions through a region-attention module: the upper eyelid, lower eyelid, internal canthus, lateral canthus, central fissure, and peri-pupil region. Dynamic weighting of these subregions is implemented based on the five facial movements. Specifically, during neutral expression and cheek puffing, the overall periorbital changes are minimal, resulting in relatively balanced weights across subregions. Eyebrow-raising elicits the highest weight in the upper eyelid region, while smiling significantly increases the weights of the lower eyelid and lateral canthus regions. Nose scrunching primarily involves elevated weights in the central fissure and lateral canthus areas, and cheek puffing emphasizes the lower eyelid and lateral canthus. Notably, the internal canthus region consistently maintains the highest or relatively high weighting across all five movements. This series of operations helps establish the internal canthus region as a critical area, and targeted cropping is applied to ensure that the extracted image preserves essential features surrounding the canthus. Subsequent adjustments are made based on the precise canthus location, focusing on capturing deformed features in this region. During color adjustment, contrast and color saturation in the canthus region are selectively enhanced to improve the distinguishability of its color distribution. Additionally, an image blur operation is applied while maintaining relative clarity in the canthus region. For other key regions beyond the canthus, conventional enhancement techniques, including cropping, flipping, and color adjustment, are employed, albeit with slightly lower intensity compared to the canthus region. In contrast, only mild enhancement operations are applied to background areas to ensure that feature optimization remains concentrated on the internal canthus region.

Feature extraction module

After preprocessing, the input image is denoted as x_p, which undergoes multiple convolutional operations to extract features at different levels. ConvNeXt is a convolution-based architecture that integrates a hierarchical modular design and efficient computational strategies, enabling superior feature extraction^[4]. It has demonstrated outstanding performance in image classification and object detection tasks^[4]. However, fully connected layers flatten the feature map into a vector, disrupting the spatial structure information. Since spatial structure is essential for accurately localizing key points in the facial region, fully connected layers are removed to preserve the original spatial layout of the feature map. Additionally, pooling layers are eliminated because they reduce the resolution of the feature map and lead to the loss of fine-grained details from the original image. In ConvNeXt^[4], the convolution operation of each layer can be expressed by a general convolution operation. For the convolution layer l, its input and output are denoted as F_l-1 and F_l, respectively. The kernel size is k × k, the stride of the convolution kernel is s, the padding at the edges of the input feature map is p, and the number of output channels is denoted as C_l. Then, the output of convolution layer l is

where w_l (m, n, d, c) is the weight of the convolution kernel l, represented as a four-dimensional tensor. m and n are position indices within the k × k convolution kernel, d is the channel index of the input feature map, c is the channel index of the output feature map, and (i, j) denotes the spatial position of the feature map. ReLU(·) denotes the activation function, i.e., the rectified linear unit.

When considering the convolution layer’s computation, the spatial size of the feature map is changed for each convolution operation^[15]. Let H_l × W_l denote the output feature map size of layer l, and its relation to the previous layer’s output feature map size is calculated as

OFKNet includes a bottleneck module, which improves the expressive capacity of the network by expanding and compressing the number of channels, enhancing the representation of characteristics while reducing the computational cost. The input feature of the bottleneck module is denoted as x_t. After a 1 × 1 convolutional layer, the number of channels is expanded to get the intermediate feature x_m. Then, the feature x_n is obtained after a 3 × 3 convolutional layer. Finally, the number of channels is restored through a 1 × 1 convolutional layer to obtain the output feature x_o:

where Ci is the number of input channels of x_t, and t is the expansion factor. ConvNeXt performs downsampling operations to convert high-resolution feature maps into low-resolution ones. Let F_h denote the input high-resolution feature map and F_l denote the output low-resolution ones. We obtain:

where r denotes the down-sampling factor. Based on this operation, the network extracts features at multiple scales, generating feature maps F₃, F₄, and F₅ at different resolutions, better capturing facial key points of varying sizes.

