🤖 machine learning

PerCaM-Health: Personalized Dynamic Causal Graphs for Healthcare Reasoning

PerCaM-Health is a novel framework that bridges the gap between population-level and patient-specific causal modeling by learning personalized, time-varying causal graphs through the conservative adaptation of a knowledge-guided population structure, thereby enabling more reliable counterfactual reasoning for individualized healthcare interventions.

Original authors: Elahe Khatibi, Ziyu Wang, Saba A. Farahani, Di Huang, Hung Cao, Ramesh Jain, Amir M. Rahmani

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Elahe Khatibi, Ziyu Wang, Saba A. Farahani, Di Huang, Hung Cao, Ramesh Jain, Amir M. Rahmani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "One-Size-Fits-All" vs. "Too Much Noise" Dilemma

Imagine you are trying to figure out what makes a specific person's blood sugar go up or down. You have two bad options right now:

The "Group Average" Map: You look at a map of how everyone behaves. This map is very clear and stable because it's based on thousands of people. But it's too general. It might say, "For most people, eating sugar raises blood sugar." But for your specific patient, maybe it's actually lack of sleep that's the main culprit. The group map misses your unique story.
The "Solo Detective" Map: You try to draw a map using only your patient's data. The problem? A single patient's history is short, messy, and full of gaps (like a torn-up diary). If you try to build a map from just that, you'll get it wrong because there isn't enough evidence.

PerCaM-Health is a new tool that solves this by acting like a smart GPS that updates in real-time.

How PerCaM-Health Works: The Three-Step Journey

The paper describes a framework that builds a "Personalized Dynamic Causal Graph." Let's break that down into three simple steps:

1. The "Group Blueprint" (The Population Prior)

First, the system looks at a huge group of people to build a master blueprint.

Analogy: Think of this as a standard architectural plan for a house. It knows that usually, the kitchen connects to the dining room, and the bedroom connects to the hallway. It's based on what works for everyone.
What it does: It uses medical knowledge and data from many patients to create a "safe" starting point. It knows which variables could possibly influence each other (e.g., stress could affect glucose).

2. The "Custom Renovation" (Patient Adaptation)

Next, the system takes that master blueprint and tweaks it for one specific patient.

Analogy: Imagine you are the architect. You look at the standard house plan, but then you look at your specific client. You see that for this client, the kitchen is actually too far from the dining room, so you move a wall. Or maybe for this client, the bedroom is the main hub, not the living room.
What it does: It doesn't throw away the group blueprint (because that's too risky). Instead, it conservatively adjusts the connections. If the group says "Sleep affects Stress," the system checks the specific patient's data. If that patient's data strongly supports it, the connection gets stronger. If the data is weak, the connection stays weak. It uses the patient's short, noisy history to fine-tune the big map.

3. The "Live Traffic Updates" (Rolling Dynamic Evolution)

Finally, the system realizes that people change over time. A patient might be stressed one month and relaxed the next.

Analogy: A static map is useless if there's a road closure. PerCaM-Health is like a GPS that uses rolling windows. It looks at the last two weeks of data, updates the map, then looks at the next two weeks, and updates it again.
What it does: It creates a movie of the causal graph instead of a single photo. It tracks how the relationships between variables (like sleep and glucose) strengthen or weaken as the patient's life changes.

What Can You Do With This? (The "What If" Engine)

Once the system has built this personalized, time-changing map, it can answer "What if?" questions.

The Analogy: Imagine you are driving. The GPS doesn't just show you where you are; it simulates the future. It says, "If you take this detour, you'll save 10 minutes. If you take that one, you'll hit traffic."
The Paper's Claim: PerCaM-Health can simulate counterfactuals. It can tell a doctor: "If this specific patient increases their walking by 1,000 steps tomorrow, our model predicts their glucose will drop by X amount."
Important Note: The paper is very careful to say this is a structured hypothesis, not a guaranteed medical fact. It's a "best guess" based on the data and the rules of the map, meant to help doctors think through options, not replace a real clinical trial.

Did It Work? (The Results)

The researchers tested this on a simulated health world (a computer-generated dataset where they knew the "true" answers because they wrote the rules).

The Competition: They compared PerCaM-Health against:
- Old methods that only look at groups (too generic).
- Old methods that only look at individuals (too noisy).
- Methods that know the rules but don't adapt to individuals.
The Winner: PerCaM-Health won. It was better at:
1. Finding the right connections: It figured out the true cause-and-effect links more accurately than the others.
2. Tracking changes: It was better at noticing when a relationship changed over time.
3. Predicting interventions: When asked "What happens if we change X?", it gave the most accurate prediction of the direction of the change.

The Bottom Line

PerCaM-Health is a tool that says: "Don't just guess based on the crowd, and don't just guess based on the messy individual. Use the crowd's wisdom as a safety net, then carefully adjust it for the individual, and keep updating it as time goes on."

