An Assistive System for Early Prediction of Microvascular Instability in Diabetic Retinopathy Using a Causal-Aware Pillar Transformer Network

4.1 Input representation

The CAPT-DR method explains that the fusion of multimodal data into a common latent space occurs right from the beginning by using a unified mathematical formulation for the diverse multimodal data inputs, which makes it easier to combine retinal imaging data and clinical data in a consistent and coherent way. As the early risk prediction for DR is, in fact, a multimodal problem, the solution involves three large sources of input data: 2D fundus image data, 3D OCTA volumetric image data, and clinical data in structured form. In the process of encoding each modality into a mathematical form, a lot of stress is placed on maintaining the modality-specific, spatial, structural, and physiological meaning while still allowing for interactive learning and fusion. A single mathematical form is used here to encode the localized fine details of microvascular morphology and the systemic risk interactions as a whole [18]. Let the set of fundus images be a batch of normalized retinal images, each of which can be modeled as a continuous spatial function on the retina. The OCTA imaging data is represented as volumetric flow and density tensors that characterize the capillary network morphology together with local depth-wise perfusion levels. Besides, clinical data like metabolic and physiological markers (e.g., HbA1c, blood pressure, lipid profile) plus disease duration are registered as structured feature vectors. These heterogeneous representations are mapped into a common embedding space by modality, using specific transformation functions, thus ensuring semantic alignment and scale consistency between different modalities prior to fusion (Figures 2-6).

This architecture enables the model to operate on a harmonized multimodal feature manifold rather than on separate data representations, thus permitting significant cross-modal interactions and hierarchically detecting microvascular instability patterns.

$F=\left\{I_i \in R^{H * W * C} \mid i=1,2, \ldots, N\right\}$            (1)

During training, Eq. (1) converts normalized fundus images into structured spatial tensors that preserve retinal vessel morphology and texture information. During inference, the same representation is used for extracting hierarchical vascular features.

This representation enables image height H, image Width W, and color channels C of a fundus sample with N samples of fundus images to be represented as structured spatial tensors that maintain pixel-level anatomical and vascular data in Eq. (1). This formulation supports spatial convolution, attention operations and hierarchical encoding across different areas of the retina; therefore, it serves as the base visual input structure for learning about microvasculature morphology and textures within the retina. The tensorized structure allows for scalability during batch learning and consistent extraction of features across large classes in Eq. (2).

$O=\left\{V_i \in R^{H * W * C} \mid i=1,2, \ldots, N\right\}$           (2)

During training, Eq. (2) encodes OCTA volumetric scans into depth-aware vascular tensors containing capillary perfusion and flow information. During inference, the representation enables the model to identify three-dimensional microvascular abnormalities and perfusion heterogeneity associated with early DR progression.

The following volumetric approach maintains depth-resolved capillary perfusion volume flow and perfusion throughout all layers of the retina by providing an anatomical framework for characterizing depth-resolved capillary perfusion volumes. It enables modeling of three-dimensional vascular connections, as well as the integrity of microcirculation within the retina. The hierarchical spatial-depth encoding of the tensor structure provides early detection of microvascular abnormalities. This spatial encoding is crucial for identifying subclinical vascular remodeling that may not be seen with two-dimensional imaging methods.

Eq. (3) transforms structured clinical measurements such as HbA1c, blood pressure, lipid profile, age, and diabetes duration into numerical feature vectors. These variables provide systemic metabolic information that complements retinal imaging features during risk prediction.

$C=\left\{c_i \in R^M \mid i=1,2, \ldots, N\right\}$                         (3)

where, M is the number of structured clinical variables in Eq. (3). This vectorized expression captures systemic metabolic and physiological risk factors in a concise form. It makes it possible to integrate imaging biomarkers directly into the prediction framework. The layout allows for causal modeling and risk factor disentanglement. It offers a context layer of physiology that goes along with the retinal microvascular images in Eq. (4).

