The causal relationship between SSRI exposure and mood destabilisation in BD has been debated for four decades, with evidence accumulating from naturalistic cohorts, post-marketing surveillance, and prospective randomised trials. In the STEP-BD randomised trial, durable recovery occurred in 23.5% of participants receiving adjunctive antidepressant therapy and 27.3% receiving placebo, while treatment-emergent affective switch occurred in 10.1% and 10.7%, respectively; the trial therefore did not show increased switch risk with adjunctive antidepressant treatment [7]. Serotonergic mechanisms have nevertheless been proposed, including 5-HT₂A receptor sensitisation, serotonin transporter occupancy effects on dopaminergic tone, and genotype-dependent differences in serotonin reuptake. Mood stabiliser co-prescription may attenuate, but does not necessarily eliminate, destabilisation risk through GABAergic, glutamatergic, and neurotrophic mechanisms.

TCNs, introduced by Bai et al. as a general-purpose sequence-modelling architecture, employ dilated causal convolutions that use only current and previous time steps while expanding the receptive field exponentially through increasing dilation. For a stack containing two kernel-size-k convolutions at each dilation d, the receptive field is 1 + 2(k − 1)Σd. With k = 3 and dilations 1, 2, 4, and 8, the architecture used here has an effective receptive field of 61 visits, exceeding the 12-visit maximum input length. Residual connections allow gradients to bypass convolutional transformations and improve optimisation stability [8].

Longitudinal EHR sequences have been modelled using LSTM networks for readmission prediction [8] Transformer encoders for ICU deterioration [10] and convolutional architectures for mortality prediction. Psychiatric EHR modelling has focused primarily on depression severity prediction from natural language processing of clinical notes and diagnostic code sequences. No deep learning study has modelled longitudinal feature sequences from psychiatric outpatient EHRs for pharmacovigilance prediction, leaving the temporal EHR signature of SSRI-induced destabilisation clinically uncharacterised.

Gradient-based saliency maps, first proposed by Simonyan et al. [11] for image classification, identify input dimensions that influence model predictions by computing the gradient of the output with respect to the input. Applied to temporal sequence models, saliency maps can indicate which observed time steps contribute most strongly to a model output. In this study, temporal saliency is used as an exploratory attribution method for generating hypotheses about visit-level model sensitivity; it does not itself establish causality or justify a clinical monitoring recommendation.

This study used an entirely synthetic proof-of-concept cohort of N = 750 independent simulated patients with bipolar disorder (BD-I: 58%; BD-II: 42%) receiving an SSRI-containing regimen. No observations were extracted from hospitals, registries, clinical databases, or individual patient records, and there was no hybrid real-synthetic component. Because the cohort was generated rather than recruited, there was no clinical site, registry source, recruitment period, or patient-consent process. One trajectory was generated per simulated patient, with a maximum of 12 scheduled follow-up visits and 14 time-varying features. The analytical cohort consisted of all 750 generated trajectories, so no patient contributed multiple sequences. Feature values were generated within clinically plausible bounded ranges using stochastic longitudinal processes that incorporated baseline heterogeneity, within-patient autocorrelation, gradual medication titration, adherence variation, symptom drift, and random measurement noise. The overall simulated destabilisation prevalence was fixed at 30.0% as a design choice intended to approximate the range discussed in the clinical literature; it is not an empirical prevalence estimate. The complete data-generation specification, parameter settings, code, and random seed should accompany the revised submission to permit exact reproduction [12]-[14].

The binary endpoint was a simulator-defined first incident hypomanic, manic, or mixed episode after SSRI initiation. Event status and event time were assigned by the synthetic data-generation process before model fitting and were not derived from model predictions. Because no clinical adjudication occurred, the endpoint should be interpreted as the simulator’s binary destabilisation state rather than a validated clinical diagnosis. For the primary sequence-level analysis, the prediction landmark was the last observed visit available to the model: the visit immediately before event onset for destabilised cases and the final available follow-up visit for stable cases. Prefix-based analyses treated each successive observed visit as a provisional landmark. Event onset was assigned within the 12-visit follow-up horizon; for an event case, model input ended at the last visit before onset. The event visit and every later visit were masked and excluded from feature aggregation, training, saliency computation, and prefix-based prediction. For stable cases, no event occurred during the 12-visit horizon. The evaluation therefore asks whether available pre-event information is associated with a later simulated event rather than whether the model can recognise measurements recorded during or after destabilisation [15].

