CareGraph: An Auditable Hybrid AI Framework for Evidence-Grounded Personalized Longitudinal Health Intelligence

Pratik Ghawate; Tanvi Patil

cs.AIAug 25, 2026

CareGraph: An Auditable Hybrid AI Framework for Evidence-Grounded Personalized Longitudinal Health Intelligence

Authors: Pratik Ghawate, Tanvi Patil

Organizations: California, United States

Abstract

Artificial intelligence is transforming personalized healthcare, yet fragmented clinical, self reported, and wearable evidence remains difficult to interpret and trace. We present CareGraph, an auditable hybrid AI framework that converts heterogeneous records into prioritized trends, missing context indicators, bounded next steps, discussion questions, and provenance linked explanations. CareGraph organizes evidence without diagnosing, predicting outcomes, selecting treatment, or making autonomous clinical decisions. Its pipeline covers deterministic analysis, context detection, graph construction, constrained language model synthesis, evidence validation, safety controls, and release gating. Tests used synthetic cohorts of 400 patients each for development, validation, and holdout. On holdout data, a frozen ordinary least squares trend rule with a sufficiency gate achieved 0.827 accuracy, 0.837 macro F1 with a 95 percent confidence interval of 0.819 to 0.854, and 0.974 insufficient data F1. Missing context detection achieved 0.815 strict micro F1 versus 0.318 for the legacy detector. On an authored holdout benchmark, safety ruleset version 1.2 achieved 1.000 precision, 0.950 recall, and 0.974 F1. An audit requiring graph retrieval across 80 patients yielded 79 syntheses and 78 presentations without fallback; one output was blocked and one failed closed because of an invalid evidence key. Against monolithic GPT 5.6 on 56 matched patients, CareGraph was faster at 40.15 versus 49.62 seconds, shorter at 661 versus 1,163 words, and showed better exploratory lexical alignment with longitudinal targets; the baseline used fewer tokens and cited more raw evidence. Graph auditing verified provenance and deterministic retrieval; incremental graph effects on generation require paired evaluation. CareGraph offers a safety bounded foundation for intelligent personalized health systems.

Figures & tables

Figure 1: CareGraph execution, evidence, and release architecture. Deterministic evidence is separated from structured synthesis and from governance. Invalid evidence, missing required graph/AI artefacts, and non-presentable safety outcomes fail closed.

Component	Logic	Input	Output	Boundary
Trend Agent	Deterministic numerical	Labs + scenario inventory	Trend class + sufficiency state	No clinical significance
Missing-Context Agent	Rule-based	Normalised patient domains + scenario	Canonical missing-context items	Scenario-conditioned benchmark
Conflict Resolver	Rule-based	Assembled record	Conflict state	Narrow family-history scope
Knowledge Graph	Deterministic transform	Patient state + upstream evidence	Nodes, edges, ranked evidence	No graph reasoning
Personalised Insight	Deterministic transform	Patient state + trends + context	Structured insight	Heuristic confidence
AI Synthesis	External LLM	Authorised evidence registry	Schema-constrained synthesis	Optional generative layer; no autonomous diagnosis or treatment

Table 1. Component Inventory

Step	Operation	Fail-closed condition or recorded state
1	Load the versioned patient record, manifests, prompt/configuration hashes, and required runtime flags.	Missing source, malformed record, or manifest/hash mismatch.
2	Apply the sufficiency gate, classify trends, detect scenario-conditioned missing context, and resolve supported conflicts.	Emit reason-coded insufficient state; do not silently coerce unusable measurements.
3	Assemble the patient state, construct typed graph nodes/edges, and deterministically rank top- $k$ evidence.	Abort when graph is required but unavailable; retain graph provenance and retrieval metadata.
4	Construct $K_{p}$ and the deterministic insight payload; freeze model, schema, prompt hash, and input hash.	Reject unknown evidence types or an empty authorised registry for required synthesis.
5	Generate schema-constrained synthesis, or use deterministic fallback only when fallback is permitted.	Raise on API/parse failure under --require-ai ; record fallback reason otherwise.
6	Validate patient identity, schema, evidence-key membership, and evidence completeness for every finding and action.	Any invalid key or unsupported required item stops safety and presentation.

