1)PsychoGraph-Net architecture:The first end-to-end, jointly trained graph neural network system designed specifically for psychiatric comorbidity analysis in liaison psychiatry, integrating graph construction, community detection, attentive message-passing, and Bayesian uncertainty in a single framework.

2)Comorbidity graph at scale:The largest liaison psychiatry comorbidity network published to date, comprising 89 ICD-10 diagnostic nodes and 1247 significant co-occurrence edges derived from 14,287 hospital admissions over six years.

3)Three novel comorbidity clusters:Community detection reveals three previously uncharacterized diagnostic clusters, the Affective-Anxiety-Sleep cluster, the Neurocognitive-Delirium complex, and the Somatic-Metabolic interface with direct implications for integrated care pathway design.

4)Primary prediction endpoint:Five-class mood-state classification (Euthymia, Depression, Hypomania, Mania, Mixed Features) on held-out test admissions, achieving AUC-ROC of 0.951 and macro-F1 of 0.887. Seven-day episode-onset prediction is evaluated as a secondary endpoint. Comorbidity structure is the input representation, not itself a prediction target.

5)Bayesian clinical integration:ECE of 0.031 and a selective prediction protocol routing 14.7% of high-uncertainty cases to clinician review, with a validated 31.4% reduction in diagnostic workload in clinical simulation.

Liaison psychiatry, the subspecialty managing psychiatric conditions arising in the context of physical illness and general hospital admission [11] confronts a distinctively complex clinical reality. Approximately 20% - 30% of general hospital inpatients have a co-occurring psychiatric condition that meaningfully affects their medical course, and systematic psychiatric consultation is associated with shorter length of stay, reduced readmission, and better physical health outcomes [12][13].

The diagnostic complexity of this population is distinctive. Delirium, one of the most common psychiatric presentations in hospital medicine, frequently overlaps with dementia and with the confusional states produced by infection, metabolic derangement, and pharmacological side effects [14]. Mood disorders in medically ill patients are complicated by somatic symptoms that mimic depression and by the bidirectional relationship between cardiovascular disease and affective illness [15]. In this environment, understanding how diagnoses co-occur is not a theoretical exercise it is a practical necessity for safe and effective care.

The application of network science to medical diagnosis established the foundational insight that diseases sharing genetic underpinnings form phenotypically coherent communities most influentially demonstrated in the Human Disease Network. Subsequent analysis showed that diseases are not independent entities but members of interconnected communities whose structure encodes biologically and clinically meaningful relationships [2][3].

In psychiatry specifically, network analysis has been applied most productively at the symptom level. The network theory of mental disorders proposes that psychiatric conditions emerge from causal interactions among symptoms rather than from latent disease entities. Empirical work demonstrated that depression and anxiety share densely connected symptom clusters [16][17] and that the network structure of symptoms predicts treatment response [18]. At the diagnosis level, temporal comorbidity trajectory analysis using large EHR databases [19] has revealed that psychiatric diagnoses frequently serve as bridge nodes connecting otherwise separate disease communities, but this descriptive work has not been connected to a predictive machine learning pipeline for individual patients.

Graph neural networks [20] operate directly on graph-structured data, learning node representations by aggregating information from neighbourhoods through differentiable message-passing. They have been applied to drug-drug interaction prediction, patient similarity networks [21], and protein-disease association mining [22]. Graph Attention Networks specifically extend standard GNNs by learning to attend differentially to different neighbours an especially valuable property in comorbidity graphs, where some co-occurring diagnoses are far more predictive of a given target than others. No prior work has applied GATs to psychiatric comorbidity networks in liaison psychiatry.

Reliable uncertainty estimation is not optional in clinical AI. A model that is 90% accurate but systematically overconfident causes greater clinical harm than an 85% accurate, well-calibrated model particularly in psychiatric settings where false alarms trigger unnecessary interventions and missed cases lead to untreated illness [23]. Monte Carlo Dropout [24] and deep ensembles [25] provide principled posterior uncertainty estimation; ensembles have been shown to produce superior calibration relative to single-model approaches. The EU AI Act (Regulation EU 2024/1689) [26] requires high-risk AI systems to provide transparent uncertainty information making Bayesian calibration legally necessary, not merely scientifically desirable.