Eq. (4) combines fundus, OCTA, and clinical feature embeddings into a unified latent representation. During training, joint optimization enables cross-modal interaction learning, whereas during inference, the fused embedding serves as the input to the PT module.

$Z_i=\emptyset_f\left(I_i\right) \oplus \emptyset_o\left(V_i\right) \oplus \phi_c\left(C_i\right)$              (4)

where, $\phi_f(\cdot), \phi_o\left(V_i\right)$ and $\phi_c\left(C_i\right)$ are modality-specific embedding functions, and $\bigoplus$ denotes feature fusion. This formulation integrates various heterogeneous modalities into a common latent embedding space. It facilitates cross-modal interaction between retinal morphology and systemic risk factors. The fusion operation keeps intact modality-specific semantics and at the same time allows for joint optimization. This embedding is the main input to the PT and causal modeling layers. After such mathematical encoding, the modalities are combined into a shared latent space of representation, which serves as the main input to the PT and causal modeling layers. The fusion embedding Z_i keeps the spatial microvascular structure from the imaging data intact while, at the same time, embedding systemic metabolic risk information. Therefore, it allows for joint optimization of visual and clinical domains. By synthesizing a single representation at the very beginning of the input stage, the framework ensures that hierarchical encoding, causal inference, and predictive modeling at the downstream levels will operate on a unified multimodal feature manifold rather than on separate modality-specific representations. This method of selecting the scheme is of utmost significance in the context of preclinical microvascular instability pattern detection by deep learning and understanding the intricate relationship between retinal morphology and systemic disease processes for the prediction of early DR.

4.2 Pillar-based hierarchical encoding

The pillar-based Hierarchical Encoding component of the CAPT-DR system is basically the backbone of the system. The purpose of this component is to preserve the continuity of the retinal microvascular patterns and, at the same time, assist the system in learning multi-scale representations. The patch-based tokenization process normally breaks the vascular patterns into spatial units that are quite distant from each other, and therefore, they normally cause difficulties in determining the topology of the vessels and the continuity of the capillaries, (Figure 2). To overcome this difficulty, CAPT-DR has developed a pillar-based spatial encoding technique in which the retinal feature maps are partitioned into vertically aligned spatial columns (pillars) that preserve anatomical continuity across scales [23]. Such a technique enables the microvascular patterns to be anatomically aggregated to the model’s learning of hierarchical vascular representations that are not only consistent with the actual retinal anatomy but also with the artificial grid divisions. The pillar is a type of biologically motivated encoding method that exemplifies longitudinal vessel continuity and spatial dependency across retinal areas.

Eq. (5) partitions the fused feature representation into vertically aligned pillar tokens. This operation preserves retinal vessel continuity and creates anatomically meaningful units for hierarchical representation learning.

$P=\psi(Z)=\left\{P_k \mid p_k=Z\left[;, x_k: x_{k+1},:\right]\right\}$            (5)

This method takes the multimodal embedding Z and divides it into vertical segments (pillars) as shown in Eq. (5). Each pillar preserves the spatial continuity of the retinal axis. It ensures that the structure is consistent along the microvascular routes. Such segmentation is the foundation of the hierarchical encoding.

Eq. (6) performs vertical feature aggregation within each pillar. During training, neighboring vascular structures contribute jointly to feature learning, thereby preserving local vessel continuity and reducing spatial fragmentation.

$a_k=\sum_{y=1}^H w_y \cdot p_k(y)$        (6)

Mathematically, this pillar-level function performs the operation of convolution vertically across the spatial dimension. It is able to encode the dependencies between blood vessels along the length of the lobule in Eq. (6). The function that determines the weights w_y maintains anatomical relevance. This makes it possible to have a biologically meaningful compression of features.

Eq. (7) extracts vascular representations at multiple spatial resolutions. This enables simultaneous learning of fine capillary patterns and large-scale retinal vascular structures important for early disease prediction.

