2.1. Ethical statement

The experimental protocols employed in this study were reviewed and approved by the Ethics Committee of Ariel University (approval number: AU-ENG-IZ-20220404). All experiments were conducted in accordance with relevant guidelines and regulations. Permission for electrophysiological recordings was granted for the period from February 4, 2022, to February 4, 2023. Written informed consent was obtained from all participants prior to their inclusion in the study.

2.2. Participant recruitment period and process

The recruitment process for this study consisted of two stages. In the first stage, from April 10, 2022 to April 20, 2022, 103 subjects completed the ECR-R [12,13] attachment questionnaires distributed to all volunteers. Based on their responses, 27 participants were selected for the second stage, which involved an EEG session in the laboratory. The EEG data collection took place over multiple days, beginning on May 17, 2022 and concluding on June 1, 2022.

2.3. EEG data collection and epoch extraction

In our investigation into the varying patterns of attachment, we employed a sequential, dual-phase methodology. Initially, we curated a sample of 96 candidates, sourced from the senior-year cohort of an engineering program, aged 20–35 years (mean age = 24.25, standard deviation = 2.0673). Criteria for selection included right-handedness and the absence of any neurological concerns. The attachment typologies for these individuals were determined using the ECR-R questionnaire, a recognized tool in psychological assessments [12,13]. Following this, a k-means clustering algorithm stratified the sample into four attachment typologies—secure, anxious, avoidant, and fearful-avoidant, as delineated in recent research [14–16].

For the EEG phase that followed, 27 subjects were chosen to ensure balanced representation from both the secure and the diverse insecure attachment groups, adhering to proportional quotas reflective of the initial clustering outcomes. The secure group consisted of 6 participants, while the insecure group included 21 participants (9 anxiously attached, 7 avoidant, and 5 fearful-avoidant). Detailed EEG procedural nuances and participant distribution specifics are documented in our recent publication [3].

During EEG monitoring, participants performed a cognitive task known as the flanker task, as detailed in [17]. The task consisted of 60 stimuli, organized into three sets of 20 (see Fig 1). In the middle set, participants were instructed to oppose their responses to the directional cues of arrows, while for the other two sets, they aligned their responses with the cues. One-minute intermissions separated these sets. Performance feedback was instantaneous, marked by a green or red signal, with an exposure of one second. The trials were interspaced with a neutral visual prompt, lasting between half a second to one and a half seconds, prepping the participant for the next stimulus. Each complete set spanned an approximate duration of one minute. An orientation session prefaced the main task to familiarize subjects with the expected actions. In total, we had 1600 epochs, 1240 samples of the insecure classes (anxious, avoidant, and fearful-avoidant) and 360 samples of the secure class.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Experimental paradigm – single block.

https://doi.org/10.1371/journal.pone.0331112.g001

In capturing the EEG data, we utilized a 16-channel EEG apparatus (USBAMP, g.tec, Austria), calibrated for a sampling frequency of 512 Hz and congruent with the 10–20 international positioning framework (see electrodes layout in Fig 2). We concentrated on monitoring the frontal and prefrontal cortical areas, specifically targeting six electrode sites: Fp1, F7, Fp2, F8, F3, and F4. To preserve key neural frequencies associated with cognitive and affective processing while minimizing low-frequency drift and high-frequency artifacts, we applied a bandpass filter set to 1–30 Hz. This range adequately captures alpha (8–12 Hz) and low-beta (13–30 Hz) activity, which prior EEG research has linked to attentional, emotional, and executive processes [18,19]. Following this, an Independent Component Analysis (ICA) was conducted to further segregate genuine neural activity from ocular and muscle artifacts, ensuring the resulting signals are more representative of true cortical processes. After preprocessing, the continuous data were chronologically segmented into one-second epochs, each aligned with the visual cues of the flanker task, thereby allowing for a more precise analysis of moment-to-moment fluctuations in neural activity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Schematic representation of the 16-electrode montage used in this study, following the standard 10–20 EEG system.

https://doi.org/10.1371/journal.pone.0331112.g002

2.4. Feature generation and dimensionality reduction

In our exploration of EEG data for attachment style prediction, we employed the ROCKET algorithm [8], utilizing its default configuration of 10,000 kernels to initially generate a rich feature set of 20,000. Although further kernel-level tuning is possible, we found the standard configuration sufficient, given that our subsequent PCA step retained over 90% of the variance in just 87 principal components, effectively mitigating the need for additional parameter adjustments. We selected ROCKET due to its proven effectiveness in time-series classification, particularly with high-dimensional EEG data, and its ability to capture a broad range of temporal features via randomized convolutional kernels [8,20]. The kernels applied in ROCKET are random and convolutional, allowing them to capture diverse temporal patterns in the EEG signals, including frequency content, transient events, signal complexity, and amplitude variations that are relevant to understanding underlying neural activity. However, because this large number of features exceeds our total number of observations, it raises a high risk of overfitting and complicates the subsequent analysis. To streamline this extensive collection of features and facilitate a more efficient analysis, Principal Component Analysis (PCA) was subsequently applied. This dimensionality reduction process refined the initial feature set down to 87 principal components while preserving over 90% of the variance, thus providing a more efficient input space for our predictive models and balancing the richness of the extracted features with the need for computational tractability and robust model performance.