Feature enhancement and fusion module

Based on the multi-scale feature maps F₃, F₄, and F₅ extracted by the ConvNeXt, an improved PAN is employed to integrate features of different scales. PAN effectively merges low-level and high-level features through top-down and bottom-up paths. The high-level feature map F₅ is upsampled to the size of F₄ and weighted summed with F₄ to obtain an intermediate feature map F_u:

where Ψ(·) denotes the upsample process. Similarly, the low-level feature map F₃ is upsampled to the size of F₄ and weighted summed with F₄ to get the feature map F_d:

Then, the top-down and bottom-up feature maps F_u and F_d are fused to obtain the final multi-scale feature map F_p:

In this case, PAN facilitates information flow between feature maps^[16]. It enables features at different scales to complement each other, ensuring that the network considers both local details and global structural information. SE is then adopted by using PAN’s output to further enhance the feature representation ability^[17]. SE applies global average pooling (ϕ) to obtain global features for each channel. It then computes channel importance z_c and channel weights s_c using fully connected layers, i.e.,

where W₁ and W₂ are the weight matrices of the fully connected layers. Each channel’s weight s_c is multiplied with the input feature to obtain the final feature map $$ F_{P}^{\prime} \\ $$:

The combination of SE and PAN enhances the representation ability of the new feature map $$ F_{P}^{\prime} \\ $$, allowing it to better adapt to facial key points at different scales and complexities.

Key points detection module

Based on the processed feature map $$ F_{P}^{\prime} \\ $$, key point detection is performed using both the heatmap and the regression branches. The heatmap branch generates more discriminative features that produce accurate 2D Gaussian heatmaps, making the probability distribution of key points locations more precise. The regression branch learns the mapping between key points’ coordinates and features, leading to more accurate key point coordinate detection. The heatmap branch detects the 2D Gaussian heatmap for each key point through convolution operations. It is assumed that the input feature F_p is denoted as x, and the heatmap is obtained through two convolution operations:

where y₁ is obtained by subjecting x to a single convolutional transformation and a ReLU activation, and y₂ is obtained by subjecting y₁ to another convolution and a ReLU activation. h is the heatmap for each key point, ranging from [0,1], indicating the probability distribution of the point in space. The regression branch directly detects the (x, y) coordinates of key points. When the input features are y₁ and y₂, the coordinates of the key points are detected through the convolutional layer as $$ \hat{x}_{i} \\ $$ and $$ \hat{y}_{i} \\ $$:

The total loss function L_t consists of heatmap loss and regression loss. The heatmap loss L_h is based on the mean squared error, ensuring that the detected heatmap is close to the true heatmap. The regression loss function L_r ensures accurate coordinate detection, with weights w_i distinguishing the importance of different key points. They are obtained as:

where x_i and y_i are the true coordinates of the samples, N represents the number of samples used for calculating the heatmap loss, and M indicates the number of samples used for calculating the regression loss.

Experimental setting and evaluation metrics

All experiments are conducted using PyTorch on a single NVIDIA RTX 4090 GPU (24 GB memory). The network backbone adopts ReLU activations in all intermediate layers, while the final output layer employs a Sigmoid activation to generate the predicted outputs. The model is optimized with Adam for 60 epochs using a learning rate of 10^-3, batch size of 32, and a dropout rate of 0.25. Our dataset comprises 60 clinically diagnosed patients with facial nerve dysfunction, including 32 males and 28 females. All participants are assessed using the H-BGS and classified as grades II to V. Inclusion criteria require patients to have a confirmed diagnosis of facial nerve dysfunction without other neurological or facial disorders. Exclusion criteria include patients who failed to complete all facial movement tests. The clinical severity is categorized as mild, moderate, severe, and very severe, with each category reasonably represented in the dataset. We adopt a patient-level stratified splitting strategy to ensure that data from the same patient appear in only one subset. The distribution consisted of 42 patients for training, 9 for validation, and 9 for testing. We maintain proportional representation of different severity levels across all subsets during this division. The distribution of patients across severity grades is balanced among the training, validation, and test sets. Additionally, we ensure that different facial movement types are evenly distributed across all three subsets to maintain dataset representativeness and training fairness.