Fourteen features were extracted per clinical visit, spanning five domains: symptom severity (MADRS total, YMRS total), functional status (GAF score), pharmacological parameters (SSRI dose, mood stabiliser serum level proxy), physiological biomarkers (HRV SDNN, sleep duration, daily step count), patient-reported outcomes (mood self-rating, anxiety severity), and prescribing behaviour (visit gap days, medication change flag, side-effect burden, prescription fill ratio). All features were normalised to [0, 1] within the feature-specific clinical range. For the tabular XGBoost baseline, three temporal aggregates were computed per feature across the 12 visits: visit mean, visit standard deviation (variability), and trend (mean of last 3 minus mean of first 3 visits), yielding 42 XGBoost features plus 14 static patient-level meta-features.

The TCN architecture comprises four sequential residual TCN blocks. Each block contains two causal dilated convolutions (kernel size k = 3; hidden channels = 64), each followed by layer normalisation and GELU activation, with dropout (p = 0.20). Block dilations are 1, 2, 4, and 8. Because each block contains two convolutions, the effective receptive field is 1 + 2(k − 1)(1 + 2 + 4 + 8) = 61 visits, which covers the complete 12-visit input window. A pointwise 1 × 1 convolution aligns the residual path when required. The network produces a hidden representation at every valid time step; however, the primary classifier is sequence-level rather than a separate classifier at each visit. Masked global average pooling aggregates only observed pre-event time steps into a 64-dimensional embedding, which is passed through a two-layer classification head (64 → 32 → 2; GELU; dropout 0.20). The risk trajectories shown later were generated by rerunning the same sequence-level model on progressively longer prefixes, not by exposing future visits to earlier predictions.

All sequence models were trained with class-weighted cross-entropy loss, AdamW optimiser (learning rate 1 × 10⁻³; weight decay 1 × 10⁻⁴), cosine-annealing learning-rate scheduling, and early stopping on validation AUC with patience = 12 epochs. Simulated patients were split once at the patient level using stratified allocation by the binary outcome into training (70%, n = 525), validation (15%, n = 112), and held-out test sets (15%, n = 113). Because only one sequence was generated per patient, no patient, duplicated trajectory, or overlapping sequence appeared in more than one split. Destabilisation was defined as a simulator-assigned incident hypomanic, manic, or mixed episode occurring after the applicable prediction landmark and within the remaining follow-up window. In event cases, the event visit and all subsequent visits were excluded and represented by a mask; stable cases retained all available visits. The primary reported models did not use SMOTE. Class imbalance was handled through class-weighted loss because flattening, oversampling, and reshaping multivariate sequences can create artificial trajectories that do not preserve within-patient temporal dependence. No claim is made for a flattened-sequence SMOTE sensitivity analysis. Five-fold out-of-fold predictions within the training data were used to fit the logistic-regression stacking meta-learner (C = 1.0), after which component models were refitted and evaluated once on the untouched test set. Temporal saliency was calculated as the L1 norm of the gradient of the destabilisation score with respect to observed input values, averaged across features at each valid visit.

Performance was evaluated on the held-out test set using AUC, F1, sensitivity, specificity, Brier score, and decision-curve analysis; 95% confidence intervals were estimated with 1000 bootstrap resamples. The decision threshold (0.32) was selected using the validation set and then fixed before test evaluation. Additional analyses included XGBoost SHAP importance for temporal aggregates, gradient-based temporal saliency for the TCN, and exploratory subgroup analyses. BIC-, WAIC-, and Bayes-factor summaries are reported as model-fit/parsimony analyses and are not interpreted as substitutes for held-out discrimination. Because the data are synthetic and no external cohort was used, all results are internal and exploratory.

Figure 1Figure 1 presents calibration curves. The ensemble achieved a Brier score of 0.082, whereas logistic regression had the lowest Brier score (0.054). Figure 2Figure 2 presents decision-curve analysis. Within the reported threshold range of 0.25 - 0.55, the ensemble showed the highest estimated net benefit. This decision-curve finding should be interpreted separately from AUC ranking and requires confirmation on real-world data before clinical use.

Figure 1. Calibration curves for all five models. Brier scores: proposed ensemble 0.082, TCN 0.149, LSTM 0.142, XGBoost 0.077, LR 0.054. Perfect calibration = dashed diagonal.

Figure 2. Decision-curve analysis. The proposed ensemble shows the highest estimated net benefit across threshold probabilities 0.25 - 0.55 in this synthetic test set. Note: This does not imply superior AUC and should be regarded as an internal, hypothesis-generating result.

