In mental health risk assessment and monitoring, systems are deployed in dynamic and evolving scenarios. In many real world scenarios, diverse signals (clinical notes, social media text, speech, and wearable biosignals) are used to detect risk states such as depression, anxiety, self-harm intent, and acute stress. However, most pipelines are trained on static datasets with fixed labels and vocabularies. When new expressions, populations, or risk patterns appear, performance is often degraded. In practice, full retraining on new data is often required, while incremental updates risk catastrophic forgetting that overwrites previously learned knowledge [13]. Therefore, continual learning is increasingly regarded as important for mental health tasks. Rather than retraining from scratch, new risk categories or domains can be incorporated as they arise while preserving recognition of earlier ones [14].

Despite rapid progress, Several unresolved challenges continue to hinder dependable mental-health risk detection from text:

Missing causal structure: Most models treat symptoms as flat labels; they do not encode directed tag→tag influences or use them during message passing.

In light of the above, we propose a continual neuro-symbolic framework that builds per-post symbolic tag graphs with directed links, encodes them via a two-layer GCN, and uses a lightweight attention head to form calibrated, interpretable post representations. For sequential datasets, we adopt a frozen-encoder, multi-head protocol to prevent forgetting while keeping adaptation lightweight. The key contributions of this work are summarized below:

A neuro-symbolic graph learning framework is proposed that combines symbolic reasoning, causal inference, graph-based neural encoding, and continual learning for mental-health risk assessment.

The remainder of this paper is organized as follows: Section 2 presents an extensive survey of related research and methods relevant to this work, Section 3 presents the proposed CNSGL framework, including symbolic graph construction, causal enrichment, the GCN–Transformer encoder, and the MH-Freeze continual-learning strategy. Section 4 describes the experimental details, and evaluation metrics. Section 5 presents the ablation study analyzing the contribution of each component. Section 6 concludes the paper and highlights future directions.

This subsection reviews key efforts on detecting mental-health signals from social-media text, ranging from traditional machine learning (ML) to deep learning (DL). Hemmatirad et al. [15] showed that lexicon and handcrafted features paired with support vector machine or logistic regression classifiers can distinguish high-risk users using linguistic and emotional cues. With the advent of contextual embeddings, hybrid models such as BERT+BiLSTM have been proposed by Zhou and Mohd [16] to better handle informal language, emojis, and sequential patterns in depression-related posts.

Prior surveys consolidate the literature and shed light on persistent challenges. Garg [17] reviewed 92 studies, introduced an updatable suicide-detection repository, and emphasized the need for real-time, responsible AI. Skaik and Inkpen [18] surveyed NLP/ML approaches for public mental-health surveillance, summarizing data collection strategies, modeling tools, and remaining gaps. Other studies explore emotion-aware and efficiency-focused systems. Benrouba and Boudour [19] proposed an emotion-aware content-filtering framework that classifies posts into basic emotions and compares them with an “ideal” lexicon to flag potentially harmful content. Ding et al. [20] compared ML models (logistic regression, random forest, LightGBM) with DL models (ALBERT, GRU) for binary and multi-class mental-health classification, finding that ML methods offer better interpretability and efficiency on medium-sized datasets, whereas DL models better capture complex linguistic patterns.

Causal reasoning in language means modeling directional influence (A→B) so that changing A would change the likelihood of B, beyond simple correlation. This has been studied using temporal precedence and directional association measures, causal discovery on event graphs, and counterfactual analyses. However, in social-media risk assessment, such causal structure is rarely embedded within the encoder itself. Choudhury and Kiciman [21] examined the causal impact of online social-support language in Reddit mental-health communities on future suicidal-ideation risk. Using human assessments within a stratified propensity-score framework to form comparable cohorts, they estimated treatment effects of support types and found that esteem and network support significantly reduce subsequent risk, with implications for tools that enhance support provision.