The dataset comprises 14,287 de-identified inpatient admissions from a tertiary academic medical centre with an active liaison psychiatry consultation service, spanning January 2018 to December 2023. The institution operates a fully integrated EHR system, with ICD-10-coded diagnosis data captured for every admission by both the primary medical team and the liaison psychiatry consultation team. Ethics approval was obtained (IRB Ref. UNIGE-2023-LPS-07); all data were processed in compliance with GDPR Article 9. Patient consent was waived under the retrospective research exemption.

Inclusion criteria: adult inpatients (age ≥ 18 years) with at least one liaison psychiatry consultation, at least five distinct ICD-10 diagnostic codes at the three-character category level (e.g., F32, I10, E11 counted as three distinct codes), and length of stay exceeding 24 hours. After exclusions (8.3% of records), the final cohort comprised 14,287 admissions from 9841 unique patients, with a mean of 3.4 ± 1.7 diagnoses per admission and the details will be found in Table 1. The apparent discrepancy between the inclusion threshold (≥5 codes) and the reported mean (3.4 ± 1.7) reflects two distinct counting conventions: the inclusion criterion counts all three-character ICD-10 codes recorded across the full administrative record for each admission, while the reported mean counts only the ICD-10 diagnostic categories that met the minimum prevalence threshold (≥ 50 admissions) and were therefore included as nodes in the comorbidity graph. Diagnoses failing this threshold were excluded from graph construction but did not trigger patient exclusion.

Table 1.Cohort sociodemographic and clinical characteristics.

The primary prediction target is a five-class mood state label assigned to each admission by the liaison psychiatry team using DSM-5 criteria, corroborated by YMRS (mania/hypomania) and HAMD-17 (depression) structured assessments documented in the EHR. Labels were assigned independently of the treating medical team and recorded at the time of psychiatric consultation. Definitions: Euthymia - YMRS ≤ 7 and HAMD-17 ≤ 7, no active affective episode; Depression - HAMD-17 ≥ 14 with clinical interview confirmation; Hypomania - YMRS 8 - 19, duration ≥ 4 days, no functional impairment requiring hospitalization; Mania - YMRS ≥ 20 or YMRS 8 - 19 with hospitalization; Mixed Features - concurrent YMRS ≥ 8 and HAMD-17 ≥ 14. The secondary endpoint, 7-day episode onset, is a binary label indicating transition from Euthymia to any non-euthymic state within the subsequent 7 days, derived from consecutive admission records for patients with ≥ 2 admissions which is shown in Table 1.

Table 1.Class distribution by split.

Note: Minor rounding to patient-stratified split boundaries; weighted cross-entropy was applied during training to address class imbalance.

The graph G = (V, E, W) was constructed as follows. Each ICD-10 diagnostic category appearing in at least 50 admissions was assigned a node v_i. A weighted undirected edge was created between nodes i and j if they co-occurred in at least 30 admissions, with weight:

where P ( d i , d j ) is the joint empirical co-occurrence probability, P ( d i ) and P ( d j ) are marginal probabilities, and N ( i , j ) is the raw count. The log-odds term penalizes spurious co-occurrence from high-prevalence diagnoses; the N term rewards statistically stable associations. Edges with negative log-odds were excluded. The comorbidity graph G, all edge weights W, Louvain community assignments, and population-level node feature statistics (mean YMRS, HAMD-17, MMSE, HRV summaries, demographic summaries) were computed exclusively from the training partition (70% of admissions, patient-stratified). Validation and test admissions contributed no information to graph construction or node feature initialization. At inference time, a new admission’s subgraph is extracted from the fixed training-derived graph; no test-time re-estimation of edge weights or community structure is performed. This design eliminates graph-level leakage and ensures that reported test-set metrics reflect genuine out-of-sample generalization.

Each node v_i was initialized with a 64-dimensional feature vector composed of six groups. All group statistics were computed from training admissions only.

