Research and Practice in Couple Therapy

The objective of this study was to develop and evaluate an artificial intelligence–based model capable of predicting intimacy decline in romantic couples using dyadic attachment orientations and individual and shared stress patterns. The study employed a longitudinal, correlational design with a predictive modeling framework. Married and long-term cohabiting couples from Turkey participated as dyads, with both partners independently completing validated self-report measures of attachment anxiety and avoidance, perceived individual stress, dyadic stress, and relational intimacy. Data were collected at baseline and at a six-month follow-up to capture changes in intimacy over time. After data preprocessing and dyadic feature construction, multiple supervised machine learning algorithms were trained and validated using cross-validation procedures to predict intimacy decline as both a categorical and continuous outcome. The predictive models demonstrated strong performance, with ensemble-based algorithms achieving the highest classification accuracy and area under the curve values in distinguishing couples with declining intimacy from those with stable intimacy. Inferential feature analyses indicated that attachment anxiety, attachment avoidance, dyadic stress, and interaction terms between attachment insecurity and stress were the most influential predictors. Models incorporating dyadic discrepancy indicators consistently outperformed those based solely on individual-level features, indicating significant partner interdependence effects. Higher combined levels of attachment insecurity and stress were associated with greater magnitude of intimacy decline over time. The findings indicate that intimacy decline can be accurately predicted using AI-based models that integrate dyadic attachment and stress variables, supporting the view that intimacy erosion emerges from complex, interactive relational processes. These results highlight the potential of artificial intelligence to inform early identification and prevention strategies in couple and family interventions.

The objective of this study was to develop and evaluate a machine learning–based early warning system capable of predicting therapy failure trajectories among high-conflict couples during the early phases of couple therapy. This longitudinal observational study was conducted with high-conflict couples undergoing outpatient couple therapy in Germany. Multimodal data were collected from both partners and therapists across early and mid-treatment sessions, including self-reported relational functioning, emotional regulation indicators, therapeutic alliance ratings, therapist session evaluations, ecological momentary assessments, and automated interactional features derived from session recordings. Therapy failure was operationalized as premature dropout, therapist-rated non-response, or reliable deterioration in relationship satisfaction over time. Multiple machine learning models, including regularized logistic regression, random forest, gradient boosting, and recurrent neural networks, were trained using longitudinal features capturing both static baseline characteristics and dynamic process indicators. Model performance was evaluated using cross-validated inferential metrics emphasizing early detection accuracy. Inferential analyses demonstrated that all machine learning models significantly outperformed chance-level prediction, with recurrent neural network models yielding the highest discriminative accuracy and sensitivity for early therapy failure detection. Dynamic process variables, particularly early-session therapeutic alliance variability, escalating conflict trajectories, emotional spillover volatility, and dyadic interaction asymmetries, showed statistically stronger predictive contributions than baseline relational characteristics. The early warning system successfully identified a substantial proportion of therapy failure cases within the first four therapy sessions, indicating robust temporal predictive validity. The findings indicate that therapy failure in high-conflict couples follows identifiable dynamic patterns that can be detected early using machine learning approaches. Implementing early warning systems in couple therapy may enable proactive, adaptive interventions that reduce dropout and non-response, thereby improving therapeutic outcomes for high-conflict couples.

The objective of this study was to examine whether deep neural network–derived indicators of emotional synchrony predict marital satisfaction among married couples in the United States. This study employed a cross-sectional, observational design with a correlational–predictive framework. A total of 286 married couples residing in the United States participated in the study. Couples completed a standardized measure of marital satisfaction and engaged in a structured dyadic interaction task during which facial expressions and vocal signals were recorded. Multimodal emotional data were extracted from synchronized facial and vocal streams and transformed into dynamic emotional synchrony indices capturing both concurrent and lagged emotional alignment between partners. These indices were used as inputs to deep neural network models specifically designed to analyze dyadic temporal data. Model training, validation, and testing were conducted at the couple level to prevent data leakage, and explainable artificial intelligence techniques were applied to identify the most influential emotional features contributing to prediction accuracy. Deep neural network models demonstrated that overall emotional synchrony significantly predicted marital satisfaction, with multimodal models outperforming unimodal models. Models incorporating both facial and vocal synchrony explained a substantially greater proportion of variance in marital satisfaction than models based on static or single-modality features. Positive affect synchrony emerged as the strongest predictor, while negative affect synchrony showed a significant inverse association with marital satisfaction. Temporal models capturing dynamic emotional alignment significantly outperformed static models, indicating the critical role of time-dependent emotional processes in marital relationships. The findings provide compelling evidence that emotional synchrony, particularly dynamic coordination of positive emotions, is a robust predictor of marital satisfaction and highlight the value of deep neural network approaches for advancing the study of emotional processes in intimate relationships.

