Chain-of-Influence: Tracing Interdependencies Across Time and Features in Clinical Predictive Modelings

TL;DR

This work tackles the challenge of predicting clinical outcomes from time-series data while maintaining interpretability. It introduces Chain-of-Influence (CoI), a framework that jointly models temporal dynamics and explicit cross-feature interactions through a time-unfolded graph, yielding both accurate predictions and a traceable chain of influence from early features to later outcomes. CoI achieves state-of-the-art AUROC on CKD progression and ICU mortality (0.960 and 0.950, respectively) and demonstrates faithful attributions via deletion-based sensitivity analyses and patient-specific influence visualizations. The approach provides clinically meaningful, transparent narratives of disease progression by mapping how interdependent variables propagate through time, potentially aiding clinical decision-making and trust in predictive models.

Abstract

Modeling clinical time-series data is hampered by the challenge of capturing latent, time-varying dependencies among features. State-of-the-art approaches often rely on black-box mechanisms or simple aggregation, failing to explicitly model how the influence of one clinical variable propagates through others over time. We propose $\textbf{Chain-of-Influence (CoI)}$, an interpretable deep learning framework that constructs an explicit, time-unfolded graph of feature interactions. CoI enables the tracing of influence pathways, providing a granular audit trail that shows how any feature at any time contributes to the final prediction, both directly and through its influence on other variables. We evaluate CoI on mortality and disease progression tasks using the MIMIC-IV dataset and a chronic kidney disease cohort. Our framework achieves state-of-the-art predictive performance (AUROC of 0.960 on CKD progression and 0.950 on ICU mortality), with deletion-based sensitivity analyses confirming that CoI's learned attributions faithfully reflect its decision process. Through case studies, we demonstrate that CoI uncovers clinically meaningful, patient-specific patterns of disease progression, offering enhanced transparency into the temporal and cross-feature dependencies that inform clinical decision-making.

Figures (5)

• Figure 1: CoI architecture overview.
• Figure 2: Temporal and feature-level importance comparison. (a) Temporal attention comparison reveals CoI's focused U-shaped pattern versus RETAIN's uniform weighting. (b-c) Feature importance profiles demonstrate consistent identification of diabetes, laboratory markers (S4, S5), and healthcare utilization as primary predictors across both models.
• Figure 3: Patient-specific Chain-of-Influence visualizations. (a) Top-$3$ influence graph for a single patient, showing the most influential feature–time pairs and their directed paths into the prediction node. (b) Path-centric view anchored at CKD Stage 5 at $t_5$, highlighting the high-influence chain through subsequent visits that drives the final ESRD risk. Additional visualization examples are provided in Appendix \ref{['appendix:coi_visualizations']}.
• Figure 4: CoI architecture overview. Detailed results are provided in Appendix \ref{['appendix:results_ablation']}
• Figure 7: Chain-of-Influence network visualization for CKD progression prediction. Nodes represent clinical features at specific time points (e.g., eGFR_t-5 indicates eGFR at 5 time steps before prediction). Directed edges capture learned temporal dependencies between features, with red edges indicating risk-enhancing influences and blue edges representing protective relationships. The network reveals complex cascading pathways from early cardiovascular events through kidney function decline to healthcare utilization escalation, demonstrating CoI's ability to capture clinically meaningful temporal-feature interactions that drive disease progression.