Zhang et al. [22] proposed a causal framework based on a counterfactual neural temporal point process (TPP) to estimate the individual treatment effect (ITE) of misinformation on user beliefs and actions at scale, using a neural TPP with Gaussian mixtures for efficient inference. Experiments on synthetic data and a real COVID-19 vaccine dataset showed identifiable causal effects of misinformation, including negative shifts in users’ vaccine-related sentiments. Cheng et al. [23] surveyed Event Causality Identification (ECI) and proposed a systematic taxonomy split into sentence-level ECI (SECI) and document-level ECI (DECI) tasks, reviewing approaches from feature/ML methods to deep semantic encoding, event-graph reasoning, and prompt/causal-knowledge pretraining, with notes on multilingual, cross-lingual, and zero-shot large language model settings.

GCN encoders capture relational structure for text via message passing on graphs, benefiting settings with explicit concept relations [24]. Hamilton et al. [25] proposed GraphSAGE, an inductive framework that extended GCNs to unsupervised learning and introduced trainable aggregation functions beyond simple convolutions. The method generated embeddings for unseen nodes by sampling and aggregating neighborhood features, leveraging node attributes for generalization. Yao et al. [26] introduced Text GCN, which built a corpus-level graph from word co-occurrence and document-word relations to jointly learn word and document embeddings. Without relying on external embeddings, Text GCN outperformed state-of-the-art methods on multiple benchmarks and showed strong robustness with limited training data.

Transformers, driven by self-attention, excel at weighting inputs and can be used as interpretable pooling over concept embeddings [27]. Vaswani et al. [28] proposed the Transformer, a sequence transduction architecture based solely on attention mechanisms, removing recurrence and convolutions. The model achieved state-of-the-art results on Workshop on Machine Translation 2014 English to German and English to French translation tasks, while being more parallelizable and significantly faster to train than prior approaches.

Devlin et al. [29] introduced Bidirectional Encoder Representations from Transformers (BERT), a bidirectional Transformer-based model pre-trained on unlabeled text by jointly conditioning on left and right context. With simple fine-tuning, BERT achieved state-of-the-art results on eleven NLP tasks. Yang et al. [30] proposed a hierarchical attention network (HAN) for document classification, which reflected the hierarchical structure of documents and applied attention at both word and sentence levels. The model outperformed prior methods on six large-scale benchmarks and provided interpretable document representations by highlighting informative words and sentences.

Continual learning addresses time-varying, patient-specific data in mental-health scenarios by incrementally updating models from electronic health records, speech/text, and wearable streams while preserving prior knowledge. Although the CL for mental health literature remains limited, we highlight a few representative systems that show feasibility under practical constraints. Gamel and Talaat [31] proposed SleepSmart, an Internet of Things (IoT) enabled continual learning framework for intelligent sleep enhancement. The system employed wearable biosensors to capture physiological signals during sleep, which were processed via an IoT platform to deliver personalized recommendations. By leveraging continual learning, SleepSmart improved recommendation accuracy over time, and a pilot study demonstrated its effectiveness in enhancing sleep quality and reducing disturbances.

Lee and Lee [32] explored the role of continual learning in medicine, where models adapt to new patient data without forgetting prior knowledge. They emphasized challenges such as catastrophic forgetting and regulatory constraints, but argued that continual learning offers advantages over non-adaptive Food and Drug Administration approved systems by incrementally improving diagnostic and decision-support performance. Li and Jha [33] proposed DOCTOR, a continual-learning framework for multi-disease detection on wearable medical sensors at the edge. The system used a multi-headed deep neural network with replay-based CL, via exemplar data preservation or synthetic data generation to mitigate catastrophic forgetting while sequentially adding tasks with new classes and distributions. In experiments, a single model maintained high accuracy, yielding up to 43% higher test accuracy, 25% higher F1 score, and 0.41 higher backward transfer over naive fine-tuning.