Each participant completes five facial movement tests under standardized conditions: NeutralFace, EyebrowRaise, ShrugNose, Smile, BlowCheek. Video capture uses a high-speed camera at 30 frames per second with 4K resolution. Each action is recorded for 5 s, generating approximately 63,000 video frames in total. Additionally, four eye movement tasks are performed: spontaneous eye blink (Spontaneous Eye Blink), voluntary eye blink (Voluntary Eye Blink), soft eye closure (Soft Eye Closure), and forced eye closure (Forced Eye Closure). For spontaneous and voluntary eye blinks, we increase the capture rate to 120 frames per second, with each action recorded for approximately 10 s.

From the collected video frames, we apply a keyframe extraction strategy. Specifically, for the annotation of 40-point eye landmarks, we select keyframes based on the following criteria: peak frames showing changes in eyelid opening and closing states (fully open, half-closed, and fully closed); transitional frames during action changes (from static to dynamic states, from closed to open); and frames where bilateral asymmetry is most pronounced. For the 14-point eye landmark annotations, the extraction criteria are relatively relaxed: in addition to the above keyframes, we include the beginning, middle, and end frames of each action sequence, as well as all frames showing significant changes in eyelid movement. We ensure that the selected keyframes possess high representativeness for subsequent analysis. Following keyframe selection, we conducted preprocessing on all videos. Initially, we perform quality screening to eliminate frames that are unsuitable for precise annotation due to blurriness, occlusion, or abnormal lighting conditions. This process yielded 3,600 frames with fine annotations of 40-point eye landmarks (approximately 60 frames per patient, covering key moments across different actions) and 5,400 frames with annotations of 14-point eye landmarks (approximately 90 frames per patient). Subsequently, two annotators with professional medical knowledge independently label all images. To evaluate annotation consistency, the inter-annotator agreement shows an average Euclidean distance of less than 2 pixels, demonstrating high consistency and reliability of the annotation results.

The main objective of this work is to evaluate the robustness and accuracy of different key point detection models for eye-related feature points under pathological conditions. Two manually annotated datasets are selected to achieve this: one with 14 eye feature points and another with 40 eye feature points^[14,18,19]. Five different facial landmark detection models are selected for comparative analysis, including MediaPipe Face Landmarker^[20,21], InsightFace^[22,23], Dlib68^[15,21] and Dlib81^[15,21]. By assessing their performance in detecting eye feature points, we aim to assess their applicability and precision in diagnosing facial nerve dysfunction. However, the models’ ability to process eye feature points varies significantly, particularly for datasets with different numbers of points. Specifically, InsightFace, Dlib68, and Dlib81 cannot detect all 40 eye feature points. To ensure fairness and comparability, a dataset with 14 eye feature points was also used for comparison with models that cannot handle the full 40 points. Detailed descriptions of the 14- and 40-point eye feature sets are provided in Table 1.

For the 40-feature-point dataset, OFKNet and MediaPipe Face Landmarker are evaluated, as they can detect a more detailed set of eye feature points, including the complete eye contour, pupil position, and other key points. In contrast, the dataset with 14 eye feature points is used for a broader comparison across all models. In particular, Dlib68 and Dlib81 do not detect the pupil center. To quantitatively compare detection accuracy, three evaluation metrics are employed: mean Euclidean distance (MED), normalized mean error (NME), and eye opening degree (EO). MED calculates the deviation between the detected and manually annotated key points using Euclidean distance, quantifying the error for each key point. NME extends MED by considering the influence of inter-individual eye size differences. It normalizes the error by using the intercanthal distance, eliminating individual differences^[16,23]. EO is an application-oriented metric that quantifies eye closure by analyzing changes in the eye fissure area, aiding in model sensitivity assessment for eye state changes in facial nerve dysfunction diagnosis^{[17-20,22,24-28]}. Specifically, MED is defined as:

