TEM represents the intersection of two distinct pharmacological phenomena: antidepressant-induced hypomanic/manic switch and antidepressant-induced cycle acceleration. In BD-II, TEM manifests most commonly as hypomanic induction, often within 4 - 12 weeks of antidepressant initiation, though cases of delayed acceleration up to 6 months post-initiation have been reported. Risk factors established by clinical studies include BD-I vs. BD-II (paradoxically, BD-II may have higher real-world TEM rates due to lower treatment safeguards), prior TEM history, antidepressant class (TCAs highest, bupropion lowest), absence of adequate mood stabilizer coverage, young age at onset, rapid cycling history, mixed episode history, and recent hypomanic episode [5][6]. Circadian biomarker studies have documented that a lower IS score in the weeks preceding antidepressant initiation predicts subsequent mood instability, suggesting a chronobiological vulnerability window [7].

Long Short-Term Memory networks [8] have demonstrated strong performance on longitudinal clinical prediction tasks, including sepsis onset, readmission prediction, and medication adherence modelling. The addition of multi-head self-attention [9] to LSTM outputs enables the model to selectively weight time points within the sequence based on their relevance to the target, complementing the LSTM’s recurrent memory with a global receptive field over the 30-day monitoring window. For mood state modelling, this is particularly relevant: the hypomanic prodrome may be concentrated in the final 5 - 10 days of the pre-antidepressant window, and attention weights over time steps should reflect this.

Episode-feature attention networks require explicit nodes, edges, and neighbourhood aggregation [10][11]. The present episode-history module does not satisfy that definition because it receives a fixed patient-level feature vector. It is therefore treated as a feed-forward feature-attention model. True graph modelling of bipolar episode histories would require episode nodes, temporally ordered edges, optional pharmacological or similarity edges, and message passing over each patient’s adjacency structure.

Hypomanic activation in BD-II produces characteristic digital signatures: increased step count and physical activity, reduced sleep duration with preserved energy, increased communication frequency, greater smartphone usage entropy, and rising circadian amplitude (RA) with declining IS score [12][13]. These signatures, detectable through passive smartphone sensing and wrist actigraphy in the weeks preceding antidepressant initiation, constitute a potentially actionable prodromal risk window for TEM prediction.

The cohort comprised N = 750 fully synthetic BD-II patients with simulated antidepressant exposure and TEM prevalence of 22.0%. No real patient records were used. Marginal ranges and relative-risk directions were informed by antidepressant-switch literature, bipolar treatment guidance, episode-recurrence theory, and digital phenotyping studies [14]-[17]. Clinical and pharmacological continuous variables were sampled from truncated normal, log-normal, or beta distributions; binary variables from Bernoulli distributions; and antidepressant and mood-stabilizer classes from categorical distributions.

Cross-variable dependence was imposed through correlated latent Gaussian factors transformed to each required marginal distribution. Prior TEM, rapid cycling, mixed-state history, recent hypomania, shorter episode intervals, and higher-risk antidepressant classes were positively correlated. Mood-stabilizer adequacy was negatively correlated with TEM risk. Circadian instability linked lower IS, higher IV, shorter and more variable sleep, increased activity, and rising YMRS sub-scores. Patient-specific random intercepts, day-level autoregressive noise, and measurement noise were added to the 30-day sequences.

TEM labels were sampled from a probabilistic function of prior TEM, antidepressant-class risk, mood-stabilizer adequacy, recent hypomania, rapid cycling, episode-interval shortening, circadian instability, and selected interactions. The intercept was calibrated to 22.0% prevalence. Because related signals occur in both predictors and the label generator, performance may partly reflect recovery of the simulation rules; stochastic noise and overlapping class distributions reduce but do not eliminate this structural optimism.

The complete cohort was synthesized first and then split once at the patient level into development, validation, and held-out test sets. The test set contained 113 patients and was not used for preprocessing, synthetic oversampling, hyperparameter selection, threshold selection, early stopping, or ensemble fitting. All partitions were stratified by TEM outcome [18].

Episode-history features (14) encoded depressive and hypomanic episode counts and severity summaries, mixed episode count and fraction, mean inter-episode interval and shortening trend, sequence entropy, depression-to-hypomania ratio, and recent hypomanic episode status. These were patient-level engineered variables, not nodes or edges in an explicit graph. Pharmacological, clinical, circadian, and 30-day longitudinal features were retained as originally specified.

