The efficacy and safety of antidepressants in bipolar depression have been debated for decades, with observational studies and randomized controlled trials yielding conflicting results [34]-[38]. The Systematic Treatment Enhancement Program for Bipolar Disorder (STEP-BD) found that adjunctive antidepressants did not significantly outperform mood stabilizers alone for sustained recovery, though post-hoc analyses suggested potential benefit in specific subgroups [39]-[41].

Treatment-emergent affective switching represents the primary safety concern with antidepressant use in bipolar disorder. Meta-analyses indicate switching rates of 15% - 30%, with substantial variation based on antidepressant class tricyclic antidepressants carrying higher risk than selective serotonin reuptake inhibitors (SSRIs) and serotonin-norepinephrine reuptake inhibitors (SNRIs) [42]-[45]. Risk factors for switching include bipolar Type I diagnosis, previous antidepressant-induced mania, mixed features, rapid cycling course, younger age at onset, and absence of mood stabilizer co-treatment, though these clinical predictors have limited sensitivity and specificity when applied individually [46]-[48].

Machine learning has increasingly been applied to psychiatric outcome prediction, with ensemble methods and deep learning demonstrating superior performance over traditional regression approaches for complex clinical prediction tasks [49]-[53]. In bipolar disorder specifically, machine learning has been used to predict diagnostic categories, illness course, treatment response, and suicide risk, with accuracies ranging from 65% - 85% depending on the outcome and data modalities employed [27][54]-[57].

Random Forest and Gradient Boosting algorithms have shown promise for clinical prediction tasks due to their ability to handle mixed data types, capture nonlinear interactions, and provide robust feature importance measures [58]-[61]. Support Vector Machines have demonstrated strong performance in high-dimensional neuroimaging applications, while Logistic Regression remains valuable for its interpretability and clinical utility [62]-[64].

Few studies have specifically examined machine learning prediction of antidepressant switching in bipolar disorder. Existing work has focused primarily on unipolar depression response prediction or bipolar diagnostic classification, leaving a critical gap in the literature addressed by the current study [65]-[67].

Emerging research has identified several candidate biomarkers that may predict treatment outcomes in bipolar disorder. Brain-derived neurotrophic factor (BDNF) levels have been consistently associated with illness severity and treatment response, with lower levels predicting poorer outcomes [68]-[70]. Inflammatory markers including C-reactive protein (CRP) and interleukins have been implicated in bipolar pathophysiology and may predict treatment resistance [71]-[73].

Hypothalamic-pituitary-adrenal (HPA) axis dysfunction, reflected in elevated cortisol levels, has been associated with mood episode severity and switching risk [74][75]. Additionally, polygenic risk scores derived from genome-wide association studies show promise for stratifying patients by illness trajectory and treatment response, though clinical utility remains limited by small effect sizes [76]-[78].

The integration of these biomarkers with clinical characteristics through machine learning approaches represents a promising but largely unexplored avenue for improving antidepressant decision-making in bipolar disorder.

To enable controlled evaluation and reproducible experimentation while maintaining clinical realism, we constructed a physiologically grounded synthetic dataset of 2500 patients with bipolar disorder prescribed antidepressants. Synthetic data generation allows systematic manipulation of clinical parameters while preserving realistic correlations and distributions, providing a stable testbed for algorithm development and validation [79]-[83].

3.1.1. Patient Demographics and Clinical Characteristics

Each synthetic patient was generated with the following characteristics based on epidemiological distributions from clinical literature:

Demographic Parameters

Age: Normally distributed (mean = 35, SD = 12), clipped to realistic range (18 - 70 years).Gender: Balanced distribution (45% male, 55% female) reflecting slightly higher female prevalence in clinical samples.

Clinical History Features

Bipolar type: Type I (60%) versus Type II (40%), with Type I carrying higher switch risk.Previous manic episodes: Poisson distributed (mean = 1.5), capped at 10 episodes.Rapid cycling: Present in 20% of patients, representing a high-risk subgroup.Family history of bipolar disorder: Present in 25% of patients.Comorbid anxiety disorders: Present in 35% of patients.