1) Network centrality measures [12 dimensions]: degree centrality, betweenness centrality, closeness centrality, eigenvector centrality - each computed on the weighted graph G and encoded as 3 values each (raw, z-scored, log-transformed), totaling 12 dimensions.

2) Clinical scale statistics [12 dimensions]: for patients in the training set carrying diagnosis i, we recorded mean and standard deviation of YMRS (2 values), HAMD-17 (2 values), and MMSE (2 values) - 6 values total. These were supplemented by the proportion of patients with YMRS ≥ 20 (mania threshold), HAMD-17 ≥ 14 (depression threshold), and MMSE ≤ 24 (cognitive impairment threshold), plus 3 additional derived statistics (YMRS–HAMD-17 correlation, MMSE–age partial correlation, scale missingness rate), yielding 12 dimensions total.

3) Biosignal HRV summary statistics [12 dimensions]: available for 61.3% of admissions (wearable-equipped subset); for the remainder, values were imputed using the training-set diagnosis-specific mean for each HRV feature. Features: SDNN, RMSSD, LF power, HF power, LF/HF ratio, mean RR interval - each as mean and SD across the admission, totaling 12 dimensions.

4) Temporal pattern features [12 dimensions]: mean calendar month of admission (sine and cosine encoded, 2 dims), mean day-of-week of admission (sine and cosine encoded, 2 dims), proportion of admissions in each meteorological season (4 dims), mean length-of-stay for this diagnosis (1 dim), readmission rate within 30 days (1 dim), mean time-since-prior-admission (1 dim), and trend slope of admission frequency over the 6-year study period (1 dim) - totaling 12 dimensions.

5) Louvain community membership [12 dimensions]: one-hot encoding of the K = 3 primary communities (3 dims) expanded to 12 by appending within-community degree rank (1 dim), bridge-node score (1 dim), internal vs. external edge ratio (1 dim), and 6 soft community membership scores from the second-best Louvain partition resolution, ensuring the 12-dimensional block is fully populated.

(6) Demographic summary statistics [4 dimensions]: mean age, proportion female, mean socioeconomic deprivation index, and proportion with known social isolation - all computed over training-set patients carrying diagnosis i.

Before training the GAT, three community detection algorithms are applied to G and evaluated by modularity Q and normalized mutual information (NMI) against expert clinical annotations:

1)Louvain Modularity Optimization:Maximizes Q = ∑ c [ L c / m − ( d c / 2 m ) 2 ] , where L_c is internal edges in community c, m is total edge count, and d_c is the sum of degrees in c. Scales efficiently to large graphs and handles weighted edges natively. Selected as primary algorithm based on best performance:Q = 0.681, NMI = 0.79 vs. expert annotations [27].

2)Girvan-Newman:Iteratively removes edges with the highest betweenness centrality, revealing community structure through structural redundancy. Computationally intensive (O(m²n)) but provides mechanistically interpretable communities. Q = 0.612, NMI = 0.71 [28].

3)Spectral Clustering:Applies k-means to the leading eigenvectors of the normalized Laplacian L = D^−1/2AD^−1/2, with k selected via the eigengap heuristic. Q = 0.594, NMI = 0.67 [29].

The K communities from Louvain are one-hot encoded and appended to each node’s feature vector before GAT training, allowing the network to condition its attention on cluster membership.

The GAT consists of two stacked attention layers. At layer l, node i aggregates from its neighbourhood N(i) as:

where W^(l) is the learnable weight matrix and σ is ELU activation. The attention coefficient α_ij is computed as:

where a is a learnable attention vector and ‖ denotes concatenation. Layer 1 uses H = 8 attention heads (64 dimensions each; outputs concatenated to 512 dimensions). Layer 2 uses H = 4 heads (128 dimensions each; outputs averaged). Between layers: batch normalization, ELU, and dropout (p = 0.30). Global mean pooling over all node embeddings produces the graph-level representation passed to the Bayesian output.

