Patient-Centered, Graph-Augmented Artificial Intelligence-Enabled Passive Surveillance for Early Stroke Risk Detection in High-Risk Individuals

This study presents a patient-centered, graph-augmented AI system that leverages patient-reported symptoms to achieve high-precision, passive early stroke risk detection in high-risk individuals, offering a valuable window for clinical intervention while minimizing false alerts.

The Gist

Imagine your body is a house, and a stroke is a sudden, catastrophic fire. Usually, by the time the fire alarm (the emergency room) goes off, the damage is already done. The problem is that many people don't realize the smoke is starting to fill the room until it's too late. They might think the smoke is just from cooking dinner or a drafty window.

This research paper is about building a super-smart, invisible smoke detector that listens to the people living in the house before the fire starts.

Here is the story of how the researchers built this detector, explained simply:

1. The Problem: The "Silent" Warning Signs

People with diabetes are like houses with older wiring; they are at higher risk of a stroke (the fire). But when these patients start feeling weird—maybe a little dizzy, a bit numb, or just generally "off"—they often don't call 911. They might think, "Oh, I'm just tired," or "It's just my diabetes acting up." They send a message to their doctor's online portal instead of calling an ambulance.

The doctors are busy and can't read every single message instantly. So, these tiny warning signs get lost in the noise.

2. The Solution: The "Digital Ear"

The researchers at Stanford decided to use Artificial Intelligence (AI) to listen to these online messages. But they didn't just use a standard computer program. They built a two-part detective team:

• The Translator (LLM): First, they used a Large Language Model (like a super-smart translator) to read thousands of messy, human messages. It figured out what patients were actually talking about. Instead of just seeing the word "dizzy," it understood the context: "I feel like the room is spinning after I stood up."
• The Pattern Hunter (Graph Neural Network): Next, they used a special type of AI called a Graph Neural Network. Imagine a giant spiderweb where every strand connects a symptom to a patient to a timeline. This AI looked at the web and asked: "When did people who eventually had a stroke start talking about these specific things? Was it 3 days before? 30 days before?"

3. The "Symptom Map"

The AI didn't just look for the classic "FAST" signs (Face drooping, Arm weakness, Speech difficulty). It found four hidden patterns that act like early smoke:

1. The "Plumbing" Issues: Messages about blood pressure, heart tests, or blood thinners. (The pipes are getting clogged).
2. The "Storm" Triggers: Messages about colds, the flu, or vaccines. (Infections can stress the heart and brain, like a storm hitting an old house).
3. The "Back-Door" Symptoms: Dizziness, nausea, or weird headaches. (These are often signs of a stroke happening in the back of the brain, which people often ignore).
4. The "Wobbly" Signs: Feeling frail, falling down, or joints hurting. (The house is becoming unstable).

4. The "High-Precision" Alarm System

Here is the clever part. The researchers knew that if their alarm went off too often for things that weren't dangerous, doctors would get annoyed and ignore it (like a smoke detector that goes off when you burn toast).

So, they tuned the system to be extremely conservative. They said: "We would rather miss a few warnings than scream 'Fire!' when there isn't one."

The Result?

• Zero False Alarms: When the system said, "This patient is at high risk," it was 100% correct. Every single time.
• The Early Warning: It could spot these risks 2 to 3 months before the stroke actually happened.
• The Trade-off: Because it was so careful, it didn't catch every single person who would have a stroke (it caught about 7 out of 10). But for the ones it did catch, it gave doctors a massive head start.

The Big Picture: From Reactive to Proactive

Think of this system as a weather forecast for your health.

• Old Way (Reactive): You wait for the hurricane to hit your house, then you call for help.
• New Way (Proactive): The AI sees the clouds gathering and the wind picking up 90 days before the storm. It sends a gentle alert to the doctor: "Hey, this patient's 'weather' is turning stormy. Let's check their blood pressure and meds now, so we can reinforce the roof before the storm hits."

Why This Matters

This study shows that we don't need new, expensive machines to save lives. We just need to listen better to what patients are already saying in their daily messages. By using AI to translate those words into a clear warning, we can give high-risk people a "golden window" of time to get help before a tragedy occurs.

It turns the patient's phone from a simple messaging tool into a life-saving early warning system.

Technical Summary

1. Problem Statement

Stroke remains a leading cause of death and disability, yet prehospital delays are common due to poor symptom recognition, particularly among high-risk populations like individuals with diabetes. Patients often misattribute stroke-like symptoms (e.g., dizziness, numbness) to non-emergency causes (fatigue, neuropathy) or fail to recognize them as emergencies.

• The Gap: While Electronic Health Records (EHRs) contain structured data (labs, vitals), they underutilize unstructured patient-reported data (secure messages/portals) where patients describe symptoms in their own language.
• The Challenge: Patient messages are heterogeneous, noisy, and lack standardization, making it difficult to extract clinically actionable signals for early risk detection using traditional methods.

2. Methodology

The authors developed a Patient-Centered, LLM-Guided Passive Surveillance System using a dual-machine learning pipeline. The study utilized data from Stanford Health Care (SHC) and 22 community partners (2014–2024).