The sequence encoder comprised a 2-layer LSTM (hidden dimension 128, dropout 0.2) processing 30  × 8 input sequences, followed by a 4-head multi-head self-attention module (attention dropout 0.1) over the full 30-day LSTM output sequence. Mean pooling over the attended sequence produced a 128-dimensional sequence embedding, projected to 64 dimensions and classified by a 2-layer head with GELU activation and dropout of 0.3. The self-attention mechanism enables the model to selectively upweight time steps corresponding to the hypomanic prodromal window rather than treating all 30 days equally.

The episode-feature attention network operated on a fixed 14-dimensional patient-level episode-history vector. A two-layer fully connected encoder projected the vector to 128 dimensions, followed by a learned gating/attention function that reweighted latent feature dimensions before classification. There were no episode nodes, adjacency matrix, edge weights, neighbourhood sets, or graph message-passing steps. Accordingly, this component is a feed-forward attention block and is reported as such throughout the manuscript.

Five-fold stratified OOF stacking combined exactly three Level-1 learners: XGBoost, LightGBM, and LSTM + Attention. The episode-feature attention network was a standalone comparator and did not contribute to the stack. For each development fold, preprocessing and SMOTE were fit only on that fold’s training partition; predictions for the untouched fold formed the OOF meta-learner inputs. After hyperparameters were selected using development/validation data, base learners were refit on the full development set, and the logistic-regression meta-learner was applied to the held-out test predictions. The classification threshold was 0.27, selected on the validation set and locked before test evaluation. Leakage control was implemented at the patient level. No patient appeared in more than one partition; test observations were never used for feature scaling, SMOTE, early stopping, model selection, SHAP fitting, or threshold optimisation. The simulator used common population-level rules but independent patient-level random draws, so no duplicated templates or trajectories were shared across partitions. Nevertheless, the shared generative mechanism remains a source of optimism and is explicitly acknowledged.

Figure 1Figure 1 presents calibration curves on the held-out synthetic test set. The proposed ensemble had the lowest Brier score (0.025), followed by LSTM + Attention (0.036). Given the modest number of TEM events and the synthetic label mechanism, these values indicate good apparent calibration within this simulation rather than near-perfect real-world calibration. Figure 2Figure 2 reports simulated decision-curve results and should not be interpreted as evidence of clinical utility.

Figure 1. Calibration curves (reliability diagrams) for all seven models. Brier scores: Proposed Ensemble 0.025 (lowest), LSTM + Attention 0.036, LightGBM 0.068, XGBoost 0.071, LR 0.093, RF 0.110, GAT 0.110. The ensemble and LSTM + Attention show the nearest tracking to perfect calibration. Lower Brier score = better.

Figure 2. Decision curve analysis: net benefit versus threshold probability (0.0 - 0.70). The Proposed Ensemble and LSTM + Attention module achieves the highest net clinical benefit across threshold probabilities 0.15 - 0.55, confirming the superior clinical utility of longitudinal sequence-informed TEM risk stratification over tabular models alone.

Table 1 presents performance on the held-out synthetic test set. The proposed three-model stack (XGBoost + LightGBM + LSTM + Attention) achieved AUC = 0.997, F1 = 0.958, sensitivity = 0.958, and specificity = 0.989. LSTM + Attention achieved AUC = 0.984. The episode-feature attention comparator achieved AUC = 0.914. The high estimates are proof-of-concept results and may be optimistic because the simulator embeds predictive temporal and pharmacological structure.

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

Note: 95% bootstrap CI (1000 resamples): AUC [0.988 - 1.000], F1 [0.885 - 1.000], sensitivity [0.863 - 1.000], specificity [0.963 - 1.000]. Threshold = 0.27 (F1-optimised on validation set).

Figure 3Figure 3 presents SHAP values from the XGBoost and LightGBM base learners after feature alignment. These attributions do not directly explain the LSTM sequence encoder or the logistic stacking meta-learner and, therefore, should not be described as unified explanations of the complete ensemble. Prior TEM history, antidepressant-class risk, mood-stabilizer adequacy, IS score, and episode-interval shortening were the largest tree-model attributions.

Figure 3. SHAP beeswarm—top 20 predictors of TEM in BD-II. Combined XGBoost + LightGBM TreeExplainer SHAP values. Each point represents one test patient. Red: high feature value; blue: low feature value. Prior TEM history, AD class risk, mood stabilizer adequacy, IS score, and episode interval shortening are the five dominant predictors.