A structured comparison is presented in Table 1 to more clearly contextualize CNSGL within existing work. Prior methods typically incorporate only one or two of the components, symbolic representations, causal edge modeling, graph-based encoders, Transformer attention mechanisms, or continual-learning strategies, rather than unifying all of them within a single framework. As shown in Table 1, approaches that employ symbolic reasoning rarely integrate GNN encoders or explicit causal edge construction; causal GNN models generally do not use symbolic tag vocabularies or Transformer-based pooling; and continual-learning systems commonly operate without symbolic graphs or causal modeling. In contrast, CNSGL combines directional PMI-derived causal edges, a symbolic tag graph, a two-layer GCN encoder, a Transformer attention pooler, and a multi-head freeze continual-learning strategy within one architecture tailored for mental-health risk detection. This integrated design forms the central novelty of the approach and demonstrates how CNSGL extends beyond existing component-wise methods.

Mental-health text from online platforms or clinical records is largely unstructured and often contains implicit cues about psychological conditions that are difficult to analyze directly. To impose structure and enhance interpretability, each post is represented as a symbolic graph that captures both semantic content and clinically meaningful indicators. A compact set of ten symbolic tags, sleep, anxiety, depression, stress, anger, lonely, health, fear, coping, and suicidal, was constructed based on well-established constructs in computational mental-health research and their frequent annotation in benchmark datasets. The vocabulary was further validated through manual inspection and an expert-informed review to ensure clinical relevance. A small, consistent set was intentionally maintained to minimize noise and support stable directional PMI estimation during causal graph construction. Let 𝒟 denote the dataset of posts, and let di∈𝒟 denote a post in this collection. A vocabulary 𝒱 of risk-related terms, including hopelessness, insomnia, and self-harm, is predefined. Using this lexicon, a set of symbolic tags is assigned to the post, τi={t1,t2,…,tm},tj∈𝒱,where each tj denotes a tag identified in di. These tags provide explicit signals of potential risk factors, linking the unstructured narrative of a post to interpretable constructs grounded in psychology.

The symbolic graph is defined as Gi=(Vi,Ei,Xi) consisting of:

Nodes: Vi={v0}∪{vj:tj∈τi}, where v0 represents the post and vj denote the symbolic tags.

At this stage, Gi uses undirected post-tag links (implemented as bidirectional pairs) to encode association rather than causality. Directional tag–tag edges are introduced in the subsequent Causal-Aware Graph Enrichment step, where order-sensitive statistics are is used to determine edge orientation.

Edges based solely on co-occurrence capture statistical associations but cannot distinguish whether one factor precedes or influences another. For example, sleep deprivation may frequently appear with stress, yet in many cases it precedes and contributes to suicidal ideation. To incorporate such directional relationships, graphs are enriched with causal edges in addition to co-occurrence links. To quantify how strongly tags co-occur in the input space, point-wise mutual information (PMI) is used here. Consider the tags sleep and anxiety. These tags may frequently appear together in posts, which would yield a symmetric, undirected edge in a standard co-occurrence graph. Directional PMI instead focuses on ordered pairs and estimates whether one tag is more likely to appear before the other. If ordered counts show that mentions of sleep problems systematically precede anxiety indicators more often than the reverse, then PMIdir(sleep→anxiety) exceeds PMIdir(anxiety→sleep), and the graph includes the directed edge sleep → anxiety. In this way, directional PMI encodes asymmetric, precedence-aware relationships that cannot be represented by undirected co-occurrence edges alone. PMI between two given tags ta and tb is calculated in Eq. (1) [45]: (1)PMI(ta,tb)=log⁡P(ta,tb)P(ta)P(tb)

Probabilities were estimated by counting how often each tag and each tag pair appeared within the same post and then normalizing by the total. A small smoothing constant was applied so that rare tags didn’t get zero probability. Because PMI treats a pair the same in either order, it captures association only and does not encode direction. Let c(ta≺tb) be the number of posts in which ta occurs before tb, and let Npairs be the total number of ordered tag pairs considered. Eq. (2) shows the mathematical definition of the directional PMI: (2)PMIdir(ta→tb)=log⁡P(ta≺tb)P(ta)P(tb),where, P(ta≺tb)≈c(ta≺tb)Npairs is the probability of ordered event. A directional PMI threshold δ is applied to filter out weak or noisy associations. In practice, δ was selected through validation by sweeping values in {0.0,0.05,0.1,0.2} and choosing the smallest threshold that removed spurious edges while preserving clinically meaningful relations. The framework was most stable for δ=0.1, which we adopt for all experiments. Sensitivity analysis showed that the graph structure remained consistent for thresholds within a small neighborhood (0.05–0.15), indicating that results are not overly sensitive to the exact choice of δ.