PsychoGraph-Net approximates Bayesian inference via Monte Carlo Dropout with T = 100 stochastic forward passes at inference time. With dropout (p = 0.30) active during inference, each pass samples a different sub-network, producing a distribution of predictions. The final prediction is the mean probability vector across passes. Epistemic uncertainty for patient i is:

Predictions with σᵢ > τ = 0.15 are flagged for clinician review rather than direct clinical use. This selective prediction protocol operationalizes the core safety principle: assert only when confident, escalate when uncertain.

Training used the Adam optimizer (lr = 1 × 10⁻³, weight decay 1 × 10⁻⁴) for up to 300 epochs with early stopping (patience = 20 epochs on validation loss). Weighted cross-entropy addressed class imbalance. Bayesian hyperparameter search covered 150 configurations with 5-fold stratified cross-validation. Implementation: PyTorch Geometric [30]. Hardware: NVIDIA A100 GPU (40 GB). Figure 2Figure 2 shows the convergence behaviour, reaching plateau at epoch 87 with final training accuracy 0.954 and validation accuracy 0.921.

Figure 2. Training and Validation Convergence. Accuracy (left) and cross-entropy loss (right) across 120 epochs for PsychoGraph-Net (navy), GCN baseline (teal), and Random Forest (coral). Solid lines = training; dashed = validation. PsychoGraph-Net achieves the highest validation accuracy; early stopping fires at epoch 87 (gold dotted line). The narrow train-validation gap (0.033) indicates well-controlled generalization.

Figure 3Figure 3 shows the full comorbidity network and its community structure. The graph exhibits hub-and-spoke topology consistent with scale-free biological networks, with major depressive disorder, anxiety disorders, delirium, cardiovascular disease, and substance use disorders occupying the highest-degree positions.

Three communities are of primary clinical interest:

1)ClusterAAffective-Anxiety-Sleep Complex:MDD, anxiety disorders, PTSD, OCD, BD-I, BD-II, and sleep disorder form a densely interconnected community (mean internal edge weight w̄ = 0.82). The MDD-anxiety edge carries the highest weight in this cluster (w = 0.92). Sleep disorder appears as a bridge node connecting the affective and anxiety sub-communities consistent with its dual role as a symptom and a risk factor across the affective spectrum.

2)Cluster B Neurocognitive-Delirium Complex:Delirium, dementia, mild cognitive impairment, and UTI cluster tightly together. The delirium-UTI edge

Figure 3. Psychiatric Comorbidity Network. Three Identified Clusters. 89 ICD-10 diagnostic nodes and 1247 co-occurrence edges. Node size ∝ weighted degree; edge width ∝ co-occurrence weight. Louvain community detection identifies three clusters: Affective-Anxiety-Sleep (navy), Neurocognitive-Delirium (coral), and Somatic-Metabolic (teal). These clusters have been described individually in the clinical literature but have not previously been characterized as unified network modules in liaison psychiatry.

carries the highest weight within this cluster (w = 1.23), computationally validating the well-established clinical relationship between urinary tract infection and acute confusional states in older adults. Heart failure appears as a bridge node to the Somatic-Metabolic community, reflecting the cerebral hypoperfusion mechanism underlying cardiac delirium.

3)Cluster C Somatic-Metabolic Interface:Somatic symptom disorder, type 2 diabetes, hypertension, heart failure, and substance use disorder cluster with high internal coherence. The somatic–anxiety link (w = 0.87) and somatic–metabolic disease links (w = 0.62 - 0.77) suggest a previously uncharacterized comorbidity module with direct implications for integrated care: patients presenting with somatic complaints carry a high metabolic burden that is systematically under-recognised in psychiatric assessment.

Figure 4Figure 4 presents the confusion matrix for PsychoGraph-Net on the held-out test set. The model performs well across all five diagnostic categories, with the highest per-class accuracy for Euthymia (94.1%) and Mania (91.8%). The most frequent misclassification is between Hypomania and Euthymia (7.3% of Hypomania cases), which is clinically unsurprising given the phenomenological overlap between these states in cross-sectional assessment.