Figure 4Figure 4 presents ROC curves with bootstrap confidence bands. The LSTM + Attention point-estimate AUC exceeded the strongest tabular baseline by 0.041 in the synthetic test set. This is consistent with the simulator assigning predictive temporal structure to the 30-day sequence, but it does not establish that the same gain will occur in prospectively collected digital phenotyping data. The episode-feature attention comparator performed below the tabular models.

Figure 4. ROC curves with 95% bootstrap CI bands (400 resamples). LSTM + Attention (purple) substantially outperforms all tabular baselines (ΔAUC = 0.041 vs. XGBoost). Proposed ensemble (dark bold) achieves AUC = 0.997. Episode-Feature Attention (teal) shows the widest CI, consistent with its limited episode-only feature scope.

Figure 5. Confusion matrix—proposed ensemble (threshold = 0.27). True TEM: 24 correctly classified (sensitivity = 0.958). True non-TEM: 87 correctly classified (specificity = 0.989). False negatives = 1; false positives = 1.

Figure 5Figure 5 presents the confusion matrix at the validation-selected threshold of 0.27. Of 113 synthetic test patients, 111 were correctly classified. Among 25 TEM cases, 24 were identified; among 88 non-TEM cases, 87 were identified. These counts are descriptive of one held-out simulated test set.

Figure 6Figure 6 presents SHAP dependence plots for the three highest-ranked predictors. The prior TEM history dependence plot confirms a bimodal SHAP distribution (positive history  →  strong positive contributions; no history  →  consistent negative contributions). The antidepressant class risk plot shows a monotonically increasing SHAP-to-risk relationship across the continuous risk index (0.12 for bupropion to 0.45 for TCAs), with an inflection above risk index 0.30 suggesting a non-linear acceleration of TEM probability at higher-risk class exposures. The mood stabilizer adequacy plot confirms a monotonically protective dose-response, with the steepest SHAP gradient in the 0.3 - 0.7 adequacy range, suggesting that partial mood stabilizer coverage provides meaningful but incomplete protection.

Figure 6. SHAP dependence plots for the three dominant predictors. Left: prior TEM history (bimodal distribution). Centre: AD class pharmacological risk index (monotonically increasing, inflection at risk > 0.30). Right: mood stabilizer adequacy (monotonically protective, steepest gradient 0.3 - 0.7). Curves: second-degree polynomial fits.

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

Note: log₁₀(BF): originally reported base-10 Bayes-factor calculation relative to the proposed ensemble. Because the ensemble complexity count excludes its base learners, these values are exploratory and should not be interpreted as decisive posterior evidence.

Table 2 and Figure 7Figure 7 report the original exploratory BIC, WAIC, and Bayes-factor calculations, with the ensemble BIC consistently reported as 35.9. For neural networks, boosting models, and stacked ensembles, likelihoods and effective parameter counts are not uniquely defined. Counting only four meta-learner coefficients for the stack omits the complexity of its three fitted base learners and mechanically favours the ensemble. Accordingly, these criteria are retained for transparency but are not treated as decisive evidence of superiority.

Figure 7.Exploratory Bayesian model comparison. Left: reported BIC values (proposed ensemble = 35.9, consistent with Table 2). Right: reported Bayes-factor calculations. Because the ensemble parameter count excludes base-learner complexity, these comparisons should be interpreted cautiously.

Table 3 presents exploratory subgroup point estimates. Several groups are small, and estimates near 1.00 are unstable and may reflect the synthetic label mechanism. Subgroup-level prediction vectors were not available in the manuscript package, so valid bootstrap confidence intervals could not be reconstructed without inventing results. The table, therefore, explicitly marks confidence intervals as not reported (NR); they should be added from the completed analysis outputs before resubmission.

Table 3.Subgroup analysis—proposed ensemble performance.

Note: Test set (n = 113). Low IS score: below test-set median. High-risk AD class: TCA, SNRI, or MAOI. All subgroup analyses are exploratory. NR = confidence interval not reconstructable from the available aggregate outputs; subgroup-level bootstrap intervals must be inserted from the analysis files.

Figure 8. SHAP waterfall for the highest-confidence TEM prediction (left) and highest-confidence non-TEM prediction (right).

The TEM case is characterised by prior TEM history, high-risk AD class, low IS score, and recent hypomanic episode. The non-TEM case is protected by adequate mood stabilizer coverage, a high IS score, and a low-risk AD class. Red: features increasing TEM risk; Blue: features reducing TEM risk (See Figure 8Figure 8).