Panic Attack Prediction for Patients With Panic Disorder via Machine Learning and Wearable Electrocardiography Monitoring: Model Development and Validation Study

Figures

• 

  Figure 1.. Prediction output of the transformer model for a single test sample visualization of the model’s predicted values compared with actual labels for the first 100 samples in the test set, based on 24-hour electrocardiogram (ECG)-derived heart rate variability (HRV) data collected from adult participants at Shaoxing People’s Hospital and Ningbo First Hospital (2021‐2023). The prediction distribution is flattened, indicating the transformer’s limited ability to differentiate between normal and abnormal patterns due to class imbalance. This figure highlights the need for improved modeling strategies or data augmentation techniques to address label sparsity in panic attack prediction tasks.

• 

  Figure 2.. Distribution of prediction scores generated by the transformer model. A histogram visualizes the predicted probabilities for each class label across the test set using a transformer model trained on 24-hour electrocardiogram (ECG)-derived heart rate variability (HRV) data from adult participants at Shaoxing People’s Hospital and Ningbo First Hospital (2021‐2023). We initially hypothesized that a transformer-based architecture would be well suited for capturing the complex temporal dependencies inherent in ECG signals, especially given its self-attention mechanism’s ability to model long-range sequential patterns. However, the model failed to learn meaningful class separations, yielding an AUC of 0.50. The histogram shows a concentration of predicted scores around 0.6, with minimal separation between negative and positive classes. This suggests suboptimal feature encoding and ineffective decision boundaries, potentially caused by class imbalance, low temporal resolution in HRV sampling, and insufficient training data for deep sequential models.

• 

  Figure 3.. Receiver operating characteristic (ROC) curve of the transformer model ROC curve generated by the transformer-based classifier, based on electrocardiogram (ECG)-derived heart rate variability (HRV) data collected from adult participants at Shaoxing People’s Hospital and Ningbo First Hospital (2021‐2023). The model achieved an AUC of 0.50, indicating performance at random chance level. This outcome reflects the model’s inability to extract meaningful signal features from the physiological data, possibly due to class imbalance, insufficient temporal sensitivity, or limitations in the feature extraction process.

• 

  Figure 4.. Boxplot of physiological features before and after clipping using interquartile range (IQR) from electrocardiogram (ECG) data collected across two hospitals (Shaoxing and Ningbo, 2021‐2023). Side-by-side boxplots show the distribution of key physiological variables—age, heart rate (HR), root mean square of successive differences (RMSSD), and R-peak count—before and after outlier clipping using the IQR method. Data were collected from adult participants using wearable ECG monitors at Shaoxing People’s Hospital and Ningbo First Hospital between 2021 and 2023. This comparison illustrates the impact of IQR-based preprocessing on improving feature reliability by mitigating extreme values.

• 

  Figure 5.. Distribution of the synthetic feature combining R-peak count and heart rate from electrocardiogram (ECG) data collected across two hospitals (Shaoxing and Ningbo, 2021‐2023). A histogram illustrates the distribution of the engineered feature (r_peak_mean_hr), created to mitigate multicollinearity between R-peak count and mean heart rate. Data were obtained from wearable ECG monitors used by adult participants at Shaoxing People’s Hospital and Ningbo First Hospital. The normalized bell-shaped distribution enhances model interpretability and feature robustness in downstream machine learning applications.

• 

  Figure 6.. Standardized feature distributions after normalization from wearable electrocardiogram (ECG) data collected in two hospitals (Shaoxing & Ningbo, 2021‐2023). Histograms show the distribution of physiological features (age, sex, root mean square of successive difference [RMSSD], target label, r_peak_mean_hr) after applying StandardScaler (mean=0, variance=1). The data were obtained from adult participants using wearable ECG monitors at Shaoxing People’s Hospital and Ningbo First Hospital. This normalization process ensures that features contribute equally during model training, preventing scale-related bias.

• 

  Figure 7.. Model performance comparison by mean squared error (MSE) across six machine learning models trained on physiological features collected from wearable electrocardiogram (ECG) data (Shaoxing and Ningbo, 2021‐2023). Comparison of MSE with standard deviation across regression models: linear regression, decision tree, XGBoost, LightGBM, explainable boosting machine (EBM), and random forest. Models were trained using data collected from adult participants monitored over 24 hours via wearable ECG sensors. Among all models, random forest achieved the lowest MSE, demonstrating superior performance in predicting physiological anomalies linked to panic attacks.

• 

  Figure 8.. Hyperparameter optimization history of the random forest model on electrocardiogram (ECG) data from Shaoxing and Ningbo hospitals (2021‐2023). Line plot shows iterative improvement of the random forest model’s performance during hyperparameter tuning using a Gaussian optimizer. Data were derived from wearable ECG monitoring conducted across Shaoxing People’s Hospital and Ningbo First Hospital between 2021 and 2023. The graph illustrates convergence behavior toward the optimal parameter set by tracking mean squared error (MSE) across trials.