The paper claims this approach creates a more reliable, interpretable, and auditable way to understand how a specific patient's health variables interact over time, helping to generate better "What if?" scenarios for healthcare decisions.

Technical Summary: PerCaM-Health

Problem Statement

Personalized healthcare decisions require reasoning about how physiological and behavioral variables influence an individual patient over time. However, existing temporal causal discovery methods face a fundamental mismatch with this setting:

Cohort-level models provide stable causal structures but fail to capture patient-specific heterogeneity.
Per-patient discovery is statistically unreliable because individual longitudinal trajectories are often short, noisy, irregular, and non-stationary.

This creates a gap between population-level causal modeling and the patient-specific, time-varying mechanisms required for effective intervention reasoning. Current healthcare predictors are largely associational, lacking the explicit causal mechanisms needed for counterfactual queries (e.g., "How would a patient's glucose change if their sleep improved?").

Methodology: PerCaM-Health

The authors propose PerCaM-Health, a framework for learning personalized dynamic causal graphs from longitudinal health data. The method operates in five modular stages:

Patient-State Construction: Heterogeneous health data (structured records, wearables, logs, etc.) are aggregated into patient-state trajectories ( $x_{i,t}$ ) containing measured and clinically inferred variables (e.g., stress, adherence, glucose).
Knowledge-Guided Population Graph Learning: A cohort-level directed graph is learned using lagged prediction, sparsity constraints, and a smooth acyclicity penalty (based on NOTEARS/DYNOTEARS principles). Crucially, this stage incorporates clinical priors (soft masks) to encourage plausible dependencies and penalize forbidden directions, establishing a "causal scaffold."
Patient-Specific Graph Adaptation: Rather than learning a graph from scratch for each patient, the framework conservatively adapts the population scaffold. It reweights population-supported edges using patient-specific temporal evidence via graph-constrained autoregressive regression. This ensures the graph remains anchored to stable population structures while reflecting individual deviations.
Rolling-Window Dynamic Evolution: To handle non-stationarity (e.g., changing sleep or stress patterns), the personalized graph evolves over time. Using rolling temporal windows, the framework estimates time-varying conditional effects and applies exponential smoothing to update edge weights, capturing regime shifts within a patient's trajectory.
Intervention Queries and Counterfactuals: The resulting sequence of dynamic graphs is coupled with temporal structural equations. This enables model-based counterfactual queries, such as estimating short-horizon outcome changes under hypothetical behavioral interventions (e.g., increasing activity or reducing stress).

Key Contributions

The paper outlines three primary contributions:

Formulation: It formulates personalized dynamic causal graph learning as a problem of population-prior adaptation for longitudinal healthcare data, bridging the gap between stable cohort models and fragile per-patient discovery.
Mechanism: It introduces conservative patient-level and rolling-window updates. These mechanisms personalize graph weights using local temporal evidence while maintaining cohort-level support, preventing overfitting to short, noisy trajectories.
Evaluation: It conducts a controlled semi-synthetic evaluation with known dynamic causal mechanisms. This allows for direct assessment of graph recovery, patient-specific adaptation, temporal edge tracking, and intervention-direction accuracy under heterogeneous and non-stationary conditions.

Experimental Results

The framework was evaluated on a semi-synthetic benchmark containing 120 patients with 60 time steps each, featuring known ground-truth causal pathways (e.g., sleep $\to$ stress $\to$ glucose) and regime shifts.

Graph Recovery: PerCaM-Health achieved superior performance compared to static, temporal, dynamic, and knowledge-guided baselines, with a Precision of 0.92, Recall of 0.97, F1 of 0.94, and the lowest Structural Hamming Distance (SHD) of 0.85.
Dynamic Tracking: The framework demonstrated strong ability to track time-varying causal strengths, achieving a direction accuracy of 0.91 and temporal correlation of 0.80, outperforming population-level dynamic methods and naive rolling correlations.
Intervention Reasoning: In counterfactual queries, the method achieved 0.96 intervention-direction accuracy and the largest mean glucose reduction (-0.31) among tested methods.
Ablation Study: Removing the population prior degraded graph recovery; removing personalization caused the largest drop in structural recovery; and removing dynamic evolution eliminated temporal tracking capabilities. This confirms that each component (anchoring, personalization, and evolution) addresses distinct failure modes.

Significance and Claims

The paper claims that jointly modeling personalization and temporal evolution yields more reliable causal structures and intervention reasoning than existing approaches. By anchoring patient-specific graphs to a knowledge-guided population scaffold, PerCaM-Health produces interpretable and auditable graph sequences.

The authors emphasize that the resulting outputs are structured counterfactual hypotheses grounded in longitudinal data and explicit assumptions, rather than definitive clinical effects. The framework is designed to support hypothesis generation and downstream intervention analysis, helping clinicians identify which variables drive a specific patient's risk and how mechanisms vary across individuals and time. The work does not claim to replace randomized clinical trials but offers a method to generate testable, patient-specific mechanistic insights from observational data.