Table 1 presents comparative performance on the held-out test set (n = 113). Logistic regression achieved the highest AUC (0.974), F1 (0.913), sensitivity (0.913), specificity (0.962), and the lowest Brier score (0.054). The proposed ensemble ranked second by AUC at 0.950 (95% CI: 0.892 - 0.989), closely followed by XGBoost at 0.948; the LSTM and TCN achieved AUCs of 0.915 and 0.844, respectively. Therefore, the results do not support a claim that the standalone TCN outperformed the tabular baselines. Instead, they show that summary-feature models were highly competitive in this synthetic dataset, while the TCN contributed a temporally interpretable component to the ensemble.

Table 1. Comparative model performance-test set (n = 113).

95% bootstrap CI (1000 resamples). Threshold = 0.32, selected on the validation set and fixed before test evaluation. XGBoost and logistic regression used temporal summary statistics (mean, SD, trend) plus 14 baseline meta-features (56 features total).

Figure 3Figure 3 summarises the complete data path used by the TCN. Each synthetic patient trajectory enters the network as a 12 × 14 normalised visit-feature tensor accompanied by a validity mask. The four residual blocks use two kernel-size-3 causal convolutions per block, with dilation factors 1, 2, 4, and 8. Under this configuration, the cumulative theoretical receptive field expands from 5 visits after the first block to 13, 29, and 61 visits after the subsequent blocks. The revised diagram reports these cumulative values explicitly, thereby distinguishing the receptive field of the full stacked architecture from the smaller increment contributed by an individual dilated layer.

Because the final theoretical receptive field exceeds the 12-visit sequence length, the deepest representation can incorporate all available observations in the input window. This does not introduce future leakage: causal padding restricts each activation to the current and preceding visits, and the event visit and all later visits remain masked for destabilised cases. Masked global average pooling then averages only valid observed pre-event positions to create a 64-dimensional sequence embedding, preventing padded or excluded time steps from influencing the classifier.

Figure 3. Leakage-aware TCN architecture for sequence-level prediction. A 12-visit × 14-feature tensor and validity mask pass through four residual blocks with dilation factors 1, 2, 4, and 8. Each block contains two kernel-size-3 causal convolutions, giving cumulative theoretical receptive fields of 5, 13, 29, and 61 visits. Masked global average pooling aggregates observed pre-event steps into a 64-dimensional embedding, followed by a 64 → 32 → 2 classifier. Prefix trajectories are produced by rerunning the same model on progressively longer observed prefixes. Note: Receptive-field values are cumulative after each residual block. A theoretical receptive field larger than the 12-visit input indicates complete coverage of the available sequence, not access to future observations.

The embedding is passed to the 64 → 32 → 2 classification head to generate a single sequence-level probability of simulated SSRI-associated mood destabilisation. Prefix-based risk trajectories are therefore obtained by repeatedly applying this same classifier to progressively longer observed prefixes, rather than by attaching an independent prediction head to every visit. The architecture provides a coherent basis for the later saliency and prefix analyses, but its structural sophistication should not be interpreted as evidence of superior discrimination: as shown in Table 1, the standalone TCN achieved lower test-set AUC than the logistic-regression and XGBoost baselines.

Figure 4Figure 4 presents ROC curves with 95% bootstrap confidence bands. Logistic regression achieved the highest AUC (0.974), followed by the proposed ensemble (0.950), XGBoost (0.948), LSTM (0.915), and TCN (0.844). The near-identical performance of the ensemble and XGBoost indicates that the temporal TCN component added limited discrimination in this synthetic test set, although it supported visit-level attribution analyses. The ROC results therefore favour the simpler logistic-regression baseline for discrimination and do not establish a standalone TCN advantage.

Figure 5Figure 5 presents the EHR trajectory heatmap of mean normalised feature values

Figure 4. ROC curves with 95% bootstrap confidence bands (400 resamples). AUC ranking: logistic regression 0.974, proposed ensemble 0.950, XGBoost 0.948, LSTM 0.915, and TCN 0.844. Note: The figure does not support a claim that TCN outperforms the tabular baselines.

Figure 5. EHR clinical trajectory heatmap. Mean normalised feature values per visit for destabilised (top, Reds colormap) and stable (bottom, Blues colormap) synthetic patients. Orange dashed line: simulated SSRI titration completion boundary. The displayed patterns include rising YMRS and declining mood-stabiliser level from approximately visit 4 and declining prescription fill ratio from approximately visit 5 in the destabilised group. Each cell is annotated with its mean normalised value.

per clinical visit for destabilised (top) and stable (bottom) synthetic patients. Three simulator-dependent patterns distinguish the groups. First, YMRS shows a progressive upward trend in destabilised trajectories from visit 4 onward while remaining comparatively flat in stable trajectories. Second, the mood-stabiliser level proxy declines in destabilised trajectories beginning around visits 3 - 4, consistent with the adherence and dose-adjustment assumptions embedded in the simulator. Third, prescription fill ratio declines from approximately visit 5 in destabilised trajectories, reflecting the simulated treatment-engagement pattern. These group-level patterns illustrate relationships encoded in the synthetic cohort and should not be interpreted as empirical clinical trajectories or causal mechanisms.