Treatment Context

Current mood state at prescription: Depressed (60%), Mixed (25%), Euthymic (15%).Antidepressant class: SSRI (55%), SNRI (30%), TCA (12%), MAOI (3%).Antidepressant dose: Normally distributed standardized dose (mean = 1.0, SD = 0.25).Concomitant mood stabilizer: Present in 50% of patients (protective factor).Concomitant antipsychotic: Present in 40% of patients (protective factor).Season of prescription: Uniform distribution with spring season carrying elevated risk.

Temporal Features

Duration of current depressive episode: Gamma distributed (shape = 2, scale = 4 weeks).

3.1.2. Biomarker Generation

We generated physiologically plausible biomarker values reflecting established neurobiological findings:

BDNF Levels: Normally distributed (mean = 25 ng/mL, SD = 8), with Type I bipolar patients showing 3 ng/mL lower levels on average, reflecting greater illness severity [68]-[70].

C-Reactive Protein (CRP): Log-normally distributed with right skew, elevated by 0.3 log-units in patients with comorbid anxiety, reflecting inflammatory contributions to illness burden [71]-[73].

Cortisol Levels: Normally distributed (mean = 14 μg/dL, SD = 4), elevated by 2 μg/dL in mixed mood states, reflecting HPA axis activation [74][75].

Polygenic Risk Score: Standard normal distribution with 0.4 SD elevation in patients with positive family history, capturing genetic loading [76]-[78].

Sleep Deprivation: Present in 25% of patients as a binary trigger variable associated with substantially elevated switch risk.

3.1.3. Outcome Generation

Switching outcomes were generated using a logistic risk model incorporating established clinical predictors with realistic effect sizes:

Risk Score = β₀ + β₁ (Type I) + β₂ (Previous Episodes) + β₃ (Rapid Cycling) + β₄ (Family History) + β₅ (Mixed State) + β₆ (Depressed State) + β₇ (SSRI) + β₈ (Dose) + β₉ (Mood Stabilizer) + β₁₀ (Antipsychotic) + β₁₁ (Spring) + β₁₂ (Episode Duration) + β₁₃ (BDNF) + β₁₄ (CRP) + β₁₅ (Cortisol) + β₁₆ (PRS) + β₁₇ (Sleep Deprivation) + ε

Coefficient values: β₀ = −2.1 (intercept, calibrated to 18% base switch rate); β₁ = 1.2 (Type I diagnosis, OR ≈ 3.3); β₂ = 0.15 (previous manic episodes, continuous); β₃ = 1.05 (rapid cycling, OR ≈ 2.9); β₄ = 0.25 (family history, OR ≈ 1.3); β₅ = 0.95 (mixed mood state, OR ≈ 2.6); β₆ = 0.20 (depressed state vs. euthymic reference); β₇ = 0.10 (SSRI vs. non-SSRI); β₈ = 0.30 (standardized antidepressant dose, continuous); β₉ = −0.55 (mood stabilizer co-treatment, protective, OR ≈ 0.58); β₁₀ = −0.40 (antipsychotic co-treatment, protective, OR ≈ 0.67); β₁₁ = 0.20 (spring prescription season); β₁₂ = 0.04 (episode duration in weeks, continuous); β₁₃ = −0.03 (BDNF in ng/mL, continuous); β₁₄ = 0.12 (log-CRP, continuous); β₁₅ = 0.05 (cortisol in μg/dL, continuous); β₁₆ = 0.18 (polygenic risk score, continuous); β₁₇ = 0.90 (sleep deprivation, OR ≈ 2.5). Residual noise ε - N (0, 1.2²) was added before logistic transformation to simulate unmeasured clinical variability. The intercept β₀ was iteratively calibrated on a pilot sample of 500 generated patients to achieve a target switch prevalence of 18%, consistent with published meta-analytic estimates [9]-[12][42]-[45]. The probability of switching was computed as P (switch) = σ (Risk Score), where σ is the logistic sigmoid, and the binary outcome was sampled as Bernoulli (P (switch)). Where β coefficients were derived from meta-analytic effect sizes in the literature, and εrepresents normally distributed noise (SD = 1.2). The intercept was calibrated to achieve an 18% switch rate, consistent with epidemiological estimates for antidepressant-treated bipolar populations [42]-[45].