$m_k=\sum_{s=1}^S \alpha_s \cdot F_s\left(a_k\right)$           (7)

The hierarchical feature extraction of the network is enabled by this formulation across different spatial scales. It captures both vascular morphology at the capillary level and the overall vascular architecture in Eq. (7). Scale functions F_s preserve the information at different resolutions. This facilitates the detection of microvascular instability at an early stage.

Eq. (8) enforces structural continuity between adjacent pillars. During optimization, the constraint penalizes discontinuities in vessel representations and encourages anatomically consistent feature learning.

$\mathcal{L}_{s c}=\sum_{k=1}^{k-1}\left\|m_k-m_{k+1}\right\|_2^2$          (8)

This constraint maintains a continuous connection between neighboring pillars. It keeps vessel connectivity and spatial smoothness as in Eq. (8). It stops the breaking up of vascular structures. This enhances the anatomical realism of the representations learned.

Eq. (9) computes attention weights between pillar embeddings. The mechanism allows the model to selectively focus on retinal regions exhibiting microvascular instability and abnormal vascular connectivity.

$e_k=\operatorname{softmax}\left(Q_k K_k^T\right) v_k$         (9)

This attention formulation models dependencies across pillar features. It enables selective focus on unstable vascular regions in Eq. (9). It captures long-range microvascular interactions. This enhances discriminative structural learning.

Eq. (10) generates the final hierarchical pillar representation by integrating continuity-preserved and multi-scale vascular features. This representation becomes the input to the PT learning stage.

$\hat{e}_k=\frac{e_k-\mu}{\sigma}$            (10)

Via this pillar-based hierarchical encoding, the CAPT-DR model develops structured microvascular representations that maintain anatomical continuity, spatial coherence, and biological realism in Eq. (10).

The combination of joint segmentation, aggregation, multi-scale encoding, continuity constraints, and attention-based embedding results in a comprehensive morphological representation that accurately reflects the real retinal vascular morphology [24]. Such hierarchical encoding provides the framework for subsequent transformer operations, enabling the model to discover subtle features of microvascular dysfunction beyond the detection capabilities [25] of previous tokenization methods. By preserving the feature of vascular structure throughout the encoding process, the model can deliver not only highly accurate predictions but also understandable results [26].

Figure 2. Pillar tokenization structure

4.3 Pillar Transformer module

The PT module is the central representation learning mechanism of the CAPT-DR framework that facilitates globally [13, 27] structured reasoning about the hierarchical encoding of retinal microvascular features in Figure 3.

Figure 3. Pillar Transformer (PT) block

In contrast to conventional transformer architectures that rely on flat, patch-based token sequences [28], the PT operates on biologically structured pillar embeddings that preserve spatial continuity and vascular topology. This mode of operation equips the model with the potential to learn local microvascular changes as well as spatial dependencies over quite a long distance between the retinal sites. By working with pillar-level tokens, the transformer can identify very subtle and even preclinical changes of the vascular structure, such as remodeling patterns, capillary loss and perfusion heterogeneity, which are at the very essence of the prediction of early DR [6]. The module integrates self-attention, multi-head attention, positional encoding, and structure-aware weighting to achieve hierarchical, anatomically coherent representation learning [29] in Formula (1). By processing pillar-level tokens, the transformer is able to detect very early preclinical changes of the vascular topology, such as remodeling, capillary drop-out, and perfusion heterogeneity, which form the very basis for the prediction of early DR [28]. The module integrates self-attention, multi-head attention, positional encoding, and structure-aware weighting to enable hierarchical, anatomically consistent representation learning [29] in Eq. (11).

$S A(X)=$ Softmax $\left(\frac{Q K^T}{\sqrt{d}}\right)$          (11)

Eq. (11) calculates contextual relationships among pillar tokens. During training, the attention mechanism captures long-range dependencies between spatially distant retinal vascular regions.

This process calculates contextual dependencies between pillar embeddings. It facilitates global interactions among spatially distant vascular regions. It helps with learning long-range microvascular relationships in Eq. (12).

$M H A(X)=\bigoplus_{h=1}^H S A_h(X)$           (12)

Eq. (12) measures similarity between query and key representations. Higher similarity values indicate stronger relationships between retinal regions and contribute more significantly to attention-based feature learning.