Figure 6Figure 6 presents XGBoost SHAP feature importance for the top 15 tabular predictors, colour-coded by aggregation type. Baseline clinical-history variables contributed strongly: prior SSRI switch history had the largest mean absolute SHAP

Figure 6. SHAP feature importance for the top 15 EHR-derived predictors. Blue bars: temporal mean features. Green bars: temporal variability features (SD). Red bars: temporal trend features (last 3 visits minus first 3 visits). Orange bars: baseline clinical meta-features. Prior switch history, mixed-episode history, and episode rate show the largest mean absolute SHAP values, while temporal trend and variability features provide additional predictive information.

value, followed by mixed-episode history and episode rate. Temporal predictors also contributed, including anxiety-severity trend, mean HRV SDNN, YMRS variability, prescription-fill-ratio trend, YMRS trend, mood-stabiliser-level variability, and side-effect-burden variability. The figure therefore indicates that the model relied on a combination of baseline vulnerability variables and longitudinal summary features rather than being dominated exclusively by temporal trend variables.

Table 2 and Figure 7Figure 7 present the reported BIC-, WAIC-, and Bayes-factor comparison. Under the stated effective-parameter assumptions, the ensemble has the lowest BIC (82.3). This parsimony-oriented result should not be described as superior held-out discrimination: logistic regression has the highest test AUC and lowest Brier score in Table 1. The Bayesian information criteria and test-set predictive metrics answer different questions, and both are reported without conflating their rankings.

Table 2. Bayesian model comparison: BIC, WAIC, and bayes factors.

k: stated effective parameter count. log₁₀(BF): base-10 Bayes factor calculated relative to the proposed ensemble under the manuscript assumptions. These model-fit summaries should be interpreted separately from held-out AUC and calibration.

Figure 7. Model-fit and parsimony comparison. The proposed ensemble has the lowest reported BIC under the stated effective-parameter assumptions. Note: This result is separate from held-out discrimination, for which logistic regression achieved the highest AUC.

Table 3 presents exploratory subgroup performance for the proposed ensemble. The highest AUC was observed among simulated patients with prior SSRI switch history (0.988), followed by the BD-I subgroup (0.975), the high-genetic-risk subgroup (0.962), and the low-IS subgroup (0.940). However, these estimates are based on small synthetic subgroups and characteristics generated by the simulation process. Differences between subgroups may therefore reflect the simulator’s assumptions and should not be interpreted as evidence of biological homogeneity, genomic causality, circadian mechanisms, or differential real-world clinical predictability.

Table 3. Subgroup analysis-proposed ensemble performance.

High genetic risk: composite genetic risk score > test-set median. Low IS: IS score < median. All subgroup n ≥ 15.

Figure 8Figure 8 presents the TCN temporal analysis in two panels. Panel A shows gradient-based temporal saliency averaged over observed, pre-event visits for destabilised and stable test cases. In the synthetic destabilised trajectories, saliency is highest at the initial visit, decreases across the early-middle visits, and then rises again around visits 7 - 10, with the later peak occurring around visits 8 - 9. Stable cases show a comparatively flatter profile. Because the data-generation process is synthetic and saliency is model-dependent, these patterns are exploratory hypotheses rather than evidence for a specific clinical monitoring window.

Panel B presents prefix-based risk trajectories for four representative synthetic patients. At each visit, the same sequence-level model was rerun using only the visits available up to that point; later visits were masked. The two destabilised examples show increasing predicted risk before their simulated event, while the two stable examples remain below the decision threshold. These plots illustrate how a causal prefix-evaluation workflow can update risk over time without allowing future observations to influence earlier estimates.

Figure 8. TCN temporal analysis in the synthetic test set. (a): gradient saliency over observed pre-event visits, TCN temporal saliency-which visits drive predictions. (b): prefix-based risk trajectories generated by rerunning the sequence-level classifier using only data available up to each visit. Visits at and after simulated event onset are excluded. Predicted risk trajectory - representative patientstcn risk score at each EHR visit; threshold shown as dashed line.