2.5. Predictive modeling and algorithm

Our predictive framework utilizes the CatBoost algorithm [5], adeptly tailored for dual regression tasks, allowing for the simultaneous prediction of the attachment dimensions ‘avoidance’ and ‘anxiety.’ We selected CatBoost for its robust performance in multi-target regression tasks and its ability to effectively handle correlated label dimensions, making it particularly suited for predicting both ‘avoidance’ and ‘anxiety’ simultaneously (e.g., [21]). To achieve the lowest validation loss, we performed a grid search over hyperparameters including “learning_rate”, “max_depth”, “l2_leaf_reg”, “bagging_temperature”, and “random_strength”, ultimately settling on learning_rate = 0.03, max_depth = 6, l2_leaf_reg = 3, bagging_temperature = 1, and random_strength = 1. Optimization of predictive accuracy was achieved through an RMSE cost function, adept at minimizing spatial distances in the label space and improving the precision of our attachment style predictions.

To fully utilize the dataset and reduce the risk of overfitting, we trained the model using the K-fold cross-validation (K = 3) method [22]. This approach ensured that all data samples contributed to both training and validation at different stages, providing a more generalized evaluation. Importantly, all reported results were obtained from instances where the relevant sample was in the test fold, ensuring unbiased performance assessment.

When using CatBoost for this dual regression task, the following adaptations and considerations were made:

1. a. Dual Regression Modeling: We employed CatBoost’s regression capabilities [6] to perform regression analysis for predicting attachment styles, with avoidance and anxiety serving as the two primary dimensions. This allowed us to model and predict attachment styles as continuous variables, providing insights into the detailed variations of these dimensions within our dataset. The algorithm has been specifically configured and fine-tuned by searching for the best combination of hyperparameters to effectively handle the simultaneous prediction of two continuous variables that are likely correlated with each other.
2. b. Cost Function Customization: In our study, we harnessed the customization capabilities of CatBoost [6,23] to tailor the cost function to our specific needs. Notably, we chose to use multi-RMSE (Root Mean Square Error) as our primary cost function, allowing us to fine-tune the model to minimize the RMSE between predicted and actual values. This customization played a crucial role in shaping the model’s training and refinement, aligning it closely with our research objectives. RMSE serves as a widely employed metric in regression tasks, quantifying the disparities between predicted and actual values. In cases involving multiple target variables for prediction, known as multi-target regression problems, the multi-RMSE is employed as an assessment measure [24]. The multi-RMSE cost function we used is a modification that accounts for the relationship between ‘avoidance’ and ‘anxiety’ (see Equation 1), minimizing errors across both dimensions simultaneously rather than treating them as separate, unrelated predictions.

(1)

Where:

- is the number of observations (epochs in the context of EEG)

- is the number of dimensions (attachment dimensions in this case)

- is the true value of the i^th observation and j^th dimension

- is the predicted value of the i^th observation and j^th dimension

2.6. Iterative refinement and data efficiency

An innovative scheme was applied to refine the EEG epoch data set further, focusing on the most representative epochs by averaging attachment values and selecting those closest to the predictive target. This iterative process aimed to filter out high-variance epochs and concentrate on data that most accurately represented participant attachment styles.

To enhance the accuracy of attachment style prediction using EEG data, we developed a bespoke two-stage algorithm inspired by the core steps in the k-means clustering procedure [25] rather than applying standard k-means clustering. Specifically, our algorithm mirrors k-means in its iterative “distance measurement” and “centroid (mean) update” steps, but it operates on a single evolving mean attachment value (instead of multiple centroids) and prunes outlier epochs at each iteration. As such, we do not perform clustering in the traditional sense; instead, we iteratively refine the dataset to emphasize the most representative epochs.

Here’s a detailed breakdown of our algorithm, assuming we have N EEG epochs

Stage One: Initial Attachment Value Estimation

1. a. Value Calculation for Each Epoch: For each of the N EEG epochs, we calculate a two-dimensional attachment value, representing ‘anxiety’ and ‘avoidance’.
2. b. Mean Attachment Value Computation: We integrate the individual attachment values from all epochs to form an aggregated mean. This mean serves as a provisional centroid akin to the k-means process, representing a collective attachment tendency.