This equation facilitates parallel attention to multiple representation subspaces. It captures various vascular patterns simultaneously. It improves representation capabilities in Eq. (13):

$P_k=\sin (w k) \oplus \cos (w k)$            (13)

Eq. (13) performs parallel attention computations across multiple representation subspaces. This allows the model to simultaneously capture diverse vascular characteristics including vessel density, connectivity, and perfusion patterns.

This encoding preserves the spatial ordering of pillars. It maintains anatomical positional consistency. It enables structure-aware spatial reasoning.

Eq. (14) integrates anatomical priors into the attention mechanism. The weighting scheme encourages learned representations to remain consistent with retinal vascular organization and biological structure:

$A_s=A \odot S$          (14)

This weighting combines anatomical structure priors with attention maps in Eq. (14). It regularizes learning with residual connections and normalization. It produces structured global retinal embeddings [30-34].

Eq. (15) produces the final transformer embedding after residual learning and normalization. This global representation summarizes retinal microvascular information for causal analysis and risk prediction.

$T = LayerNorm (M H A(X)+X)$              (15)

This formulation produces the final pillar-transformer representation in Eq. (15). It stabilizes learning through residual connections and normalization. It generates structured global retinal embeddings [30-34].

4.png

Figure 4. Raw retinal image vs AI-extracted feature representation

Figure 4 represents, in a graphical way, the conversion of the study [5] original retinal images into AI, mechanism, aware representation learning, feature maps, via hierarchical encoding and transformer-based processing. It shows how important microvascular patterns, vessel structures, and pathological information are identified and represented in a way that is of high utility [35]. It also shows how the model's internal knowledge and feature abstraction are achieved. By doing so, the PT, as the key component in the structured transformer architecture, facilitates the smooth integration of microvascular features at different levels with global context reasoning [36]. The incorporation of biologically structured tokens, multi-head attention, and structural weighting allows the model not only to accomplish anatomical realism but also to go deep into semantic abstraction. As a result, the generated feature representations can very effectively predict the onset of the disease and can be interpreted in terms of vascular structure and connectivity. Making the transformer's learning consistent with retinal anatomy, CAPT-DR is able to perform mechanism-aware representation learning that is way (more than) the conventional ViT methods in recognizing the first slight preclinical microvascular instability patterns [37].

To assess the clinical relevance of transformer attention maps, a subset of 200 retinal images was independently reviewed by two board-certified ophthalmologists. Regions highlighted by the CAPT-DR attention mechanism were compared against clinically annotated microvascular abnormalities, including microaneurysms, capillary dropout regions, vessel-density reduction areas, and perfusion defects observed in corresponding OCTA scans.

Region-level agreement was quantified using Intersection-over-Union (IoU) and Dice Similarity Coefficient (DSC). The proposed model achieved an average IoU of 0.76 and a DSC of 0.82 across all lesion categories. Inter-observer agreement between specialists yielded a Cohen’s κ coefficient of 0.84, indicating substantial agreement.

These findings suggest that the attention maps generated by CAPT-DR correspond closely with clinically meaningful retinal abnormalities and provide interpretable evidence supporting model predictions.

4.4 Structural Causal Modelling layer

The SCM layer is the interpretability and reasoning core for the CAPT-DR framework. It allows the model to move beyond models that only rely on correlation, thus, prediction and mechanistic understanding of the disease progression. Deep neural networks, however, while being highly efficient in pattern recognition tasks, are unable to properly separate causal associations from correlations. Thus, in this way, they become less trustworthy when applied to the clinical domain. To remedy this issue, the model introduced a causal inference layer that formally captures the directed mechanisms of the systemic metabolic risk factors, features of microvascular instability obtained by the network, and the onset of DR. In addition to allowing risk to be broken down into direct and indirect effects, a causal inference layer also enables counterfactual reasoning and, most importantly, intervention simulation which is of clinical relevance. By embedding a causal structure in the training process, the proposed method achieves both high prediction accuracy and the capability to offer mechanistic interpretability.