Table 3

Cohort	Seed	Patients	Trend labels	Missing/safety labels	Role
Development	20260731	400	1,700	567	Method/rule development
Validation	20260801	400	1,700	587	Selection and freeze verification
Held-out test	20261802	400	1,700	564	One-shot final component evaluation
Excluded nominal test	20260802	400	1,700	—	Not analysed; preview exposure
Safety development	Authored split	—	—	49 examples	Ruleset development
Safety held-out	Locked authored split	—	—	39 examples	One-shot safety evaluation

Table 4

Class	Precision	Recall	F1	Support
Decreasing	0.783	0.863	0.821	351
Increasing	0.826	0.870	0.847	469
Insufficient data	0.949	1.000	0.974	350
Stable	0.767	0.651	0.704	530

Table 3. Held-Out Trend Metrics by Class

Figure 2: Held-out trend F1 by class under the frozen OLS rule. The explicit sufficiency gate yields strong insufficient-data performance; stable classification remains the weakest class.

Figure 3: Development-only threshold selection for OLS. The active threshold was frozen at 0.08 before replacement held-out evaluation.

Detector	TP	FP	FN	Precision	Recall	Micro-F1
Legacy	177	371	387	0.323	0.314	0.318
Version 2	542	224	22	0.708	0.961	0.815

Table 4. Held-Out Strict Missing-Context Results

Figure 4: Strict held-out missing-context performance. Version 2 substantially improves precision, recall, and micro-F1 over the legacy detector.

Ruleset	TP	TN	FP	FN	Precision	Recall	F1
v1.1	2	19	0	18	1.000	0.100	0.182
v1.2	19	19	0	1	1.000	0.950	0.974

Table 5. Held-Out Safety Benchmark

Figure 5: Held-out authored safety benchmark. Ruleset v1.2 improves recall and F1 while retaining zero false positives in the 19 safe examples.

Measure	Value
Patients audited	400
Available graph-evidence items	19,531
Retrieved top-10 items	4,000
Mean nodes per patient	46.0
Mean saved edges per patient	48.8
Mean retrieval fraction	0.207

Table 6. Graph Structural and Provenance Results

Figure 6: Graph provenance, retrieval, and realised citation. Structural lineage metrics are high, but the final-prompt cohort cited no graph-specific evidence.

Study element	Evaluated sample	Evidence in this study	Role in research program
Trend held-out test	1,700 labels	Frozen one-shot evaluation	Core analytical validity
Missing-context held-out test	400 patients / 564 labels	Strict generator-truth evaluation	Core context-detection validity
Safety held-out test	39 examples	Frozen authored benchmark	Bounded safety-rule validity
Graph provenance audit	400 patients	Structural and lineage evaluation	Provenance infrastructure
Condition C deterministic comparison	80 patients	Schema and key-contract evaluation	Deterministic output contract
80-patient graph-required AI batch	80 scheduled	79 syntheses; 78 presentations	Operational reliability and failure analysis

Table 14

Figure 7: Matched generation trade-offs across 56 patients. CareGraph wins most latency pairs and every output-length pair; the monolithic baseline wins every total-token pair and most evidence-reference pairs. Mean ratios use CareGraph as the denominator.

Metric	Monolithic mean	CareGraph mean	Mean diff. M $-$ C	Holm $p$	Rank-biserial
Duration (s)	49.62	40.15	+9.46	$2.74\times 10^{-7}$	0.821
Input tokens	5,172	10,143	$-$ 4,971	$6.79\times 10^{-10}$	$-$ 1.000
Output tokens	4,010	3,033	+976	$6.63\times 10^{-9}$	0.936
Total tokens	9,182	13,176	$-$ 3,994	$6.79\times 10^{-10}$	$-$ 1.000
Output words	1,163	661	+502	$6.79\times 10^{-10}$	1.000
Findings	4.86	4.41	+0.45	$1.37\times 10^{-3}$	0.794

Table 8. Matched 56-Patient Generation Comparison

Category	Threat
Synthetic coupling	Generator scenarios and missingness may favour schema-aligned rules and do not reproduce clinical prevalence, ambiguity, or documentation practice.
Trend construct	The scenario-to-expected-lab inventory remains coupled to the generator schema; stable and selected biomarker results are weaker than directional classes.
Missing-context truth	Strict labels are scenario-conditioned and omit some plausible or contradictory states; current metrics measure agreement with generator truth rather than complete clinical relevance.
Safety benchmark	Only 88 authored examples were available across development and held-out sets; the benchmark has no independent expert annotation and retains one held-out treatment-plan miss.
Operational safety specificity	Medication-change rules matched benign documentation phrases and cross-sentence proximity; generated-output specificity requires sentence-bounded refinement.
Matched comparison cohort	The 56-patient final-prompt cohort represents six scenario groups and is not balanced across all eight; paired results describe this matched subset.