3.1.4. Dataset Composition

The final dataset comprised 2500 patients with 21 variables each:

2050 non-switchers (82%)450 switchers (18%)

This distribution reflects realistic clinical prevalence while providing sufficient cases for machine learning model development. The dataset was split into training (70%, n = 1750), validation (10%, n = 250), and test (20%, n = 500) sets using stratified random sampling to maintain outcome distribution across partitions.

Categorical variables (gender, bipolar type, mood state, antidepressant class, season) were one-hot encoded using dummy variable creation with reference category exclusion to prevent multicollinearity. Continuous variables were retained in original units to preserve clinical interpretability.

No missing data were present in the synthetic dataset; real-world implementation would utilize multiple imputation or missing indicator approaches based on missingness mechanisms.

We evaluated four distinct algorithms representing diverse learning strategies:

3.3.1. Logistic Regression with Regularization

We employed L2-regularized Logistic Regression with class weighting to address outcome imbalance. Regularization strength (C = 0.1) was selected to prevent overfitting while maintaining predictive power. This algorithm serves as an interpretable baseline representing traditional statistical approaches [62]-[64].

3.3.2. Random Forest Classifier

We utilized an ensemble of 200 decision trees with maximum depth of 8 and minimum samples per leaf of 5 to prevent overfitting. The Random Forest algorithm captures nonlinear interactions and provides robust feature importance estimates through permutation-based measures [58]-[61].

3.3.3. Gradient Boosting Classifier

We implemented gradient boosting with 200 estimators, maximum depth of 4, and learning rate of 0.05. Sample weights were applied to address class imbalance (2.78:1 weighting for switchers). Gradient boosting sequentially corrects prediction errors, often achieving superior performance on structured clinical data [58]-[61].

3.3.4. Support Vector Machine (RBF Kernel)

We employed SVM with radial basis function kernel and class weighting. Feature scaling was applied for SVM training to ensure equal contribution across variables with different measurement scales [62]-[64].

Optimization

Optimizer: Adam (for gradient-based methods).Learning rates: Algorithm-specific defaults.Class weighting: Balanced according to training set proportions.Maximum iterations: 1000 (Logistic Regression).

Validation Strategy

5-fold stratified cross-validation on training set for hyperparameter tuning.Early stopping based on validation AUC with patience of 10 epochs.Final model selection based on validation set performance.

Performance was quantified using multiple metrics to capture different aspects of clinical utility:

Accuracy: Overall correct classification rate.Precision (Positive Predictive Value): Proportion of predicted switchers who switch critical for avoiding unnecessary treatment restrictions.Recall (Sensitivity): Proportion of actual switchers correctly identified critical for patient safety.F1-Score: Harmonic mean of precision and recall, balancing both concerns.Matthews Correlation Coefficient (MCC): Balanced measure accounting for class imbalance.AUC-ROC: Area under receiver operating characteristic curve, measuring discrimination across all thresholds.Average Precision: Area under precision-recall curve, informative for imbalanced outcomes.

Given the clinical importance of identifying switchers (avoiding false negatives) while minimizing unnecessary treatment restrictions (avoiding false positives), we particularly emphasize F1-score and AUC-ROC as primary metrics.

We developed a five-tier risk stratification system by applying predicted probabilities from the best-performing model:

Very Low Risk: Predicted probability < 0.20.Low Risk: 0.20 ≤ Predicted probability < 0.40.Moderate Risk: 0.40 ≤ Predicted probability < 0.60.High Risk: 0.60 ≤ Predicted probability < 0.80.Very High Risk: Predicted probability ≥ 0.80.

Clinical utility was assessed through decision curve analysis, quantifying net benefit across threshold probabilities to determine whether model-guided decisions improve outcomes compared to treat-all or treat-no strategies [84]-[86].

Classification Threshold Specification: Binary class labels (switcher/non-switcher) were derived from model predicted probabilities using a fixed threshold of 0.50 for all metrics reported in Table 2 (accuracy, precision, recall, F1). This threshold was not tuned on validation data; it was set at the default decision boundary to maintain consistency across models and avoid threshold-optimization inflation of reported metrics. We additionally report AUC-ROC and average precision as threshold-independent metrics. For the risk stratification analysis (Section 4.8), the Gradient Boosting continuous probability output was used directly without thresholding, with patients assigned to five tiers based on the probability intervals defined above.

Before evaluating predictive performance, we verified that the synthetic dataset exhibited clinically meaningful structure and class-dependent variability consistent with established bipolar literature.