where N is the total number of key points. $$ \left(x_{i}^{\mathrm{p}}, y_{i}^{\mathrm{p}}\right) \\ $$ denote the predicted coordinates of the key point i at a given frame. $$ x_{i}^{\mathrm{g}} \\ $$, and $$ y_{i}^{\mathrm{g}} $$ represent the corresponding ground truth coordinates. NME is defined as:

where d_a is the average predicted deviation for the given image, P₁ and P_r are the coordinates of the left and right internal canthus points, and dist(P₁, P_r) is the Euclidean distance between the internal canthus points of both eyes. Finally, EO is computed as:

where P₁ is the external canthus point, P₁₀ is the internal canthus point, and S(P₁, P₂, …, P₁₇) is the area of the eye fissure formed by these points.

Experiment results

One patient is selected to show eyebrow raise and voluntary blink actions as experimental data. Figure 3A and B compares the errors between the detected 40-point eye landmarks and the manually annotated points for OFKNet and MediaPipe Face Landmarker. It is shown that the results of the two models in detecting the eye contour and pupil location are different, with OFKNet exhibiting higher detection accuracy.

Figure 3. Error comparison for 40-point eye landmarks detection. Red points indicate the ground truth (manual annotation), while green points represent the model’s predictions. (A) OFKNet; (B) MediaPipe Face Landmarker.

Figure 4A-E presents the detection results of different models using the 14-point eye feature dataset. OFKNet and MediaPipe Face Landmarker demonstrate higher accuracy, particularly in detecting the pupil center and the internal and external canthus points.

Figure 4. Error comparison for 14-point eye landmarks detection. Red points indicate the ground truth (manual annotation), while green points represent the model’s predictions. (A) OFKNet; (B) MediaPipe Face Landmarker; (C) InsightFace; (D) Dlib68; (E) Dlib81. OFKNet: Ocular-enhanced Face Keypoints Net.

To complement the single-image overlays in Figures 3-4, we present dataset-level mean NME heatmaps for the eye landmarks in Figure 5A and B and Figure 6A-E. Each panel visualizes the canonical mean landmark positions after intercanthal alignment. The color and size of each marker encode the NME at that landmark (blue means low error, red means high error). The heatmaps show that OFKNet produces tightly localized errors with consistently low mean NME across both the 14-point and 40-point eye landmark sets, whereas competing methods exhibit higher errors at specific periocular landmarks, particularly near the pupil centers and the internal canthus.

Figure 5. Mean error distribution of eye landmarks (40 points). (A) OFKNet; (B) MediaPipe Face Landmarker. OFKNet: Ocular-enhanced Face Keypoints Net.

Figure 6. Mean Error Distribution of Eye Landmarks (14 points). (A) OFKNet; (B) MediaPipe Face Landmarker; (C) InsightFace; (D) Dlib68; (E) Dlib81. OFKNet: Ocular-enhanced Face Keypoints Net.

For completeness, Tables 2 and 3 report the landmark-wise median NME under the 40-point and 14-point eye landmark protocols, respectively. These tables reveal that the largest discrepancies between OFKNet and the baseline models are concentrated at the pupil centers and around the internal canthus, rather than being uniformly distributed along the eyelid contours. Competing models exhibit notably higher errors at these locations, whereas OFKNet maintains consistently low NME. This observation is consistent with the fact that pupil centers are highly sensitive to specular highlights and low iris-sclera contrast, and that the internal canthus is frequently affected by eyelid occlusions.