Table 17

Clinical Decision Support

Clinical Decision Support uses data, medical knowledge, and AI to help clinicians assess patients and choose tests, diagnoses, and treatments.

See all 242 papers

May 8, 2026

PerCaM-Health: Personalized Dynamic Causal Graphs for Healthcare Reasoning

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.

Elahe Khatibi, Ziyu Wang, Saba A. Farahani +4

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

Ask PDF

Aug 15, 2026

Framework for Grounding Healthcare LLMs in a Causal Knowledge Graph: A Cardiovascular Example Pilot

Large language models (LLMs) are increasingly proposed for healthcare decision support, but their evaluations still reward single-answer accuracy rather than reasoning about interventions, mechanisms, harms, evidence, and uncertainty. We propose a reproducible, graph-centered evaluation framework for intervention-oriented LLM behavior in healthcare and stress-test it in a cardiovascular pilot. The framework has four components: (i) a domain causal knowledge graph in which assertions are first-class, provenance-preserving nodes with stable identifiers; (ii) a scenario-conditioned subgraph extraction step that, given any clinical scenario, retrieves the relevant reified-assertion subgraph; (iii) four controlled grounding conditions that vary how the retrieved subgraph is composed into the model's context (ungrounded C1, knowledge-graph C2, causal-graph C3, integrated C4); and (iv) an automated scoring pipeline, anchored on assertion identifiers, that computes intervention accuracy, and other evaluation measures on a single pass. To test the framework, we built a category-balanced scenario generator across eight reasoning failure modes and instantiated it on a cardiovascular graph. The metric panel discriminates conditions along interpretable, non-redundant axes: C4 obtains the strongest causal edge F1 (0.838), adverse-effect F1 (0.833), evidence accuracy (0.738), and unsupported claim rate (0.114), while C1 obtains the highest raw intervention accuracy (0.948) with no measurable causal or evidential grounding.

Ummara Mumtaz, Aimen Noor, Awais Ahmed

LLM Grounding Clinical Language Model Evaluation Causal Reasoning in Language Models Biomedical KGs

Ask PDF

Jun 12, 2026

Personal Care Utility: Health as Everyday Infrastructure

Healthcare is essential, expert, and episodic by design - built around the roughly one hour per year a person spends with a clinician. The 8,759 hours outside clinical settings, where eating, sleeping, movement, medication, and stress actually shape long-term health, have no comparable infrastructure. The bottleneck for personalized health is not raw data or reasoning capability; it is the absence of that infrastructure layer. This paper introduces the Personal Care Utility (PCU): a layered, event-driven architecture proposed as the missing utility for everyday health, in the way that payments, networks, and power are utilities for their domains. PCU organizes continuous personal signals into semantically meaningful life events through a Personicle, estimates dynamic health state against personal baselines, reasons about cause and context, and routes guidance through an orchestrator that separates clinical decision logic, behavioral strategy selection, and natural-language expression. This separation lets large language models support reasoning and communication while keeping safety-critical clinical decisions grounded in validated evidence. We instantiate PCU for Type 2 Diabetes - turning CGM, meal, activity, medication, sleep, stress, and clinical data into glycemic events, individualized state estimates, causal explanations, and knowledge-grounded interventions. A day-in-the-life scenario shows the same infrastructure producing real-time nudges, weekly summaries, medication check-ins, silence, or deterministic safety alerts depending on context and risk. We close with how PCU generalizes to other chronic conditions and the governance questions any always-on personal health utility must address. The result is a blueprint that treats personalization not as a final messaging layer, but as an architectural property of everyday health guidance.

Mahyar Abbasian, Elahe Khatibi, Saba A. Farahani +5

University of California, Irvine

Type 1 Diabetes Healthcare Clinical Decision Support

Ask PDF

Abstract

Clinical Decision Support

Explore similar work

PerCaM-Health: Personalized Dynamic Causal Graphs for Healthcare Reasoning

Framework for Grounding Healthcare LLMs in a Causal Knowledge Graph: A Cardiovascular Example Pilot

Personal Care Utility: Health as Everyday Infrastructure