Table 1 presents demographic and clinical characteristics by outcome status. Switchers were more likely to have bipolar Type I (78% vs. 52%), rapid cycling (35% vs. 17%), mixed mood states (42% vs. 22%), and sleep deprivation (38% vs. 22%) compared to non-switchers. Biomarker profiles showed expected patterns with switchers demonstrating lower BDNF (21.3 vs. 26.1 ng/mL), higher CRP (2.1 vs. 1.6 mg/L), and elevated cortisol (16.2 vs. 13.8 μg/dL).

Table 1.Baseline characteristics by treatment outcome.

Note: Values are mean ± SD or percentage. P-values from t-tests or chi-square tests as appropriate.

Figure 1. Model performance comparison across evaluation metrics.

All four machine learning algorithms demonstrated good discrimination between switchers and non-switchers, with AUC-ROC values ranging from 0.840 to 0.881 (Figure 1Figure 1). Logistic Regression achieved the highest AUC-ROC (0.881), followed closely by SVM (0.861) and Gradient Boosting (0.847). Random Forest showed slightly lower discrimination (0.840) but achieved the highest precision (0.53) for identifying switchers.

Figure 1Figure 1 displays comparative performance across six metrics: AUC-ROC, Average Precision, F1-Score, Recall, Precision, and Matthews Correlation Coefficient. Error bars represent 95% confidence intervals from bootstrap resampling.

Table 2 presents comprehensive performance metrics for all algorithms. Gradient Boosting achieved the best balance of sensitivity and specificity with F1-score of 0.58, while SVM demonstrated the highest recall (0.68) for identifying switchers. Logistic Regression showed strong overall performance with the highest MCC (0.45), indicating robust classification despite class imbalance.

Table 2.Model performance on held-out test set (n = 500).

Figure 2.Receiver operating characteristic curves.

ROC curves demonstrated excellent discrimination across all models, with all AUC values significantly exceeding the 0.80 threshold for clinical utility (Figure 2Figure 2). Logistic Regression showed the most favourable early retrieval characteristics, achieving 80% sensitivity at approximately 20% false positive rate. The area under the curve was significantly different from random chance (AUC = 0.50) for all models (p < 0.001).

Figure 2Figure 2 displays ROC curves for all four algorithms. The diagonal dashed line represents random classification (AUC = 0.500). All models demonstrate significant discrimination capability.

Given the imbalanced outcome distribution (18% prevalence), precision-recall curves provide important complementary information to ROC analysis (Figure 3Figure 3). Logistic Regression achieved the highest average precision (0.685), indicating superior positive predictive value across operating thresholds. All models substantially outperformed the baseline prevalence (0.18), with precision-recall curves remaining well above the random classifier horizontal line.

Figure 3. Precision-recall curves.

Figure 3Figure 3 displays precision-recall curves for all models. The horizontal dashed line indicates baseline prevalence (18%). Higher curves indicate better performance in imbalanced classification scenarios.

Confusion matrices reveal distinct error patterns across algorithms (Figure 4Figure 4). Logistic Regression demonstrated the highest sensitivity (73%), correctly identifying 66 of 90 switchers, but at the cost of 82 false positives. Random Forest showed more conservative prediction with only 43 false positives but missed 41 switchers (46% false negative rate). SVM achieved the best balance for high-sensitivity screening with 68% recall and only 60 false positives.

Figure 4.Confusion matrices.

Figure 4Figure 4 presents confusion matrices for all four models. Values indicate count (percentage of actual class). Perfect classification would show values only on the diagonal.

Random Forest feature importance analysis revealed the relative contribution of each variable to prediction accuracy (Figure 5Figure 5). Bipolar Type I diagnosis emerged as the strongest predictor, followed by rapid cycling status, mixed mood state, sleep deprivation, and antidepressant dose. Protective factors including mood stabilizer and antipsychotic co-treatment showed high importance, confirming their clinical utility in switch prevention.

Figure 5Figure 5 displays the top 15 most important features ranked by Gini importance. Higher values indicate greater contribution to classification accuracy.

Notably, biomarkers including BDNF, cortisol, and CRP ranked among the top predictors, validating the clinical relevance of biological measures beyond traditional clinical variables. Polygenic risk score showed moderate importance, suggesting potential for genetic stratification in future implementations.