The SCM layer is modeled as a directed acyclic graph (DAG), where the nodes are clinical variables, representations of microvascular instability, and disease outcome states, and the edges are directed causal dependencies. Along with this graph, the PT learned representations are integrated, facilitating a joint optimization of data-driven learning and causal reasoning. The structural equations that are associated with each variable in the graph define the generative process, thus, the model can simulate the effect of changes in the systemic factors on retinal microvascular instability and the subsequent risk of disease in Eq. (16). This fusion allows a clinically interpretable breakdown of the different components of risk and results in scenario-based predictive modeling that can be used in preventive intervention planning.

$G=(v, \epsilon)$           (16)

Eq. (16) formally defines the structural causal graph used in CAPT-DR. Nodes represent systemic risk factors and retinal biomarkers, whereas directed edges represent clinically established cause-effect relationships.

The structural causal graph was designed using ophthalmological and endocrinological domain knowledge. Nodes represent HbA1c level, diabetes duration, blood pressure, lipid profile, retinal microvascular instability features, and DR outcome. Directed edges were established according to clinically validated causal pathways reported in DR literature.

Confounding effects were controlled using back-door adjustment and stratification over age, gender, and baseline disease severity. The graph assumes causal sufficiency, consistency, positivity, and conditional exchangeability. Counterfactual estimation was performed using Pearl’s intervention framework under the do-calculus formulation.

Sensitivity analysis was conducted by randomly perturbing graph edges and evaluating the stability of causal effect estimates across 100 bootstrap iterations.

The causal graph consists of the following nodes:

HbA1c (H), Diabetes Duration (D), Systolic Blood Pressure (B), Lipid Profile (L), Retinal Microvascular Instability Features (M), and Diabetic Retinopathy Outcome (R).

Directed causal relationships are defined as:

H → M, D → M, B → M, L → M,

H → R, D → R, B → R, L → R,

M → R.

Age, gender, and baseline retinal severity are treated as confounding variables and are controlled using stratified adjustment procedures. The graph structure was derived from established ophthalmological and endocrinological evidence rather than being learned directly from observational data.

This formulation defines the causal structure as a directed acyclic graph. Nodes represent variables and learned retinal features. Edges encode directed cause–effect relationships.

$X_i=f_i\left(P a\left(X_i\right)\right)+\varepsilon_i$             (17)

Eq. (17) models each causal variable as a function of its parent variables and stochastic noise. During training, the formulation captures causal dependencies among metabolic factors and retinal abnormalities.

This equation models each variable as a function of its causal parent in Eq. (17). It incorporates both deterministic and stochastic influences. It defines the generative causal process of the system.

$D C E(X \rightarrow Y)=\frac{\delta Y}{\delta X}$           (18)

Eq. (18) estimates the direct causal influence of a risk factor on disease progression while excluding mediation effects. This helps quantify independent clinical contributions of individual variables.

This formulation quantifies the immediate causal impact of a variable in Eq. (18). It isolates direct influence without mediation. It supports targeted risk factor analysis.

$I C E(X \rightarrow Y)=\sum_{m \in \mathrm{M}} \frac{\delta Y}{\delta m} \frac{\delta m}{\delta X}$            (19)

Eq. (19) quantifies causal effects transmitted through intermediate variables. It captures multi-step biological pathways involved in retinal disease development.

This expression models mediated causal pathways in Eq. (19). It captures multi-step influence propagation. It reflects systemic risk transmission mechanisms.

$TCE(X \rightarrow Y)= DCE (X \rightarrow Y)+I C E ?(X \rightarrow Y)$             (20)

Eq. (20) combines direct and indirect causal influences to estimate the total impact of a clinical factor on DR risk.

This equation accounts for both direct and indirect effects together in Eq. (20). It is the total causal impact on the risk of disease. It allows for the evaluation of interventions on a comprehensive level.