A directed edge, ta→tb is introduced whenever PMIdir(ta→tb)>δ. This indicates that the occurrence of ta increases the likelihood of tb, suggesting a causal tendency rather than a mere correlation. After adding the causality factor, the resulting enriched graph becomes: Gi=(Vi,Eico∪Eicausal,Xi),where Eico denotes post–tag associations, and Eicausal represents directed tag–tag relations. For example: If mentions of sleep problems are frequently followed by references to anxiety, and the directional PMI exceeds the threshold δ, an edge vsleep→vanxiety is added. In this way, the representation allows the model to reason not only about what terms appear together, but also about which factors may act as potential precursors of others. During the execution of the GCN, the directed causal edges are converted into an undirected form (with self-loops) so that message passing remains symmetric. Throughout this work, the directed edges are treated as precedence-aware statistical associations rather than definitive cause–effect links; accordingly, the term “causal” is used in an operational sense to describe directional risk-tendency patterns observed in mental-health text.

The proposed CNSGL framework is depicted as an end-to-end architecture in Fig. 1, emphasizing the left-to-right progression from symbolic/causal structuring to representation learning and, then to sequential adaptation. The diagram marks where the shared encoder is frozen and where dataset-specific heads are attached, making clear how prior knowledge is preserved while new tasks are added. The Transformer Encoder used for graph-level pooling in the encoding stage is described separately in Fig. 2.

The symbolic graphs enriched with causal relations are processed by a two-layer GCN. The GCN propagates information across connected nodes so that each representation reflects both its own features and those of neighboring nodes. In this way, a post node aggregates signals from its tags, while tag nodes incorporate both symbolic and causal context. Eq. (3) shows how the node representations are updated at layer l [46]. (3)H(l+1) = σ(D~−12A~D~−12H(l)W(l)),where A~=A+I is the adjacency with self-loops, D~ is its degree matrix, W(l) are trainable weights, and σ is an element-wise ReLU activation function. The input is H(0)=Xi, the node-feature matrix of graph Gi. Since causal edges are directed, their weights are first assembled in a directed matrix W and then symmetrized to form A (e.g., A=12(W+W⊤)) prior to normalization.

Although the final GCN uses a symmetrized adjacency matrix for stable message passing, the causal interpretability of the framework is retained because symmetrization occurs only after causal tendencies have been encoded in the edge-selection stage. Directional PMI determines which tag pairs are connected and the strength of those connections, thereby shaping the underlying causal structure even if the GCN operates on an undirected form of the graph. The interpretability comes from this directed edge construction and from the subsequent analysis of causal paths and attention weights, whereas symmetrization serves primarily as a computational requirement of the canonical GCN rather than a removal of causal information. After two layers, the node embeddings H(2) are passed to a learned classification token [CLS] as a global query over the tag nodes to produce a graph-level representation zi with tag-level importances. This pooled vector zi∈Rdz encodes symbolic, semantic, and causal structure and is used for classification.

While mean pooling provides a simple mechanism for aggregating node embeddings into a graph-level representation, it treats all nodes equally and fails to highlight which risk factors are more influential in a particular post. To address this limitation, the node embeddings produced by the GCN are passed through a Transformer encoder to perform attention-based pooling [47]. Fig. 2 illustrates the Transformer encoder block employed as an attention pooler over GCN-derived node embeddings. The [CLS] token attends to tag embeddings to produce a pooled representation while simultaneously providing interpretable tag-level importance scores through attention weights.