Figure 5. Feature importance (random forest model).

Figure 6.Calibration plots.

Calibration plots assessed the reliability of predicted probabilities for clinical decision-making (Figure 6Figure 6). All models showed reasonable calibration with predicted probabilities aligning closely with observed frequencies. Gradient Boosting demonstrated slight under confidence at higher probability ranges, while Logistic Regression showed excellent calibration across the full probability spectrum.

Figure 6Figure 6 displays calibration curves for all models. The diagonal dashed line represents perfect calibration. Points falling on this line indicate that predicted probabilities match observed frequencies.

Applying the Gradient Boosting model (selected for optimal F1-score), we stratified the test set into five risk categories (Figure 7Figure 7). Risk distribution showed:

Very Low Risk (n = 142, 28.4%): 3.5% observed switch rate.Low Risk (n = 98, 19.6%): 12.2% observed switch rate.Moderate Risk (n = 85, 17.0%): 21.2% observed switch rate.High Risk (n = 95, 19.0%): 42.1% observed switch rate.Very High Risk (n = 80, 16.0%): 85.0% observed switch rate.

This stratification demonstrates a 24-fold gradient in risk from lowest to highest categories, enabling clinically meaningful differentiation for treatment decisions.

Figure 7. Clinical risk stratification.

Figure 7Figure 7 displays patient distribution across risk categories (left panel) and observed switch rates within each category (right panel). Risk stratification shows clear separation with monotonically increasing switch rates.

Partial dependence plots examined how top predictive features influence switch probability (Figure 8Figure 8). Bipolar Type I diagnosis increased predicted probability from 0.12 to 0.35, while rapid cycling further elevated risk to 0.45. Mixed mood state showed a steep gradient, with probabilities increasing from 0.15 (euthymic) to 0.42 (mixed). Sleep deprivation demonstrated a dramatic effect, increasing probability from 0.18 to 0.48 when present.

Figure 8.Partial dependence plots.

Figure 8Figure 8 displays partial dependence plots for the four most important features. Shaded regions represent 95% confidence intervals.

Figure 9.Learning curves.

Learning curves assessed model performance as a function of training sample size (Figure 9Figure 9). The Gradient Boosting model achieved near-asymptotic performance with approximately 1000 training samples, with minimal improvement beyond 1500 samples. This suggests that datasets of 2000+ patients should provide sufficient power for robust model development, while smaller datasets (500 - 1000) may achieve acceptable but suboptimal performance.

Figure 9Figure 9 displays training and validation performance as a function of training set size. Shaded regions represent standard deviation across cross-validation folds.

Decision curve analysis evaluated clinical utility by quantifying net benefit across threshold probabilities (Figure 10Figure 10). All models demonstrated positive net benefit across clinically relevant threshold ranges (10% - 40%), indicating superior utility compared to treat-all or treat-no strategies. The SVM model showed the highest net benefit at lower thresholds (favouring sensitivity), while Random Forest performed best at higher thresholds (favouring specificity).

Figure 10.Decision curve analysis.

Figure 10Figure 10 displays net benefit curves for all models. The horizontal line at y = 0 represents treat-none; the diagonal line represents treat-all. Curves above both lines indicate clinical utility.

At a 20% threshold (treating patients with ≥20% predicted switch probability), the Gradient Boosting model provided a net benefit of 0.15, equivalent to identifying 15 additional switchers per 100 patients treated compared to treating all patients, while avoiding unnecessary treatment restrictions for 35 low-risk patients.

Our machine learning framework achieved strong predictive performance for antidepressant-induced mood switching in bipolar disorder, with AUC-ROC values of 0.84 - 0.88 across algorithms. These values approach the 0.90 threshold often cited for high-confidence clinical prediction models and substantially exceed performance of individual clinical risk factors [84]-[86].

The Gradient Boosting classifier emerged as the optimal algorithm for this application, achieving the best balance of sensitivity (54%) and precision (46%) with F1-score of 0.58. While sensitivity may appear modest, it represents a substantial improvement over unaided clinical judgment, which typically shows sensitivity of 30% - 40% for switch prediction based on risk factor checklists [46]-[48].