By combining SCM with deep representation learning, CAPT-DR converts predictive modeling into a mechanism-aware inference system. The causal layer here makes it possible to separate systemic metabolic effects from retinal structural instability and this provides clinically understandable explanations of the risk pathways. The power of modeling direct and mediated influences gives the doctors the possibility to get clarity not only on the fact if the patient is at risk, but also on the reasons behind the risk and the ways it can be reduced. Causal interpretability of such kind is one of the main issues for an interpretable predictive framework in clinical decision support. In this way, it will allow for setting up personalized prevention strategies, policymaking, and intervention planning based on the evidence of early DR management.

4.5 Counterfactual Reasoning Engine

The Counterfactual Reasoning Engine is a component that allows the CAPT-DR system to perform a hypothetical intervention analysis which is illustrated in Figure 5. In other words, it is from predictive modeling to clinical intelligence that is actionable. Typically, AI risk prediction systems make predictions from the distribution of data observed; however, they lack the ability to answer "what if" type questions which are fundamental to preventive medicine and phenotypic care. Thus, the counterfactual element in the account of the problem can be used to simulate changes that may occur in the systemic risk factors and be able to estimate the expected effect of those changes on microvascular instability and the risk of disease progression. Therefore, the framework has the ability not only to describe the prediction but also to provide decision support that is prescriptive and hence it can be used for planning a proactive intervention and individual risk management.

What if thinking is a way to understand the effects of changes? This is done by using a limited number of variables in the causal graph and keeping the underlying generative structure. When the clinical variables that are controllable, such as glycemic control, blood pressure, and lipids levels, are intervened, the model generates new patient histories and calculates the expected results of the different clinical scenarios. The simulated paths resulting in microvascular instability and the risk of DR through the causal network and the learned retinal representations. This approach of causal inference combined with deep learning representations allows the system to produce not only clinically meaningful counterfactual explanations but also personalized intervention insights in Eq. (21).

$P(Y \mid d o(X=x))$          (21)

Eq. (21) establishes the mathematical basis for counterfactual simulation. The formulation enables evaluation of hypothetical clinical interventions under alternative patient conditions.

This operator represents an external intervention on variable X in Eq. (22). It breaks natural causal dependencies in the graph. It enables controlled causal manipulation.

$R_{c f}=E\left[Y \mid d o\left(X=x^{\prime}\right)\right]$         (22)

Eq. (22) applies an external intervention using Pearl’s do-operator. This operation disconnects natural causal dependencies and simulates controlled clinical modifications such as HbA1c reduction.

Figure 5. Counterfactual simulation flow

This formulation computes risk under a hypothetical intervention. It estimates outcomes for altered clinical states. It supports personalized risk projection.

$\Delta R=R_{o b s}-R_{c f}$           (23)

Eq. (23) computes the predicted disease risk under an intervention scenario. The resulting value estimates how risk changes after a hypothetical clinical action.

This function quantifies the impact of intervention in Eq. (23). It measures potential risk reduction. It enables benefit–impact analysis.

$\pi=\arg \min _\pi E[Y \mid \operatorname{do}(X=\pi(X))]$            (24)

Eq. (24) identifies intervention strategies that maximize projected risk reduction. This formulation supports personalized prevention planning and decision support.

This formulation maximizes strategies for intervention in Eq. (24). It is a great help to decision policy design. It makes it possible to plan prevention in an automated way.

By means of a counterfactual reasoning engine, CAPT-DR is not limited to risk prediction, but it also offers proactive clinical decision support. The system is able to imagine different physiological trajectories, assess how effective the intervention is, and create personalized prevention strategies based on causal reasoning instead of mere correlation. In this way, the model is turned into a digital decision assistant that can help clinicians choose the best intervention paths for disease prevention at the early stage. The addition of counterfactual intelligence to the prediction pipeline renders the system a clinically actionable, explainable system for the early management of DR risk.