Let H(2)=[h1,…,h|Vi|] be the node embeddings (post + tags) for post i after the 2-layer GCN. The input sequence formed is presented in Eq. (4): (4)X=[CLS;H(2)]where [CLS] is a learnable pooling token. The encoder computes linear projections Q=XWQ,K=XWK,V=XWV,with trainable WQ,WK,WV and per-head key dimension dk. We apply a pooler mask so that only the [CLS] row of Q issues queries (tag↔tag attention is masked). Let qCLS denote the [CLS] query. The attention weights from [CLS] to all tokens are calculated in Eq. (5): (5)α=softmax(qCLSK⊤/dk)

These are then used to form a weighted sum of the values for [CLS] (computed per head, concatenated, and projected). Each encoder block applies Add&LayerNorm around multi-head attention and a Feed-Forward Network (FFN). The final [CLS] vector is taken as the pooled embedding zi∈Rd.

The weights α (averaged over heads) serve as tag importances, providing a transparent summary of which tags influenced the decision. The pooled vector zi is passed to a linear sigmoid head (pi) to obtain the risk label, explained in the subsequent section in detail. The attention pooler improves aggregation of symbolic and causal information while preserving interpretability via tag-level weights.

The pooling module is implemented as a single Transformer-style encoder block with multi-head self-attention and a position-wise feed-forward network (FFN). In practice, the CLS attention Pooler uses an embedding dimension of d=128, four attention heads, and an FFN hidden size of 256 with dropout 0.1. A learned [CLS] token is prepended to the node embeddings, and its final output vector is used as the graph-level representation, while the CLS-to-tag attention weights provide interpretable tag importances. The same configuration is used across datasets to maintain consistency in the continual-learning setup.

To illustrate how the Transformer attention pooler provides qualitative interpretability, two example posts are shown below. In each case, the model highlights the most influential tags and their causal relations when producing a risk prediction.

Example 1 (High-risk post): “I have not slept properly for days, and the constant anxiety is making everything feel overwhelming. Lately I keep thinking that things would be easier if I just disappeared.” The attention pooler assigns high importance to the tags sleep, anxiety, and suicidal, with a strong causal pathway sleep→ anxiety→suicidal. These attended tags correspond to clinically salient risk indicators, leading the classifier to assign a high-risk label.

Example 2 (Low-risk post): “Feeling a bit stressed about exams next week, but talking to friends has helped and I’m trying to stay positive.” The model focuses primarily on stress, with low attention weights on other tags and no causal escalation toward depression or suicidal. The attention pattern reflects a non-escalatory emotional state, leading to a low-risk prediction.

In real-world applications, data arrive in stages {𝒯1,…,𝒯K} with evolving language, populations, and even label definitions. Sequential training on 𝒯k risks catastrophic forgetting of knowledge learned on earlier datasets. Continual learning is incorporated to address the sequential arrival of mental-health datasets and the risk of catastrophic forgetting. This ensured that knowledge acquired from earlier domains was preserved while adaptation to new sources was achieved, thereby enhancing robustness and practical applicability of the framework. A simple, effective continual learning technique, MH-Freeze is used, which preserves a shared encoder while adding a small task-specific head per dataset. Let Gi be a post graph and let fθ denote the encoder mapping Gi to a pooled embedding. The pooled embedding is computed in Eq. (6). (6)zi = fθ(Gi)∈Rdwhere fθ comprises causal-aware graph construction, two GCN layers, and the lightweight Transformer encoder. Thus zi integrates symbolic structure, causal links, and attention-based tag weighting.

For dataset 𝒯k={(Gi(k),yi(k))} with a single binary risk label y∈{0,1}, the task head gϕk is a linear–sigmoid classifier unit that produces a probability p^i(k) = gϕk(zi) = σ(wk⊤zi+bk) ∈ [0,1],with parameters ϕk={wk∈Rd,bk∈R} and sigmoid σ(⋅). A threshold τk (calibrated on a small validation split) yields the binary decision. At inference, predictions are labeled as positive if p≥τ.

MH-Freeze

At first, the shared encoder fθ and the initial linear–sigmoid head gϕ1 are jointly trained on 𝒯1 by minimizing the average BCE loss, which is the negative log-likelihood of a Bernoulli target and directly trains calibrated probabilities p^∈[0,1] from the sigmoid head. Eq. (7) presents formulation of BCE for a single label y∈{0,1} and predicted probability p^∈[0,1] is [48]: (7)BCE(y,p^)=−[ylog⁡p^ + (1−y)log⁡(1−p^)].