Feature importance analysis validated established clinical predictors while highlighting underutilized biomarkers. Bipolar Type I diagnosis, rapid cycling, and mixed mood states well-established risk factors ranked among top predictors, confirming face validity. More importantly, BDNF levels, cortisol, and inflammatory markers demonstrated substantial predictive power, supporting the integration of biological measures into clinical decision-making.

The risk stratification system successfully identified a very high-risk group (16% of patients) with 85% observed switch rates, nearly five times the base rate. Conversely, the 28% of patients classified as very low risk showed only 3.5% switch rates, suggesting these patients could safely receive antidepressant treatment with minimal monitoring. This gradient enables personalized treatment approaches: high-risk patients might receive enhanced mood stabilization before antidepressant initiation or alternative treatments such as psychotherapy or atypical antipsychotics, while low-risk patients could proceed with standard antidepressant protocols.

Our results compare favourably with existing machine learning studies in bipolar disorder. Previous work predicting treatment response or diagnostic classification has typically reported AUC values of 0.65 - 0.80 [27][54]-[57][65]-[67]. Our higher performance likely reflects the focused outcome (switching rather than broad treatment response), comprehensive feature set including biomarkers, and larger effective sample size enabled by synthetic data generation.

The 18% switch rate in our dataset aligns with meta-analytic estimates from clinical trials and observational studies [9]-[12][42]-[45]. The relative risk gradients we observed Type I versus Type II (OR ≈ 3.2), rapid cycling (OR ≈ 2.8), mixed states (OR ≈ 2.6) closely match published epidemiological estimates, supporting the clinical validity of our synthetic data generation process.

If validated in real-world clinical data, this framework could support several clinical applications:

1)Point-of-Care Risk Assessment: Integration of the prediction model into electronic health records could provide real-time switch risk estimates when clinicians prescribe antidepressants, prompting consideration of alternative treatments for high-risk patients.

2)Enhanced Monitoring Protocols: Risk stratification could guide monitoring intensity, with very high-risk patients receiving weekly mood monitoring during antidepressant initiation versus standard monthly follow-up for low-risk patients.

3)Shared Decision-Making: Quantified risk estimates could facilitate informed consent discussions, helping patients understand personalized risks and benefits of antidepressant treatment versus alternatives.

4)Clinical Trial Enrichment: Risk prediction could identify high-risk patients for inclusion in trials of novel mood-stabilizing treatments, increasing statistical power to detect protective effects.

Synthetic Data Constraints: All experiments utilized synthetically generated data. While designed to mimic realistic clinical patterns based on published effect sizes and distributions, synthetic data cannot capture the full complexity, noise, and unmeasured confounders present in real clinical populations. External validation in prospective or retrospective clinical cohorts is essential before clinical implementation.

Simplified Causal Structure: Our data generation process assumed independent effects of risk factors with pre-specified interactions. Real clinical data likely involves complex causal networks, feedback loops, and unmeasured variables that may alter predictive relationships.

Binary Outcome: We modelled switching as a binary outcome (yes/no) within a fixed timeframe. Real-world switching represents a time-to-event phenomenon with variable onset, duration, and severity that would require survival analysis approaches in clinical implementation.

Treatment Heterogeneity: We did not model specific antidepressant medications, dosing titration schedules, or treatment adherence patterns that may modify switching risk in clinical practice.

Real-World Validation: Priority should be given to validating these models on existing clinical datasets with detailed treatment outcome data. The Systematic Treatment Enhancement Program for Bipolar Disorder (STEP-BD), large electronic health record databases, and international registries provide opportunities for external validation.

Temporal Extension: Development of longitudinal models that predict switching risk as a time-dependent hazard, incorporating changes in clinical status, biomarker trajectories, and treatment modifications during follow-up.

Deep Learning Integration: Exploration of deep learning architectures (recurrent neural networks, attention mechanisms) for sequential clinical data, potentially capturing temporal patterns in illness course that predict switching vulnerability.

Multimodal Fusion: Integration of neuroimaging (fMRI, EEG), genetic, and digital biomarker data with clinical variables to enhance prediction accuracy beyond the clinical-biomarker framework presented here.

Implementation Science: Development of clinical decision support interfaces, workflow integration strategies, and prospective trials assessing the impact of algorithm-guided prescribing on patient outcomes and healthcare utilization.