Limited discriminative utility of resting-state heart rate variability for chronic pain–depression–anxiety comorbidity subtypes: an exploratory SHAP-based machine-learning analysis

Technical Summary: Limited Discriminative Utility of Resting-State HRV for CPDA Subtypes

Problem Statement
Chronic pain frequently co-occurs with depression and anxiety, forming a complex comorbidity known as Chronic Pain–Depression–Anxiety (CPDA). While clinical evidence suggests that distinct subtypes exist—specifically pain-depression dominant, pain-anxiety dominant, and pain-depression-anxiety triple comorbidity—current clinical assessment relies heavily on subjective symptom reports. There is a lack of established objective physiological indicators capable of differentiating these specific CPDA subtypes. Although Heart Rate Variability (HRV) is a well-documented marker of autonomic nervous system regulation associated with pain and emotional disorders, its specific utility in distinguishing between these heterogeneous CPDA presentations remains uncertain. This study investigates whether resting-state HRV can serve as a reliable biomarker for subtype classification or if it merely reflects a broad, transdiagnostic pattern of autonomic dysregulation.

Methodology
This exploratory cross-sectional study utilized data from 172 adult inpatients with chronic pain recruited from a hospital in Southwest China.

• Participant Stratification: From the total cohort, 82 patients were classified into three mutually exclusive CPDA subtypes based on symptom burdens relative to the sample mean:
  • Pain-depression dominant ($n=8$): High depression, low anxiety.
  • Pain-anxiety dominant ($n=11$): High anxiety, low depression.
  • Pain-depression-anxiety triple comorbidity ($n=63$): High levels of both depression and anxiety.
  • Note: The remaining participants did not meet the specific threshold criteria for these three subtypes and were excluded from the classification models.
• Data Collection:
  • Clinical Metrics: Pain severity/interference (Brief Pain Inventory–Short Form), depressive symptoms (Hamilton Depression Rating Scale, HAMD-24), and anxiety symptoms (Hamilton Anxiety Rating Scale, HAM-A).
  • Physiological Metrics: Resting-state HRV was recorded via a 5-minute non-invasive fingertip sensor. Extracted features included time-domain (e.g., SDNN, RMSSD, pNN50, TINN, Triangular Index) and frequency-domain indices, as well as heart rate metrics.
• Machine Learning Approach:
  • Three weighted binary "one-vs-rest" Random Forest classifiers were constructed to distinguish each specific subtype against the other two.
  • Class Imbalance Handling: Class weights were applied based on inverse frequency to penalize misclassification of minority classes.
  • Validation: Models were evaluated on a held-out test set (stratified 8:2 split) using ROC-AUC, accuracy, confusion matrices, and classification reports.
  • Interpretability: SHAP (Shapley Additive Explanations) analysis was employed to rank feature importance and visualize the contribution patterns of HRV indices, including dependence plots for non-linear relationships.

Key Results
The study found that resting-state HRV demonstrated uneven and generally limited discriminative utility across the CPDA subtypes:

• Pain-Depression Dominant Model: Performed poorly with an AUC of 0.483. Crucially, the model failed to correctly identify any of the 2 positive cases in the test set, misclassifying them as negative.
• Pain-Anxiety Dominant Model: Also showed poor discrimination with an AUC of 0.433. Similar to the depression-dominant model, it failed to identify the 2 positive cases in the test set.
• Triple Comorbidity Model: Showed comparatively higher discrimination with an AUC of 0.712. However, the model exhibited a bias toward positive classification, correctly identifying 13 positive cases but misclassifying 4 negative cases as positive. The authors caution that this performance likely reflects the higher overall comorbidity burden rather than a specific subtype signature.
• Feature Importance (SHAP): The leading HRV features varied by model:
  • Pain-Depression: Max HR, Mean HR, pNN50.
  • Pain-Anxiety: TINN, pNN50, HRV Triangular Index.
  • Triple Comorbidity: HRV Triangular Index, SDNN, CV.
  • SHAP distributions for the minority subtypes were narrow and centered near zero, indicating limited marginal contribution of features to the model output.

Significance and Claims
The paper explicitly claims that resting-state HRV cannot be used as a stand-alone biomarker for the precise classification of CPDA subtypes, particularly for distinguishing between pain-depression dominant and pain-anxiety dominant presentations.

• Primary Conclusion: The inability of HRV to differentiate the single-dominant subtypes suggests that HRV reflects a shared, transdiagnostic autonomic dysregulation common to pain and affective disorders, rather than a disorder-specific signature.
• Nuanced Finding: While the triple comorbidity model showed better exploratory discrimination, the authors attribute this to the detection of a "high-burden" state characterized by severe global autonomic dysregulation, rather than a unique physiological fingerprint for triple comorbidity.
• Methodological Contribution: The study provides a rigorous, SHAP-based evaluation of HRV's limitations in this context, moving beyond simple correlation to test actual classification performance. It highlights the risks of relying on small, imbalanced samples for subtype discovery.
• Future Direction: The authors conclude that HRV may have value as a supplementary screening indicator for high-burden comorbidity but requires validation in larger, balanced cohorts with multimodal indicators (clinical, inflammatory, sleep-related) before it can inform CPDA stratification.

Limitations Acknowledged by the Authors
The study emphasizes several constraints:

1. Subtype Definition: Subtypes were defined using sample-mean thresholds rather than established clinical diagnostic cut-offs, limiting generalizability to formal diagnostic categories.
2. Sample Size: The small number of participants in the pain-depression ($n=8$) and pain-anxiety ($n=11$) groups severely limited statistical power and model stability.
3. Validation: The lack of external validation and reliance on a single train-test split means the reported AUCs are exploratory estimates.
4. Contextual Sensitivity: Short-term resting HRV is sensitive to factors like respiration, posture, and medication, which may reduce its stability as a classifier.