4.6 Microvascular Instability Index

The MII is proposed as a quantitatively interpretable biomarker that can be directly sensed from model outputs. It is intended to distinguish very early changes in the retinal microvasculature before dysfunction, as illustrated in Figure 6. Furthermore, whereas typical clinical biomarkers are either lesions visible or late-stage disease manifestations, MII was intended to quantify the very subtle changes in vascular morphology, connectivity, and flow stability that are preclinical. It can be considered as a gathering of information from the transformer attention mechanism, hierarchical pillar representations, and structural continuity constraints into a single instability measure. In short, this index is a biological equivalent of retinal microvasculature health in the service of title disease prediction; hence it facilitates early patient stratification and timely targeted intervention.

The model internally (MII) formulation exploits internal signals of a model such as attention distributions and structural connectivity representations to estimate the degree of instability at a system level. Attention entropy is used to evaluate the uncertainty and disorder of the microvascular focus patterns that have been learned, while the loss in vascular connectivity quantifies how much spatial continuity within the retinal network has been broken. These three types of signals are aggregated at various levels of the hierarchy and then normalized to produce a stable, interpretable index value. The MII score thus obtained provides a continuous risk metric that can be mapped to clinically important risk categories, thereby enabling tiered screening and personalized monitoring approaches in Eq. (25).

$H=-\sum_i p_i \log \left(p_i\right)$         (25)

Eq. (25) calculates attention entropy as a measure of uncertainty within learned vascular representations. Higher entropy values indicate greater microvascular instability.

5.png

Figure 6. Microvascular Instability Index (MII) computation pipeline

The experimental setup of the CAPT-DR framework is designed to conduct a comprehensive, reliable, and effective clinical assessment of the ability to predict early DR. The training procedure adopts a stratified data splitting approach that prevents data leakage, and hence the model will be able to generalize effectively to new patients. The dataset is split into training, validation, and testing sets based on patient-level separation, such that the temporal consistency is preserved for the longitudinal prediction task.

Table 2. Reproducibility settings

The training of the model is conducted under the supervision of multimodal inputs, where the image and clinical data are synchronized and optimized together. To avoid convergence to local minima and ensure stable convergence, early stopping and model checkpointing techniques are used. The training is conducted using mini-batch stochastic gradient descent with an adaptive learning rate schedule, which enables effective large-scale learning without compromising numerical stability. Model robustness is tested by cross-validation with the aim of reducing sampling bias, and statistically valid performance evaluation is ensured. Hyperparameter tuning is achieved by systematic searching techniques such as grid search and Bayesian optimization to figure out the optimal settings of the key parameters. Learning rates are dynamically changed with decay schedules, and both dropout and weight decay are implemented to increase the model's generalization. The transformer model parameters, such as the number of layers, the size of the hidden states, and the number of attention heads, are modified so as to increase the representational power and, at the same time, keep computational cost to a minimum. The calibration-related hyperparameters are adjusted to make sure that the risk predictions are probabilistically reliable, thus facilitating clinical interpretability and decision-making. The parameter settings are given in Table 2.

To prevent temporal leakage, patient-level separation was strictly enforced before dataset partitioning. All images belonging to a single patient were assigned exclusively to one subset. Confidence intervals were estimated using bootstrap resampling with 1000 iterations. Statistical significance between models was evaluated using paired DeLong testing for AUROC comparisons and McNemar tests for classification outcomes. Repeated experiments were performed using five independent random seeds and results are reported as mean ± standard deviation.

Baseline models have been implemented to make a fair comparison evaluation. A CNN baseline has been set up by using typical deep convolutional architectures for retinal image classification, thus representing the conventional deep learning approaches. A ViT baseline has been realized using patch-based tokenization and global self-attention mechanisms, thus representing state-of-the-art transformer models in medical imaging. Furthermore, a hybrid CNN–Transformer model has been constructed, which combines convolutional feature extraction with transformer-based global reasoning. These baselines facilitate comprehensive performance benchmarking and also prove the benefits of the proposed pillar-based causal-aware transformer architecture for early microvascular instability prediction.