Figure 5Figure 5 presents the per-class ROC curves and the AUC-ROC ranking across all models. PsychoGraph-Net achieves the highest macro-AUC of 0.951.

Table 2 presents the full performance comparison. PsychoGraph-Net achieves statistically significant improvements over all baselines across all four primary metrics.

Figure 4. Confusion Matrix PsychoGraph-Net Mood State Classification. Raw counts (left) and row-normalized accuracy (right) on the held-out test set (n = 2,143 admissions). Diagonal elements represent correct classifications. The primary confusion is between Hypomania and Euthymia consistent with the inherent clinical difficulty of distinguishing these states in snapshot assessment.

Figure 5. ROC Analysis and AUC Comparison. Left: Per-class ROC curves for PsychoGraph-Net. Each class achieves AUC > 0.93. Right: Macro-AUC comparison across all models. PsychoGraph-Net (0.951) significantly outperforms GCN baseline (0.912) and XGBoost (0.872) (DeLong test, p < 0.01 for both).

Table 2.Comparative model performance on held-out test set (n = 2143 admissions).

Statistically significant improvement over all baselines (DeLong test for AUC, p < 0.01; McNemar test for F1, p < 0.001). ECE = Expected Calibration Error (lower is better). All metrics on patient-stratified test set.

Figure 6Figure 6 presents the feature group importance decomposition and the learned GAT attention weight matrix. Network centrality measures and clinical scales contribute the most predictive information, followed by biosignal HRV features. The community embedding encoding contributes ~15% of predictive information, validating that the Louvain pre-processing step provides signal not available in raw node features.

The attention weight matrix is directly interpretable. The highest learned coefficients correspond to: MDD-Cardiovascular disease (ᾱ = 0.89), Delirium-UTI (ᾱ= 0.84), and Anxiety-Somatic Symptom Disorder (ᾱ = 0.81). These are precisely the associations that experienced liaison psychiatrists consider most clinically salient. (See Figure 7Figure 7)

Figure 6. Feature Importance and Graph Attention Weights. Left: Feature group importance from Random Forest decomposition. Network centrality and clinical scales dominate, followed by biosignal HRV. Right: Learned GAT attention weight matrix α_ij across the ten most central diagnostic nodes. Clinically coherent pairs MDD→CVD, Delirium→UTI, Anxiety→Somatic receive the highest weights, providing face validity for the model’s learned attention.

Figure 8Figure 8 presents the calibration reliability diagram and ablation results. PsychoGraph-Net achieves ECE = 0.031 near-perfect calibration, substantially outperforming the GCN baseline (ECE = 0.061) and Logistic Regression (ECE = 0.128).

Figure 7. All-Metrics Comparison and Bayesian Uncertainty Distribution. Left: Grouped bar comparison of Accuracy, Macro-F1, Precision, and Recall across all models. PsychoGraph-Net leads on every metric. Right: Distribution of epistemic uncertainty (MC-Dropout standard deviation) for correct (teal) and incorrect (coral) predictions. Correct predictions cluster at low uncertainty; incorrect predictions have higher variance. The τ = 0.15 review threshold (navy dashed) effectively separates the reliable from the uncertain prediction regime.

Figure 8. Calibration Reliability Diagram and Module Ablation. Left: PsychoGraph-Net (navy) lies closest to the perfect calibration diagonal (dashed), confirming ECE = 0.031. Right: Ablation study quantifying the independent contribution of each module. Removing the GAT contributes the largest accuracy drop (−3.0 pp); removing the Bayesian layer the largest calibration degradation. All modules contribute independently, and their combination yields super-additive benefit.

The ablation study confirms that every module contributes independently. Removing the Bayesian layer: accuracy −1.5 pp, F1 −1.7 pp. Removing the GAT: accuracy −3.0 pp, F1 −3.3 pp. Removing Clinical BERT: accuracy −4.6 pp. Removing biosignal TCA: accuracy −5.8 pp. Graph-only model (no text or biosignal): accuracy 0.838, F1 0.802 still competitive with tabular baselines but substantially below the full system.