Compared with existing facial landmark detection architectures, OFKNet is specifically designed for fine-grained analysis of ocular movements in neurological dysfunction assessment. MediaPipe Face Landmarker maintains high-resolution feature representations, but its parallel multi-branch design results in high computational overhead, limiting its practicality in clinical environments requiring lightweight and real-time inference. InsightFace is optimized for identity recognition and global facial embeddings and thus lacks the sensitivity to subtle ocular movement variations and local asymmetry, making it inadequate for capturing clinically relevant neuromuscular dysfunctions. To address these challenges, OFKNet focuses on high-precision ocular landmark representation by employing a denser, anatomically-informed annotation scheme with 40 key points in the periorbital region. This fine-grained key point layout enables more accurate characterization of eyelid contours, canthus morphology, and pupillary kinematics, significantly enhancing sensitivity to subtle neuromuscular impairments. Building on this design philosophy, OFKNet incorporates three architectural improvements to further enhance performance. To evaluate the effectiveness of each component in OFKNet, we conduct a series of ablation experiments by progressively incorporating the proposed modules into the baseline framework. Specifically, ConvNeXt is adopted as the baseline feature extraction network, upon which we sequentially add the MobileNetV3-based local geometric enhancement module and the SE-PAN cross-scale adaptive feature fusion module. This results in three model variants: (a) ConvNeXt backbone; (b) ConvNeXt + MobileNetV3; and (c) ConvNeXt + MobileNetV3 + SE-PAN (i.e., the full OFKNet). The quantitative results are summarized in Table 4.

Table 4 shows that the ConvNeXt backbone exhibits relatively limited performance on the eye landmark detection task, particularly on the 40-point fine-level landmark set. This suggests that relying solely on high-level semantic features is insufficient for capturing the subtle geometric variations present in the eye region. Introducing the MobileNetV3 enhancement module yields a notable reduction in NME for both 40-point and 14-point settings. This improvement demonstrates that local geometric and contour-aware feature amplification is crucial for modeling the eyelid curvature, palpebral fissure shape, and canthus detail, thereby improving localization accuracy. Further incorporating the SE-PAN module leads to consistent performance gains. SE-PAN guides the network to adaptively fuse multi-scale features based on structural importance, enabling the model to dynamically focus on the most discriminative local regions under varying eye-opening states and individual anatomical differences. This confirms the effectiveness of the proposed cross-scale attention-guided fusion strategy. In summary, the ablation results indicate that the performance improvement of OFKNet does not stem merely from backbone selection or increased network capacity.

Figure 7A shows the cumulative error distribution curve for each model in the 14-eye feature points detection task, while Figure 7B presents the corresponding results for the 40 eye feature points. It is shown that the iteration curves for OFKNet and MediaPipe Face Landmarker are relatively smooth with low error, with OFKNet being the best. In contrast, comparative models show larger errors and a faster increase in error. Table 5 summarizes the per-image NME statistics for all models under the 40-point and 14-point eye landmark settings. For each baseline model, we perform a two-sided paired comparison against OFKNet on the common image subset. Two-sided paired Wilcoxon signed-rank tests between OFKNet and each baseline model on the common image subset using per-image NME as paired samples. Holm correction is applied to control the family-wise error rate across multiple baseline comparisons. The resulting p-values are reported in the rightmost columns. For both protocols, OFKNet achieves the lowest median NME, and all pairwise comparisons against OFKNet are statistically significant (P < 0.001), confirming that the improvements are not due to random variation. Beyond accuracy, we also benchmark the efficiency of all models in terms of model size (MB) and inference time (ms). The results are summarized in Table 6.

Figure 7. Cumulative error distribution curve for 14 and 40-point eye landmarks. (A) 14 eye feature points; (B) 40 eye feature points.

The EO is a key indicator of the severity of facial nerve dysfunction, providing an objective measure of eyelid closure for each eye. Figure 8(A–E) illustrates the EO degree over time for representative voluntary blinking sequences, comparing OFKNet with all four baseline models. It is shown that OFKNet produces a more precise and nuanced eye-opening degree curve than other methods. In all cases, the OFKNet curve closely tracks the subtle changes in eyelid position throughout the blink cycle, including minor partial closures and re-openings, resulting in a smooth, detailed waveform for the eye-opening degree. In contrast, the compared methods yield coarser or less responsive curves. Their outputs tend to miss or flatten minor variations in eye-opening, showing only significant dips corresponding to complete blinks but not finer oscillations during slight or incomplete blinks. These visual comparisons highlight the superiority of OFKNet in capturing minor variations in eye-opening degrees during voluntary blinking.

