Peer-Grounded Counterfactual Path Planning for Chronic Health Management

Saman Khamesian; Hassan Ghasemzadeh

cs.LGSep 25, 2026

Peer-Grounded Counterfactual Path Planning for Chronic Health Management

Authors: Saman Khamesian, Hassan Ghasemzadeh

Organizations: College of Health Solutions, Arizona State University, Phoenix, AZ, USA · School of Computing and Augmented Intelligence, Arizona State University, Tempe, AZ, USA

Abstract

Effective behavioral intervention in chronic disease management requires not a single prescription but a sequence of incremental steps, each grounded in what real, similar individuals have demonstrably achieved. Counterfactual explanation offers a natural computational route to such guidance, answering what change in behavior would have produced a better outcome. But existing methods return a target state without a route to it, guarantee no monotone health improvement along the way, and draw no evidence from peer behavior -- asking a patient to close a wide gap in one move, which is precisely the recommendation structure least likely to be attempted. We propose POROS (Peer-Grounded Optimal Routes Over States), a domain-agnostic framework rooted in Bandura's self-efficacy theory and Festinger's social comparison theory that constructs a Behavioral Progression Graph -- a directed acyclic graph over observed patient states in which every edge requires both peer-grounded behavioral proximity and strict health outcome improvement. Every edge is therefore a behavioral change that individuals in the cohort have demonstrated is achievable within a single period. Minimum-cost paths through this graph decompose otherwise inactionable behavioral gaps into incremental, peer-grounded steps. We evaluate POROS on two independent longitudinal cohorts of patients with diabetes. For patients below the 70% clinical threshold for time in range (TIR, blood glucose within 70-180 mg/dL), it reduces the mean gain required per step from 26.3 percentage points (pp) to 5.5 pp on one cohort and from 31.1 pp to 5.7 pp on the other, decomposing large behavioral jumps into the incremental steps that self-efficacy requires. Across both cohorts, 97-98% of multi-hop paths cross patient boundaries, embedding social comparison by construction.

Figures & tables

Figure 1: Traditional counterfactual explanation versus POROS path planning, instantiated on type 1 diabetes management. Traditional methods return a single target state whose behavioral gap violates self-efficacy. In contrast, POROS decomposes the same gap into incremental, peer-grounded steps that embed Bandura’s self-efficacy and Festinger’s social comparison theories.

Figure 2: POROS methodology and the Behavioral Progression Graph it constructs over a study cohort. Each node is one behavioral observation, colored by health outcome from red (poor) to blue (better), positioned by two behavioral features

f_{1}

and

f_{2}

. (1) Node Labeling: observations are labeled relative to a clinical threshold

\tau

. (2) Graph Construction: a directed edge

x_{i}\to x_{j}

is admitted under Eq. ( 1 ), requiring

d(\mathbf{v}_{i},\mathbf{v}_{j})\leq\varepsilon

and strict outcome improvement, so every connection points toward better health. (3) Counterfactual Path: Dijkstra’s algorithm recovers the minimum-cost path from source

x

to target

x^{\prime}

, traversing peer observations as incremental steps. The dashed green arc is the infeasible direct transition (

d>\varepsilon

), which the graph decomposes into a multi-hop peer-grounded path.

Feature	Unit	$\delta_{k}$
Daily carbohydrate intake	grams	5
Carbohydrates per meal	grams	5
Inter-meal interval	minutes	30
Daily insulin dose	units	1
Bolus dose per meal	units	1
Bolus timing offset	minutes	10

Table 1: Behavioral features comprising

\mathbf{v}_{i}

and their MCID values

\delta_{k}

Figure 3: Behavioral Progression Graph over 483 patient-days and 25,556 directed edges (ExActHealth). Node color indicates TIR value from red (low) to blue (high).

	ExActHealth	T1D-UOM
Subjects	18	15
Days	483	854
RED nodes	97 (20.1%)	388 (45.4%)
BLUE nodes	386 (79.9%)	466 (54.6%)
Directed edges	25,556	104,404
Mean out-degree	52.9	122.3

Table 2: Cohort and graph summary.

	$n$ pairs	Direct (pp)	MH/step (pp)	Reduction
ExActHealth
All	111,289	$15.2\pm 12.3$	$3.9\pm 5.2$	3.9 $\times$
RED	37,442	$26.3\pm 13.4$	$5.5\pm 7.0$	4.8 $\times$
T1D-UOM
All	288,887	$23.2\pm 16.6$	$4.8\pm 6.3$	4.9 $\times$
RED	180,808	$31.1\pm 15.5$	$5.7\pm 7.0$	5.5 $\times$

Table 3: Comparison of direct transitions and POROS multi-hop paths. All includes every source node, whereas RED includes only source nodes with TIR

<70\%

. Direct is the TIR difference between source and target; MH/step is the mean TIR gain per step along the minimum-cost path. Reduction

=

mean Direct

\div

mean MH/step.