The shared encoder fθ and the initial linear–sigmoid head gϕ1 are then jointly trained on 𝒯1 by minimizing the average BCE, as shown in Eq. (8): minθ,ϕ1 1|𝒯1|∑(G,y)∈𝒯1BCE(y, p^(1)(G)), (8)p^(1)(G)=σ(w1⊤fθ(G)+b1).

After convergence, the encoder parameters θ are frozen, so that zi=fθ(Gi) provides a fixed representation for all subsequent heads. Freezing is applied after Task 1 because the first dataset provides the broadest and most diverse distribution of symbolic tags, allowing the encoder to learn generalizable representations before domain-specific heads are introduced. We also examined variants where the encoder is frozen after Task 2 or Task 3. These alternatives showed higher forgetting on earlier datasets and reduced overall stability, as the encoder continued adapting toward the later-task distributions and drifted away from the symbolic-causal structure learned initially. Freezing after Task 1 therefore offered the best balance between preserving prior knowledge and supporting effective multi-head adaptation.

For each subsequent dataset 𝒯k (k≥2), a lightweight head is instantiated and only its parameters are optimized while keeping the encoder fixed, as mentioned in Eq. (9): minϕk 1|𝒯k|∑(G,y)∈𝒯kBCE(y, p^(k)(G)), (9)p^(k)(G)=σ(wk⊤fθ(G)+bk),∇θℒk=0.

As θ is fixed, adaptation reduces to fitting task-specific linear separators in the common embedding space z, which avoids cross-task interference and sharply limits forgetting. Intuitively, the shared encoder captures domain-general structure (symbolic and causal relations plus attention-based tag weighting), while each head accounts for dataset-specific prevalence, wording, or scope. Given dataset, the corresponding head gϕk is selected and its probability p^(k) is thresholded to yield the label: risky ifp^(k)≥τk,non-risky otherwise.

To evaluate the effectiveness and generalizability of the proposed Continual Neuro-Symbolic Graph Learning framework, six diverse datasets were used spanning Reddit-based mental health discourse and clinician-guided interviews. A brief summary is provided in Table 2. These corpora differ in annotation protocols, linguistic style, and risk indicators, enabling a comprehensive assessment of both the symbolic reasoning components and the graph-based learning modules.

Data Analytics for Smart Health (DASH-2020) [49]: It consists of reddit posts annotated for substance use, addiction, and recovery. For our binary setup, we merged all recovery-related categories into a single non-addicted class, while posts explicitly labeled as addicted are retained as the positive class.

The proposed MH-Freeze framework is compared with several representative continual learning techniques, each reflecting a distinct strategy for mitigating catastrophic forgetting in sequential task scenarios.

Elastic Weight Consolidation (EWC) [56]: EWC addresses catastrophic forgetting in sequential learning by estimating the importance of each parameter for previously learned tasks (via a Fisher-based approximation) and selectively slowing changes to those important weights when learning a new task. This preserves prior expertise while allowing plasticity on less critical parameters.

The effectiveness of the proposed framework and the baseline continual learning techniques is evaluated using multiple performance metrics. These metrics capture not only classification accuracy but also robustness under class imbalance and calibration of probabilistic outputs.