Stage Two: Refinement and Iterative Process

1. c. Distance Measurement: The Euclidean distance from each epoch’s attachment value to the mean attachment value is calculated.
2. d. Selection of Representative Epochs: We refine our epoch set by retaining only the subset comprising the closest half (N/2) of epochs, effectively reducing noise and high-variance outliers.
3. e. Recalculation of Mean Attachment Value: With the pruned subset, we compute a new average attachment value, yielding a more precise representation of the underlying attachment styles.

These two stages loosely parallel the assignment and update steps in k-means, but our approach focuses on refining a single evolving mean attachment value rather than partitioning data into multiple clusters.

2.7. Evaluating algorithm performance across different N values and attachment classes

In this final step of our evaluation, we assess the algorithm’s performance across a range of increasing values of N, exploring aggregate analysis applied within attachment classes. This comprehensive evaluation yields critical insights into the model’s predictive accuracy and its ability to distinguish between attachment styles. Conducting the analysis within each attachment class individually offers a more detailed examination of the model’s performance for specific attachment styles. This granular approach helps to identify potential variations in predictive accuracy among different attachment classes, shedding light on the model’s strengths and weaknesses in distinguishing between specific attachment profiles.

2.8. Assessing robustness with permutation testing

To ensure reliable results for each value of N and participant, we performed 10,000 permutations of epochs for each run of the algorithm. The results were then aggregated for all participants within each attachment class. For example, the secure attachment class consisted of 6 participants, resulting in a total of 60,000 permutations for each value of N (10,000 permutations per participant). This comprehensive approach accounts for all participants within an attachment class, enhancing the model’s performance assessment by reducing the impact of outliers or random variations in individual permutations.

Stage One of our model focuses on establishing a baseline mean attachment value, which serves as an initial approximation. This baseline helps prevent underfitting, where the model might oversimplify and provide inaccurate results. In Stage Two, we undertake an iterative refinement process. During this stage, we systematically assess the robustness of the model by validating its performance across all attachment classes. This iterative refinement, akin to a tuning process, incrementally improves the fidelity of attachment style predictions. By focusing on the most consistent and representative epochs, we minimize the influence of outliers or random variations that could otherwise skew the overall average. This step helps mitigate the risk of overfitting, where the model becomes too specific to the training data and may not generalize well.

2.9. Summary of overall methodology

Fig 3 illustrates the integrated workflow of our study, drawing together the procedures described in Sections 2.1–2.7. In brief, we recruited participants and acquired EEG data (Section 2.1–2.2), extracted and refined features using ROCKET and PCA (Section 2.3), applied CatBoost for initial classification (Section 2.4), and further honed predictions using our iterative refinement algorithm (Section 2.5). Finally, we evaluated robustness and performance across different conditions (Sections 2.6–2.7). This holistic perspective demonstrates how each component of our methodology contributes to improved attachment style prediction accuracy.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Overall methodological pipeline.

https://doi.org/10.1371/journal.pone.0331112.g003

4.1. Dual-stage framework for enhanced prediction

The core of our approach centers around a dual-stage framework aimed at improving prediction precision while streamlining computational aspects. In the first stage, we leverage the CatBoost algorithm [5], known for its predictive strength across domains, to establish baseline attachment style predictions using EEG data. This forms the foundation for our subsequent refinement process, which constitutes the second stage of our model. This refinement process involves data pruning [26] (Luo and Wang, 2018) and employs a data sampling scheme within our framework to extract informative neural patterns while adhering to principles of smart sampling [27].

4.2. Feature extraction and model analysis

For feature extraction, we utilized ROCKET [8] to process the time-series data, and CatBoost [6] was employed to analyze these features. CatBoost’s proficiency in handling datasets with numerous features was crucial in managing the complexity of attachment style data. We used RMSE (Root Mean Square Error) as a cost function [28] to guide model refinement. The number of epochs was set at 11, as determined by the trend in RMSE observed in Fig 5, balancing model accuracy and computational efficiency. This approach aided in distinguishing between different attachment styles, as demonstrated in Fig 6.

4.3. Implications for attachment theory

Our findings hold significance within the context of the dimensional theory of attachment styles [29]. Advanced machine learning techniques allowed us to address the data’s complexity, reflecting the continuous nature of attachment styles. Fig 6 visually underscores the variability in predicting Anxious and Avoidant attachment styles, indicative of broader error ranges in more complex attachment dimensions. These findings support the dimensional perspective of attachment theory, emphasizing that attachment styles span multiple dimensions, enriching our understanding of human attachment and relationships.