Case	Condition	ExActHealth ( $n{=}97$ )	T1D-UOM ( $n{=}388$ )	Interpretation
1	$d\leq\varepsilon$	81 (83.5%)	294 (75.7%)	POROS single-hop — both methods agree
2	$d\leq\varepsilon$	10 (10.3%)	53 (13.6%)	POROS multi-hop — $d^{2}$ prefers gradual steps
3	$d>\varepsilon$	4 (4.1%)	24 (6.2%)	Endpoint-only infeasible — POROS routes via peers
4	$d>\varepsilon$	2 (2.1%)	17 (4.4%)	Both methods fail — target behaviorally isolated

Table 4: Endpoint-only counterfactual methods (DiCE, Wachter et al.) versus POROS on the RED

\rightarrow

nearest-BLUE setting. Each RED source is classified by whether the direct jump is feasible (

d\leq\varepsilon

) and whether POROS resolves it as a single- or multi-hop path.

n

is the number of RED source nodes per cohort; cells give the count and share of sources in each case. POROS finds a feasible path for 97.9% (ExActHealth) and 95.6% (T1D-UOM) of RED sources.

Figure 4: Per-feature behavioral change (MCID units) for the RED sources where POROS returns a multi-hop path while endpoint-only methods prescribe a single jump. Dashed contours denote the endpoint-only single jump; solid contours denote the mean POROS change per step. On both cohorts POROS reduces the change demanded of every behavior.

ExActHealth ( $n{=}33{,}632$ )
	FACE	POROS
Infeasible paths ∗ (%)	3.5	0.0
Monotone paths † (%)	9.2	100.0
Paths below source ‡ (%)	22.7	0.0
Hops per path	$6.48\pm 3.77$	$\mathbf{4.90\pm 2.43}$
TIR gain per step (pp)	$4.2\pm 17.0$	$\mathbf{5.6\pm 7.2}$

Table 5: FACE versus POROS on identical source–target pairs. Sources are RED nodes and targets are BLUE nodes with strictly higher TIR. FACE routes over an undirected

k

-NN graph (

k=10

) with density-weighted edges, while POROS routes over the

\varepsilon

-graph with

d^{2}

weights.

n

is the number of pairs both methods reach.

Explore similar work

PerCaM-Health: Personalized Dynamic Causal Graphs for Healthcare Reasoning
May 8, 2026Elahe Khatibi, Ziyu Wang, Saba A. Farahani +4Structural Causal Models Healthcare
University of California, Irvine · Washington University in St. Louis
Personal Care Utility: Health as Everyday Infrastructure
Jun 12, 2026Mahyar Abbasian, Elahe Khatibi, Saba A. Farahani +5Type 1 Diabetes Healthcare
University of California, Irvine
Patient-Centered Treatment Planning for Chronic Multimorbidity: A Hierarchical Reinforcement Learning Framework for Preference Modeling
Sep 30, 2026Nafiseh Payani, Soham Das, G. Anthony Wilson +1Healthcare Hierarchical RL
Department of Industrial and Systems Engineering at the University of Tennessee, Knoxville, TN, USA · Department of Industrial Engineering at the University of Arkansas, Fayetteville, AR, USA

May 8, 2026, cs.LGJ/K move · Enter open · S save

PerCaM-Health: Personalized Dynamic Causal Graphs for Healthcare Reasoning

Elahe Khatibi, Ziyu Wang, Saba A. Farahani, Di Huang+3

University of California, Irvine · Washington University in St. Louis

Structural Causal Models Healthcare Time-Series Causal Discovery Causal Discovery Causal Structure Learning Clinical Reasoning+1

Open paper AskPDF arXiv

Personalized healthcare decisions require reasoning about how physiological and behavioral variables influence an individual patient over time. Existing temporal causal discovery methods are poorly matched to this setting: cohort-level models provide stable but non-personalized structures, while per-patient discovery is unreliable because individual trajectories are short, noisy, irregular, and non-stationary. This creates a fundamental gap between population-level causal modeling and the patient-specific, time-varying mechanisms needed for intervention reasoning. We introduce PerCaM-Health, a framework for learning personalized dynamic causal graphs from longitudinal health data. The framework learns a knowledge-guided population temporal graph, then conservatively adapts and evolves it using patient-specific temporal evidence and rolling-window updates, producing interpretable and auditable graph sequences. By coupling these graphs with temporal structural equations, the framework enables patient-level counterfactual queries, such as estimating short-horizon outcome changes under hypothetical behavioral interventions. Experiments on a semi-synthetic dynamic health benchmark show that PerCaM-Health improves graph recovery, dynamic edge tracking, and intervention direction accuracy compared to cohort-level, per-patient, and non-personalized temporal baselines. These results demonstrate that jointly modeling personalization and temporal evolution yields more reliable causal structure and intervention reasoning.

Abstract

Explore similar work

PerCaM-Health: Personalized Dynamic Causal Graphs for Healthcare Reasoning

Personal Care Utility: Health as Everyday Infrastructure

Patient-Centered Treatment Planning for Chronic Multimorbidity: A Hierarchical Reinforcement Learning Framework for Preference Modeling