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

## TL;DR

This study shows that combining wearable device data with blood markers can better predict inflammation risk after heart procedures, highlighting the role of daily behaviors like step count and sleep.

## Contribution

The novel integration of wearable behavioral data with physiologic biomarkers improves inflammatory risk prediction in post-PCI patients.

## Key findings

- Patients with improved inflammation had significantly better wearable-derived behavioral metrics like step count and sleep efficiency.
- The Transformer model outperformed other methods in predicting inflammatory outcomes (AUC 0.88).
- Behavioral features were confirmed as strong predictors of inflammation via SHAP analysis.

## Abstract

Persistent low-grade inflammation following percutaneous coronary intervention (PCI) is a known contributor to major adverse cardiovascular events (MACE). While biomarkers such as high-sensitivity C-reactive protein (hs-CRP) are routinely assessed, the predictive role of behavioral factors derived from wearable devices remains underutilized.

This study aimed to develop and validate a multimodal predictive model integrating wearable-derived behavioral data and physiologic biomarkers to assess sustained inflammatory risk in post-PCI patients.

In this prospective observational study, data from 312 adult patients who underwent PCI between January 2022 and December 2024 were analyzed. Data sources included electronic health records, blood-based inflammatory markers (hs-CRP, IL-6, NLR), and continuous wearable-based lifelog variables (step count, sleep efficiency, HRV, SpO₂) collected for up to 6 months. Four machine learning approaches—including logistic regression, random forest, LSTM, and Transformer—were compared for predicting ≥1.0 mg/L reduction in hs-CRP. SHAP and attention weight analyses were used to assess feature importance and model interpretability.

Participants with improved inflammation (59.3%) demonstrated significantly higher step count (8,050 vs. 6,140 steps/day), sleep efficiency (87.1% vs. 78.2%), HRV (64.7 vs. 51.1 ms), and SpO₂ (97.1% vs. 95.2%) compared to non-responders (all p < 0.001). The Transformer model yielded the best performance (AUC 0.88, F1-score 0.81), outperforming other models. SHAP results confirmed the strong predictive contribution of modifiable behavioral features.

Multimodal integration of wearable-informed behavioral and physiologic data enhances the prediction of inflammatory outcomes after PCI. The strong association of behavioral metrics with inflammation supports the development of patient-centered, self-regulatory interventions for long-term cardiovascular risk management.