The proposed MH-Freeze framework demonstrates strong continual-learning behavior across heterogeneous datasets as seen in Table 3. MH-Freeze performs continual learning by freezing a shared GCN-Transformer encoder and training lightweight, task-specific heads as new tasks arrive. MH-Freeze is compared against six continual-learning baselines across six tasks, where each task corresponds to a different dataset in a fixed sequential order. The six tasks correspond to distinct datasets: T1 = DASH, T2 = Dreaddit, T3 = SWMH, T4 = Go_emotions, T5 = DAIC-WOZ, and T6 = Kaggle-MH. For brevity and consistency, these datasets are referenced as T1-T6 throughout the remainder of the paper. In all experiments, tasks are encountered in the fixed order T1-T6; balanced mini-batches are used to counter dataset-level imbalance, and the MH-Freeze architecture prevents dominance by any single dataset due to differing label distributions. The discriminative capability remains uniformly high, with AUPRC ≥ 0.934 and AUROC ≥ 0.942 throughout the task sequence, indicating robust separability between risk and non-risk classes. Both Accuracy (0.898 to 0.925) and F1-Score (0.886 to 0.923) follow a steady upward trajectory from T1 (DASH) to T6 (Kaggle-MH), reflecting positive forward transfer without degradation of earlier competencies. The MCC also improves from 0.829 to 0.873, confirming balanced predictive behavior under label imbalance. In parallel, Brier and ECE scores decline from 0.069 to 0.060 and 0.023 to 0.014, respectively, demonstrating progressive improvement in probability calibration. These metrics affirm that MH-Freeze effectively preserves prior knowledge while adapting to new domains, achieving well-calibrated, generalizable predictions with minimal catastrophic forgetting. It maintains a clear advantage in both discrimination and calibration.

Among the CL baselines used for comparison, GEM performs better, attaining moderately high AUROC (0.94–0.95) and balanced MCC values, though it gains plateau beyond mid-sequence tasks. LWF exhibits reasonable F1-scores but suffers from high calibration error and inconsistent reliability across datasets. Experience Replay and Synaptic Intelligence provide stable yet lower performance, with AUROC typically below 0.91 and limited robustness to domain shifts. EWC achieves comparable mid-range results but shows greater sensitivity to task transitions, while Finetuning performs worst overall, displaying rapid accuracy decay (0.73–0.76) and high Brier/ECE values indicative of severe forgetting. In contrast, MH-Freeze sustains near-optimal metrics across all six tasks, confirming that its frozen encoder with task-specific heads yields superior retention, adaptation, and calibration in continual-learning environments. The detailed results for each continual-learning baseline are presented in Tables 4–9, providing a comprehensive comparison across all tasks. Experience Replay and Synaptic Intelligence show early performance saturation, as seen in Tables 6 and 7, respectively. This likely reflects limited forward transfer and calibration instability under domain shift. A plausible cause is that ER’s small replay buffer cannot adequately represent later datasets, while SI’s weight-importance penalty restricts the flexibility needed to adapt. Consistently higher Brier and ECE on the final tasks (T4-T6) reinforce this interpretation, indicating less reliable probabilities and weaker calibration as the data distribution changes. While MH-Freeze exhibits a monotonic increase in MCC from T1 to T6, GEM stabilizes at a slightly lower range (Table 4). Brier and ECE generally decline for MH-Freeze, indicating progressively better calibration, whereas replay and regularization-based baselines (ER, EWC, SI) show smaller or inconsistent reductions.

Methods that incorporate explicit memory mechanisms or parameter regularization, such as GEM and EWC, demonstrate better retention than Finetuning or LWF across all six tasks, confirming that constraining weight drift mitigates forgetting. However, these approaches still exhibit limited calibration stability, as indicated by elevated Brier and ECE values across late tasks. Synaptic Intelligence achieves moderate balance between accuracy and calibration, but its adaptation saturates beyond mid-sequence datasets, revealing difficulty in scaling to domain shifts. In contrast, MH-Freeze consistently maintains high discriminative accuracy while achieving the lowest calibration errors.

From T1 (DASH) to T6 (Kaggle-MH), most baselines show mild fluctuations in F1-Score and AUROC due to changing dataset characteristics and label imbalance. However, MH-Freeze exhibits smooth performance progression, achieving improvements in accuracy and F1-Score compared with the strongest baseline (GEM). This trend demonstrates strong forward transfer and minimal backward interference. Moreover, the consistently low Brier (≈0.06) and ECE (≈0.014) values emphasize its reliability in producing well-calibrated probabilities, critical in sensitive applications such as mental-health risk prediction, where overconfident misclassifications can have severe consequences.