Specifically, the Secure cluster exhibits the tightest grouping, indicating the model’s exceptional performance in predicting this attachment style (Fig 6). Conversely, the Anxiously Attached and Avoidant clusters, with larger radii, demonstrate varying prediction errors and data variability, suggesting less consistency and precision. The Fearful Avoidant cluster falls in between, indicating comparable predictive accuracy to the Secure group. These findings harmonize with established literature that posits Secure and Fearful Avoidant attachment occupy opposite ends of the attachment style spectrum [30]. This highlights the complex and continuous nature of attachment styles, supporting the dimensional approach within attachment theory.

4.4. Insights for AI researchers

Our study offers insights for AI researchers working with complex, dimensional data. We present a practical example of using ROCKET for time-series data feature extraction, paired with CatBoost for analysis. The use of RMSE as a cost function for model refinement, particularly in determining the number of epochs, demonstrates a systematic approach to model tuning. Indiscriminate increases in the number of epochs may not significantly enhance model performance, introducing unnecessary complexity without gains in accuracy. The error distribution visualizations in our results can serve as a template for transparently displaying prediction challenges in AI research with complex datasets. These two stages, comprising CatBoost-driven baseline predictions and iterative data refinement, form the cornerstone of our approach to improving attachment style prediction through EEG data analysis, with the use of ROCKET for feature extraction.

4.5. Broader theoretical implications

Our findings add to the growing understanding that attachment styles exist on a spectrum rather than as fixed categories (e.g., [31,32]). The variation in our model’s predictive accuracy across different attachment styles reflects how attachment is shaped by a mix of psychological and neural factors. This supports the idea that attachment isn’t static, it can shift based on life experiences [33], emotional regulation [34], and social context [35]. By applying EEG-based predictive modeling, we provide further evidence that attachment styles, especially anxious and avoidant patterns, have overlapping and complex neural signatures. This reinforces the notion that attachment is a dynamic process rather than a simple label.

4.6. Limitations and future directions

1. Limited Sample Size: Our study’s sample size, especially within the secure attachment group, was limited due to the proportional allocation method used. To ensure broader applicability of the results, future studies should strive for larger and more varied sample sizes.
2. Homogeneity of Participants: We concentrated on university students, potentially limiting the wider applicability of our findings. Future work should seek to incorporate a demographically broader participant cohort, with varying ages, cultural contexts, and life experiences to bolster the universality of the results.
3. Arbitrary Selection of Data Pruning Threshold: One potential limitation in our study lies in the arbitrary selection of the N/2 threshold for data pruning. While this threshold was chosen to balance reliability and precision, it may not be the optimal choice for all datasets and research objectives. To address this limitation and refine our approach, future studies should explore the selection of data pruning thresholds in a more systematic and data-driven manner. This could involve conducting sensitivity analyses with different threshold values to identify the thresholds that maximize model performance, particularly with respect to reliability and precision. Such investigations can contribute to a more evidence-based and robust data pruning process in the context of attachment style prediction through EEG data analysis.
4. Engineering Enhancements: Technologically, it would be advantageous for subsequent studies to investigate the creation of real-time EEG data processing systems to predict attachment styles. Such systems could prove invaluable in clinical settings for immediate interventions. Improving the model’s computational efficiency for use on devices with limited resources would also be beneficial, facilitating wider deployment and access.
5. Longitudinal Analysis: Incorporating a longitudinal study design could provide insights into the stability of attachment styles over time and the model’s predictive power across different life stages.
6. Interdisciplinary Approaches: Combining psychological theory with advanced machine learning techniques can be further explored. For example, integrating findings from neurobiology and psychodynamic theory might refine the model’s predictive capabilities.
7. Cross-Validation with Behavioral Data: To enhance the validity of our model, future studies could incorporate behavioral measures of attachment, providing a multimodal approach to validation and potentially capturing a more holistic picture of attachment patterns.

Addressing these limitations and incorporating the proposed future directions could improve the understanding and prediction of attachment styles. This progress may benefit psychological practices and contribute to more tailored therapeutic approaches.

4.7. Conclusion

In conclusion, our study offers a tangible example of how AI can be harnessed to predict complex psychological constructs with a high degree of accuracy, using EEG data. It underscores the significance of precision and careful model engineering when handling the complex signals typically found in psychological assessments. These two stages, comprising CatBoost-driven baseline predictions and iterative data refinement, form the cornerstone of our approach to improving attachment style prediction through EEG data analysis, with the use of ROCKET for feature extraction.

Our approach—balancing data inclusivity against the necessity for data pruning—reflects an ongoing trend in AI development towards more tailored, specialized, and efficient systems. By detailing the parameters and decision-making processes behind our model, we hope to add a useful perspective to the collective effort of applying AI in complex scientific fields. Our findings may inspire further exploration and cautious application of AI within the realms of neuroscience and psychology, potentially leading to new avenues for research and understanding in these interdisciplinary areas.