The objective of this study was to develop and interpret explainable artificial intelligence models capable of forecasting divorce risk based on dyadic communication patterns among married couples. This quantitative, observational study was conducted with legally married couples residing in Canada. Couples completed validated self-report measures assessing communication quality, emotional responsiveness, and relational characteristics, and participated in a structured dyadic interaction task designed to elicit naturally occurring conflict-related communication. Interaction transcripts were processed using natural language processing techniques to extract linguistic and interactional features reflecting positivity, negativity, contempt, emotional validation, and conversational balance. A composite divorce risk indicator was constructed from self-reported divorce proneness and separation intentions. Multiple supervised machine learning models, including regularized regression and tree-based ensemble methods, were trained and evaluated using nested cross-validation. Explainable artificial intelligence techniques were applied to identify global and local feature contributions to model predictions. Ensemble-based models demonstrated significantly higher predictive performance than linear models, achieving superior accuracy and area under the receiver operating characteristic curve. Negative communication features, particularly contempt markers and overall communication negativity, were the strongest positive predictors of divorce risk, while emotional validation and balanced turn-taking showed significant protective effects. Demographic variables contributed comparatively less to prediction once dyadic communication patterns were included. Explainability analyses revealed consistent and interpretable pathways through which specific interactional behaviors increased or reduced predicted divorce risk. The findings indicate that explainable artificial intelligence models can accurately and transparently forecast divorce risk using dyadic communication patterns, highlighting communication behaviors as central, modifiable indicators of marital instability.

This study aimed to examine and classify trauma-related emotional patterns in couples facing PTSD using multimodal deep learning models to predict high-risk emotional episodes during conflict interactions. The study employed a cross-sectional design involving 112 couples (224 individuals) from the United States in which one partner met diagnostic criteria for PTSD. Each couple participated in a 40-minute conflict-discussion task recorded through synchronized audio, video, and physiological sensors. Linguistic transcripts, facial expressions, acoustic features, and autonomic indicators (electrodermal activity, heart-rate variability, and peripheral temperature) were extracted and temporally aligned. These modalities were analyzed using a multimodal transformer architecture integrating text embeddings, CNN-LSTM visual features, physiological time-series data, and acoustic spectral representations. Additional analyses included t-SNE–based latent clustering of emotional patterns and risk-surface modeling of predicted high-risk episodes as a nonlinear function of physiological arousal and negative emotional language. Model performance was evaluated using accuracy, precision, recall, F1-score, and AUC metrics with ten-fold stratified cross-validation. The multimodal transformer significantly outperformed baseline and unimodal models in predicting high-risk PTSD-related emotional episodes (AUC = .90; F1 = .83; accuracy = .86; p < .001 vs. clinical baseline). All unimodal models performed above chance but remained significantly weaker than the multimodal approach (p < .01). Latent clustering revealed four statistically distinct emotional interaction patterns (hyperaroused escalation, avoidant disengagement, mixed volatile–repairing, and numbed coexistence), each differing significantly in PTSD severity, relationship quality, and proportion of high-risk segments (all ps < .05). Nonlinear risk-surface analysis demonstrated a strong interaction effect between negative emotional language and physiological arousal in predicting emotional risk (βinteraction = .41, p < .001). Multimodal deep learning provides a highly sensitive and integrative method for identifying trauma-related emotional risk states in couples facing PTSD, revealing clinically meaningful emotional clusters and nonlinear escalation patterns that may inform assessment and intervention.