Figure 8. Comparison of eye opening degree curves for different models during voluntary blinking. (A) OFKNet; (B) MediaPipe Face Landmarker; (C) InsightFace; (D) Dlib68; (E) Dlib81. OFKNet: Ocular-enhanced Face Keypoints Net.

In addition, offering a data-driven representation of the eye’s opening and closing states enables a quantitative assessment of eye-opening ability throughout the blinking cycle. This measurement is applicable across different blinking stages and facilitates bilateral comparisons between the left and right eyes, thereby characterizing neural motor function asymmetries. The dynamic variations in EO degree over time offer a clear visual representation of facial nerve impairment and recovery, making it particularly valuable in the clinical assessment of facial nerve dysfunction. In this work, we dynamically measure the EO degrees of both eyes in patients exhibiting significant facial nerve dysfunction symptoms. To enhance interpretability, we generate time-series graphs to illustrate temporal changes in EO degree. Furthermore, keyframes extracted from patient data highlight the accuracy and reliability of the measurements, reinforcing their potential utility in neurological evaluation and treatment monitoring. Therefore, we conduct measurements of four characteristic eye movements in patients with facial nerve dysfunction: Spontaneous Eye Blink, Voluntary Eye Blink, Soft Eye Closure and Forced Eye Closure. These movements effectively simulate the majority of eye activities performed by individuals with facial paralysis in daily life, enabling a comprehensive analysis of EO degree variations in both eyes. A detailed introduction to each movement is provided below.

• Spontaneous Eye Blink refers to the involuntary blinking of the eyes that occurs without external stimuli or conscious control. In this work, patients are instructed to relax their facial muscles through verbal cues, allowing them to blink naturally for over 10 s while maintaining a relaxed state.

• Voluntary Eye Blink refers to a consciously controlled and deliberately initiated blink. In this study, patients are instructed to keep their heads still, maintain direct eye contact with the camera, and blink three times consecutively.

• Soft Eye Closure is a natural, unconscious action, and neurological abnormalities often first manifest in such subtle movements. By assessing the ability to close the eyes softly, we can preliminarily evaluate whether the orbicularis oculi muscle is functioning properly.

• Forced Eye Closure is a conscious action that serves as an assessment of the strength of the eyelid muscles. In patients with nerve damage or muscle weakness, insufficient muscle tension may be observed during forceful eye closure, resulting in incomplete eyelid closure or difficulty closing the eyes. By having the patient perform a forceful eye closure, we can evaluate the intensity of the action and analyze the speed, symmetry, and smoothness of the closure through video analysis.

The line charts derived from the EO degree data clearly illustrate the temporal variation in eye openness under different eye movement tasks. In each chart, the x-axis denotes the video frames, while the y-axis represents the EO degree. The blue and red lines correspond to the EO degrees of the left and right eyes, respectively. These plots reflect the patient’s ability to control facial muscles and serve as a basis for the quantitative assessment of facial nerve dysfunction. From these line charts, one can intuitively observe dynamic changes in EO degrees across time during various standard eye movements. Notably, at specific moments, the difference between the EO degrees of the two eyes reaches a maximum, indicating a marked asymmetry in binocular motor function. The most pronounced asymmetry consistently occurs during the peak amplitude phase of eyelid closure and exhibits a reproducible pattern across different patients and tasks. These phases, serving as spatiotemporal windows where features of muscle weakness or paralysis are most evident, hold important diagnostic value. Identifying and selecting these time points as key frames facilitates accurate characterization of the condition’s severity and provides reliable reference data for clinical evaluation. To further validate this observation, we extracted key frames from three patients diagnosed with facial nerve dysfunction across four movement types and visualized their EO degree time series. Figures 9-20 show that in each case, one eye is fully closed while the contralateral eye remains partially open. This is consistent with the typical clinical manifestations of facial nerve dysfunction.

Figure 9. Spontaneous Eye Blink in Patient 1.

Figure 10. Voluntary Eye Blink in Patient 1.

