Multi-modal inflammatory risk modeling in post-PCI patients using behavioral and physiologic data.

Area of Science:

• Cardiovascular Medicine
• Biomarkers and Inflammation
• Digital Health and Wearable Technology

Background:

• Persistent low-grade inflammation post-percutaneous coronary intervention (PCI) is linked to major adverse cardiovascular events (MACE).
• Current risk assessment relies on biomarkers like high-sensitivity C-reactive protein (hs-CRP), underutilizing wearable-derived behavioral data.

Purpose of the Study:

• To develop and validate a multimodal predictive model for sustained inflammatory risk in post-PCI patients.
• The model integrates wearable-derived behavioral data with physiologic biomarkers.

Main Methods:

• Prospective observational study of 312 adult patients undergoing PCI.
• Collected data included electronic health records, inflammatory markers (hs-CRP, IL-6, NLR), and wearable lifelog variables (step count, sleep efficiency, HRV, SpO₂) for up to 6 months.
• Compared four machine learning models (logistic regression, random forest, LSTM, Transformer) for predicting hs-CRP reduction, using SHAP and attention analyses for interpretability.

Main Results:

• Patients with improved inflammation showed significantly higher step count, sleep efficiency, HRV, and SpO₂.
• The Transformer model achieved the highest performance (AUC 0.88, F1-score 0.81).
• SHAP analysis highlighted the strong predictive contribution of modifiable behavioral features.

Conclusions:

• Integrating wearable-derived behavioral and physiologic data enhances prediction of inflammatory outcomes post-PCI.
• Behavioral metrics strongly associate with inflammation, supporting patient-centered, self-regulatory interventions for cardiovascular risk management.