All baseline models were trained using identical patient-level splits, preprocessing pipelines, optimization schedules, and evaluation protocols to ensure fair comparison. CNN, ViT, and Hybrid CNN-Transformer architectures were selected because they represent conventional convolutional approaches, pure transformer-based architectures, and hybrid feature-learning paradigms, respectively. These baselines collectively provide a representative benchmark for evaluating the contribution of pillar encoding and causal reasoning. For fairness, all baseline models received identical multimodal inputs consisting of fundus images, OCTA scans, and structured clinical variables whenever architecture compatibility permitted. Patient-level dataset partitions, preprocessing procedures, augmentation strategies, optimization schedules, and evaluation protocols were kept identical across all experiments. Therefore, observed performance differences primarily reflect architectural characteristics rather than differences in available information.

6.1 Predictive performance analysis

The predictive performance of the CAPT-DR framework displays strong discrimination power to locate the early DR risk as shown in Table 3. In Figure 7, one can see that the proposed model has obtained a large area under the ROC curve, which is indicative of the robustness of sensitivity–specificity trade-offs over the whole spectrum of decision thresholds. The deep ROC curve acquires the model's capacity to discern high-risk individuals correctly. Thus, it allows for keeping the number of false positives at a minimum. On the other hand, Figure 8 (a)-(g) illustrates the model's strength in regulating the imbalance of classes while maintaining high precision at the same time, even at recall levels, which is extremely important for early disease testing conditions. Figure 1 goes on to show that classification is quite evenly performed, the negative results are kept to a minimum, while the positives practically show the highest level of metamorphosis; thus, cynical early-stage risk identification is assured. The combination of these findings reveals that the CAPT-DR performs equally well in terms of the standard of accuracy and clinical reliability for microvascular instability prediction at an early stage.

Table 4 presents the ablation analysis of the proposed CAPT-DR framework by systematically removing key architectural components and evaluating their impact on predictive performance. The complete CAPT-DR model achieves the highest AUROC (0.934), AUPRC (0.911), and sensitivity (0.892) while maintaining the lowest Expected Calibration Error (ECE = 0.023). The removal of SCM, Pillar Encoding, MII, and Multimodal Fusion results in consistent performance degradation, demonstrating that each component contributes to predictive accuracy, calibration quality, and clinical reliability. The largest performance reduction is observed when multimodal fusion is removed, highlighting the importance of integrating retinal imaging and clinical information for early DR risk assessment.

Figure 7. Receiver Operating Characteristic (ROC) analysis for multimodal diabetic retinopathy (DR) risk prediction. Error bars and shaded regions represent 95% confidence intervals estimated using 1000 bootstrap resampling iterations

8a.png

8b.png

8c.png

8d.png

Table 3. Dataset characteristics

6.2 Comparative analysis

Table 4 presents the outcomes of our model in comparison with baseline models and shows the efficacy of the proposed framework. According to Figure 2, CAPT-DR has continuously outperformed the CNN, ViT, and hybrid CNN Transformer architectures in terms of all evaluation metrics, AUROC, sensitivity, specificity, and F1 score. Standard CNN models are capable of extracting local features but lack the ability to provide global reasoning; on the other hand, ViT models are able to grasp the overall context but do not have the structural coherence in the vascular system. Although hybrid models that integrate CNN-Transformer and conventional CNN have attained better performance, they are not able to preserve microvascular continuity. However, with pillar-based hierarchical encoding and causal-aware learning, CAPT-DR is able to give a more accurate depiction of retinal microvasculature and therefore result in significantly better predictive performance and robustness.

Table 4. Ablation analysis of CAPT-DR components for early diabetic retinopathy (DR) risk prediction

Note: CAPT-DR = Causal-Aware Pillar Transformer-Diabetic Retinopathy (CAPT-DR); DR = Diabetic Retinopathy; SCM = Structural Causal Model; MII = Microvascular Instability Index; AUROC = Area Under the Receiver Operating Characteristic Curve; AUPRC = Area Under the Precision–Recall Curve; ECE = Expected Calibration Error.