Figure 11. Soft Eye Closure in Patient 1.

Figure 12. Forced Eye Closure in Patient 1.

Figure 13. Spontaneous Eye Blink in Patient 2.

Figure 14. Voluntary Eye Blink in Patient 2.

Figure 15. Soft Eye Closure in Patient 2.

Figure 16. Forced Eye Closure in Patient 2.

Figure 17. Spontaneous Eye Blink in Patient 3.

Figure 18. Voluntary Eye Blink in Patient 3.

Figure 19. Soft Eye Closure in Patient 3.

Figure 20. Forced Eye Closure in Patient 3.

This phenomenon carries two important academic implications. First, it may extend beyond the scope of traditional clinical observation; our method is capable of effectively capturing and documenting it, thereby addressing a gap in current diagnostic and therapeutic approaches within specific observational dimensions. Second, even if this phenomenon has been qualitatively recognized in clinical practice, conventional methods often lack the precision and reproducibility required for quantitative analysis, which constrains the development and application of targeted treatment strategies. Our approach overcomes these limitations by enabling objective and repeatable assessment, thereby offering novel support for the refinement and scientific advancement of clinical diagnosis and intervention.

Limitations

Our analysis focuses on the ocular region. While OFKNet can detect keypoints across the entire face, we have not yet fully examined how features from other facial areas, such as the mouth corners or forehead, might contribute to neurological assessment. Given that facial nerve dysfunction affects multiple muscle groups, a more comprehensive analysis integrating these regions could yield richer diagnostic insights. In addition, testing the system’s performance with standard medical imaging equipment would help determine its practical deployment potential. Finding ways to reduce hardware requirements without sacrificing accuracy represents an important direction for future development.

We demonstrate OFKNet’s effectiveness in facial paralysis assessment, but its utility for other neurological conditions (Parkinson’s disease and Alzheimer’s disease) remains to be established. Different neurological disorders may manifest distinct ocular movement patterns, requiring algorithm adaptations and dedicated clinical validation. Additionally, longitudinal studies tracking patients over extended periods would clarify whether this approach can effectively monitor disease progression and treatment response.

In addition, our study has several limitations arising from the controlled nature of data collection. All recordings are captured under standardized indoor clinical conditions with stable lighting, fixed patient positioning, and high-resolution imaging equipment. As a result, the current evaluation does not fully reflect factors such as non-uniform illumination. Furthermore, the patient cohort consisted primarily of East Asian individuals, and periocular morphology varies substantially across ethnic groups. Therefore, cross-ethnic validation is required to establish generalizability. Finally, because the dataset is derived from continuous video sequences, the frame-level redundancy inherent in temporal sampling may introduce subtle correlations that increase the risk of model overfitting, even though patient-level data splitting is applied to prevent identity leakage.

In conclusion, this work presents OFKNet, a fine-grained ocular landmark detection framework that substantially improves the accuracy of eye-related keypoint localization during dynamic eyelid motion. Beyond technical performance, OFKNet provides quantitative eye-opening trajectories and visualized asymmetry indices that can support clinical decision-making. These measurements help clinicians evaluate blink completeness, eyelid closure fatigue, and interocular imbalance, which are important factors for treatment planning, prognostic assessment, and follow-up monitoring. OFKNet operates in real time on standard clinical computing hardware, indicating that it can be deployed within existing examination rooms without major infrastructure changes. This supports practical integration into routine neurological assessments and facial nerve clinics.

Future work will focus on several translational directions. The next step is prospective multicenter validation to ensure robustness across diverse imaging devices, ethnic groups, and recording environments, which is necessary for broader clinical adoption and regulatory evaluation. Quantitative blink dynamics obtained from OFKNet may also be incorporated into current grading systems to complement House-Brackmann or Sunnybrook scales with objective and continuous measurements. In addition, expanding analysis beyond the ocular region and integrating multimodal signals such as surface electromyography, electroencephalography, or professional eye-tracking will help establish a more comprehensive and data-driven framework for neurological function assessment.