This study aims to identify the most influential psychological, communicative, and digital-behavioral predictors of relationship burnout in couples using machine learning feature-importance analysis. A cross-sectional predictive design was implemented with a sample of 206 couples (412 individuals) from Turkey, collected through online recruitment. Participants completed validated measures assessing emotional exhaustion, communication avoidance, stress and emotional dysregulation, relational motivation, conflict behaviors, sexual and relational satisfaction, and digital interaction patterns such as shared online activity and response latency during conflictual exchanges. Data preprocessing included normalization, encoding, missing-value correction, and outlier management. Multiple machine learning models—random forests, gradient-boosted trees (XGBoost), multilayer perceptron networks, support vector machines, and logistic regression—were trained on a stratified 80/20 train–test split. Model performance was evaluated using accuracy, precision, recall, F1-score, and AUC. Feature importance was assessed using permutation importance, model-specific variable importance scores, and SHAP (SHapley Additive exPlanations) values to identify consistent early warning indicators. XGBoost achieved the highest predictive performance (Accuracy = 0.89, AUC = 0.94), followed by random forests (Accuracy = 0.86, AUC = 0.91). SHAP analysis revealed emotional exhaustion as the strongest predictor, followed by communication avoidance, response latency during conflict, emotional dysregulation, weekly conflict episodes, relational motivation, and digital disengagement ratio. Interaction effects showed that high emotional exhaustion combined with high communication avoidance produced a multiplicative increase in predicted burnout probability, confirming nonlinear relational deterioration patterns captured by the models. Machine learning modeling effectively identified early warning signs of relationship burnout, demonstrating that emotional, communicative, and digital-behavioral variables jointly predict relational decline. These findings highlight the need for integrating computational analytics into clinical screening and preventive relationship interventions.

This study aimed to identify and compare generational differences in the predictors of relationship stability using a large-scale machine-learning framework applied to adults in the United States. This cross-sectional study analyzed data from 4,812 adults representing Baby Boomers, Generation X, Millennials, and Generation Z. Participants completed a multidimensional online assessment capturing psychological, relational, socioeconomic, and digital-behavioral variables. Quantitative scales measured communication clarity, emotional support, attachment anxiety, conflict recovery time, financial stress, and additional relational factors, while open-ended items provided qualitative textual data. Preprocessing included imputation, normalization, categorical encoding, and NLP-based embedding of narrative responses. Machine-learning models—including gradient-boosted trees and random forest algorithms—were used to predict relationship stability, with performance evaluated via accuracy, AUC, precision, recall, and F1-scores. SHAP analysis was conducted to interpret feature importance globally and within generational subgroups. Machine-learning models achieved strong predictive performance across generations (AUC range: 0.84–0.92). SHAP values revealed significant generational differences: communication clarity was the strongest predictor for Baby Boomers and Generation X, financial stress and attachment anxiety were dominant predictors for Millennials and Gen Z, and digital comparison exposure showed sharply increasing influence from older to younger cohorts. Relationship length exhibited high predictive value for older generations but minimal influence among younger adults. Across all models, higher emotional support and shorter conflict recovery time significantly increased predicted relationship stability, while elevated financial stress and attachment anxiety significantly reduced stability probabilities. Generational cohorts differ markedly in the factors that predict relationship stability, with younger adults exhibiting heightened sensitivity to financial strain, emotional insecurity, and digital comparison pressures. Machine-learning modeling reveals that relationship stability is not governed by universal predictors but instead emerges from generationally distinct psychological, socioeconomic, and technological influences.

This study aimed to determine whether natural language patterns extracted from early couple therapy sessions can accurately predict premature therapy dropout. The study used a mixed quantitative–qualitative design involving 148 couples from Canada who participated in community, private, and university-based couple therapy settings. High-quality audio recordings of the first two therapy sessions were transcribed and preprocessed using advanced natural language processing techniques, including tokenization, lemmatization, turn-level segmentation, and affective, syntactic, and semantic feature extraction via transformer-based models. Communication variables such as emotional disengagement, interruption frequency, partner-focused pronoun use, and demand–withdraw sequences were quantified. Machine-learning models—logistic regression, random forest, gradient boosting, Bi-LSTM networks, and transformer architectures—were trained to predict dropout, defined as termination prior to session four without therapist-approved discontinuation. Model performance metrics included accuracy, precision, recall, F1 score, and ROC-AUC. SHAP values were used to interpret model-level decision patterns. Reflexive thematic analysis of therapist notes complemented quantitative findings to contextualize communication deficits. Inferential analyses revealed significant differences between dropout and treatment-completion groups across multiple linguistic variables, including higher negative affect, lower partner-focused pronouns, greater interruption frequency, and elevated demand–withdraw cycles among dropout couples (all p < .001). Transformer-based models achieved the strongest predictive accuracy (92%) and highest ROC-AUC (0.96), outperforming all traditional and neural baselines. SHAP interpretability demonstrated that emotional disengagement markers, interruption frequency, topic abruptions, and conversational asymmetry were the most influential predictors of dropout. Communication reciprocity declined over time in dropout couples, whereas it increased in treatment completers. Early-session communication deficits captured through natural language processing serve as powerful predictors of premature dropout in couple therapy. Incorporating automated linguistic assessment tools into routine clinical practice may enable earlier identification of at-risk couples and support targeted intervention strategies to reduce attrition.