A. Data Source & Cohort

• Population: 71,358 individuals with diabetes.
• Cases: 4,465 patients with diabetes who subsequently experienced an ischemic stroke (ICD-10: I63).
• Controls: 1,947 matched patients with diabetes but no stroke, stratified by age, race, ethnicity, sex, and marital status.
• Input Data: 29,084 secure patient messages from cases and 13,258 from controls.

B. Iterative LLM-Guided Topic Mapping

To structure unstructured text, the team built a hierarchical symptom taxonomy:

1. LLM Bootstrapping: Used Gemini Pro 3 to generate an initial seed set of 10 main categories.
2. Hierarchical Clustering: Streamed messages in batches to create a three-level hierarchy (Main $\to$ Sub1 $\to$ Sub2), resulting in 35 main topics, 109 sub1, and 495 sub2 topics.
3. Validation: Two investigators reviewed LLM reasoning (Gwet's AC1 agreement: 0.93–0.97). The LLM taxonomy was validated against BERTopic (showing ~60% overlap), confirming the LLM's ability to handle noisy, subjective data.

C. Dual ML Pipeline (Graph-Augmented AI)

To identify symptom patterns associated with stroke, the authors employed a hybrid approach:

1. Elastic Net (EN) / LASSO: Used for feature selection to identify statistically significant predictor symptoms.
2. Heterogeneous Graph Neural Network (GNN):
   • Structure: Modeled relationships between patients, symptoms, and events in a tree-like structure.
   • Enhancement: Used Sentence Transformer embeddings to create semantic edges between LLM-labeled symptoms.
   • Architecture: GraphSAGE-based HeteroConv layers optimized via Adam.
   • Goal: Capture complex, non-linear dependencies that linear models (EN/LASSO) might miss.
3. Risk Scoring: Combined Event Association Scores (statistical relevance via GNN delta and EN permutation) and Short-Term Temporal Scores (probability of symptom occurrence within 7 and 30 days of stroke).

D. Hybrid Risk Screening Simulation

• Logic: A hybrid algorithm combined an interpretable symptom-count rule with a logistic regression risk score.
• Optimization: Intentionally tuned for high specificity (>0.90) to minimize false positives and clinician alert fatigue.
• Temporal Windows: Evaluated across 3, 7, 14, 30, 60, and 90-day pre-stroke windows.
• Validation: Internal temporal validation across three distinct time blocks (Cohorts I, II, III).

3. Key Contributions

1. Novel Data Venue: Demonstrated that patient-reported language from secure messages is a valid, high-precision source for early stroke risk detection, complementing structured EHR data.
2. Methodological Innovation: Developed a Graph-Augmented LLM framework that combines the semantic reasoning of Large Language Models with the relational modeling of GNNs to handle heterogeneous, unstructured clinical text.
3. Patient-Centered Taxonomy: Created a hierarchical symptom ontology grounded in patients' own language rather than clinical jargon, capturing under-recognized prodromal signals (e.g., frailty, specific pain patterns).
4. Temporal Granularity: Quantified the temporal proximity of specific symptoms to stroke events, revealing that risk signals can be detected 2–3 months in advance.

4. Key Results

• Symptom Discovery: Identified four high-risk clinical domains:
  1. Classic Vascular Risk: BP management, anticoagulation, cardiac testing.
  2. Acute Biological Triggers: Flu-like symptoms, vaccine concerns, taste/smell loss.
  3. Early Neurological Manifestations: Dizziness/vertigo, nausea/vomiting, migraines (often misdiagnosed).
  4. Frailty & Functional Decline: Joint pain, gait instability, falls, fatigue.
• Performance Metrics (90-day window):
  • Specificity: 1.00 (Perfect; no false positives).
  • Prevalence-Adjusted PPV: 1.00 (Every alert corresponded to a true stroke case).
  • Sensitivity: 0.72 (Captured ~72% of high-risk individuals).
  • Alert Burden: 0.35 (35% of evaluated patients received an alert).
• Temporal Trends: Sensitivity increased with longer windows (0.45 at 7 days vs. 0.72 at 90 days), while specificity remained at 1.00 across all windows.

5. Significance and Implications

• Shift to Proactive Care: The system moves stroke care from reactive (treating after onset) to proactive (intervening based on early digital phenotypes).
• Clinical Utility: By prioritizing precision over sensitivity, the system ensures that when an alert is generated, clinicians can trust it represents a genuine, immediate risk, thereby reducing alert fatigue.
• Intervention Window: The 2–3 month lead time provides a critical window for clinicians to optimize "ABC" (A1C, Blood Pressure, Cholesterol) and for patients to prepare for emergencies.
• Scalability: The framework is designed to be reproducible for other diseases, leveraging existing patient portal infrastructure to create passive surveillance systems without additional patient burden.

Limitations: The study lacked external validation (though internal temporal validation was robust) and relied on patients who actively use secure messaging (though this group is highly representative of the diabetic population at the study site). Future work will focus on clinician workflow integration and patient acceptance of risk alerts.