Fig. 3 shows accuracy trends for all methods across the six datasets in sequence. The accuracy curves show a clear and persistent margin for the proposed MH-Freeze approach on every task. From T1 to T6, the accuracy of MH-Freeze rises from 0.898 to 0.925 (an increase of 3.0%), which indicates that knowledge gained on earlier tasks is retained while useful information from later tasks is added. In contrast, Finetuning changes only slightly (0.725 to 0.731; 0.8% increase) and shows signs of forgetting as new tasks are introduced. Regularization methods yield smaller gains, EWC improves by 2.1% and SI decreases by 1.7%, suggesting limited ability to adapt to the domain shifts in this sequence. Replay (ER), distillation (LWF), and gradient-projection (GEM) make training more stable, but they still perform worse than the proposed method. The highest final accuracy of 0.925 and the steady rise across datasets suggest that freezing the shared GCN-Transformer encoder and using separate heads for each dataset reduces interference and supports reliable forward transfer.

Similarly, a comparison based on F1-Score is presented in Fig. 4. This metric is informative under class imbalance and helps assess whether decisions remain balanced as new datasets are introduced. The F1-Score of MH-Freeze increases from 0.886 at T1 to 0.923 at T6, a 4.2% increase, indicating that earlier decision boundaries are preserved while new patterns are learned. The baselines follow a consistent ordering: Finetuning decreases from 0.775 to 0.722 (a 6.8% decrease), reflecting forgetting; EWC shows a small improvement of 1.5%; SI decreases slightly by 0.7%; ER and LWF achieve moderate increases of 2.0% and 4.0%, respectively; and GEM improves by 3.1% but remains below MH-Freeze. These results suggest that the multi-head freeze strategy maintains balanced predictions across datasets while enabling steady gains, which is the intended behavior in a continual-learning environment.

In addition, ROC curves are presented to provide an intuitive visualization of classification trade-offs across varying decision thresholds. Unlike single-value metrics, ROC curves reveal how each model balances the true-positive and false-positive rates, offering a deeper understanding of their discriminative behavior. The ROC curves shown in Fig. 5 illustrates the comparative classification performance of all continual-learning baselines and the proposed MH-Freeze model across six sequential tasks (T1–T6). Each subplot corresponds to a specific task, where the x-axis represents the False Positive Rate (FPR) and the y-axis represents the True Positive Rate (TPR). The ROC trajectories of all models are plotted within each panel, while only the area under the ROC curve (AUC) of the MH-Freeze model, representing the top-performing method, is explicitly annotated. The solid black curve corresponds to MH-Freeze and demonstrates its consistently superior separability across all tasks. In contrast, baseline models such as Finetuning, EWC, and SI are depicted with thinner colored curves to provide visual benchmarking and highlight relative performance differences. This visualization clearly demonstrates that MH-Freeze maintains stable and near-optimal discriminative capability across all incremental tasks, confirming its strong resistance to catastrophic forgetting and enhanced adaptability in continual-learning environments.

The CNSGL framework is designed to remain computationally efficient while still using the same encoder as the continual-learning baselines. The shared encoder, comprising a two-layer GCN (hidden size 128) and a single Transformer attention block with four heads (dmodel=128, FFN size 256), contains approximately 482 k learnable parameters, which are used in all models, including the proposed MH-Freeze. Each task also has a linear–sigmoid classifier head with 129 parameters, giving a total of about 482 k+129 parameters per model. The major difference is not in how many parameters exist, but in how many are updated during each new task. In the proposed framework, MH-Freeze freezes the encoder after first task (T1), and trains only the 129-parameter head, whereas baseline models continue to update all 482 k+129 parameters for every new task. Training on T1, where the encoder and head are jointly optimized, takes about 19 min, while training on subsequent tasks (T2–T6), where only the head is updated, completes in roughly 2 to 4 min per task. A single forward pass requires approximately 1.47 ms per post and about 1.37 million floating point operations per second (FLOPs).

All continual-learning baselines use the same encoder architecture for a fair comparison, so their inference FLOPs are identical to CNSGL. However, they differ substantially in how many parameters are updated during each new task and in the resulting training time. CNSGL (MH-Freeze) updates only 129 parameters per new task, yielding the lowest per-task training time, whereas baselines must update the full encoder (≈ 482 k+129 parameters), leading to longer training times despite identical inference complexity.