Inflammatory Biomarkers & Interpretable ML for SAP Risk Stratification in AIS Patients Undergoing Bridging Therapy

This study developed an interpretable machine learning model integrating inflammatory biomarkers that outperforms traditional clinical methods in accurately stratifying the risk of stroke-associated pneumonia among acute ischemic stroke patients undergoing bridging therapy, thereby enabling early identification and targeted intervention.

The Gist

Imagine your brain is a bustling city. When a stroke happens, it's like a major traffic jam blocking a key highway (an artery), causing parts of the city to shut down. Doctors have a "rescue plan" called Bridging Therapy: they first send in a chemical cleanup crew (clot-busting drugs) and then follow up with a mechanical tow truck (a tiny device to physically pull the clot out). This saves the city from total collapse.

However, there's a sneaky villain that often shows up after the rescue: Stroke-Associated Pneumonia (SAP). Think of this as a "ghost infection" that sneaks into the lungs of patients while they are recovering. It's dangerous, hard to spot early, and can undo all the good work the rescue team just did.

This paper is like a team of detectives trying to build a super-smart crystal ball to predict who is most likely to get this lung infection, so doctors can protect them before it even happens.

Here is how they did it, broken down into simple stories:

1. The Clues: Inflammatory "Smoke Signals"

When the brain is injured, the body's immune system goes into overdrive, like a city putting all its fire trucks on high alert. This creates "smoke signals" in the blood. The researchers looked at specific types of smoke:

• NLR, PLR, SII, and SIRI: These sound like complicated acronyms, but think of them as scorecards that measure the balance between the body's "attackers" (white blood cells) and its "peacekeepers" (lymphocytes).
• The Discovery: They found that patients who got pneumonia had much higher "smoke signals" (inflammation) in their blood at 24 and 48 hours after the stroke compared to those who stayed healthy. It's like seeing a massive fire alarm go off in the lungs before the fire is even visible on an X-ray.

2. The Problem with Old Maps

Previously, doctors used simple checklists (like asking "Is the patient older?" or "Do they have a fever?") to guess who might get pneumonia. But these checklists are like using a paper map in a GPS world. They are often too slow or rely on things that are hard to see until it's too late.

3. The Solution: The "Super-Brain" (Machine Learning)

The researchers decided to build a digital detective using Machine Learning (AI).

• The Training: They fed this AI the medical records of 135 patients who had undergone the "Bridging Therapy." They showed the AI 63 different clues (age, blood sugar, stroke severity, blood test results, etc.).
• The Filter: The AI used a tool called LASSO to act like a sieve, filtering out the noise and keeping only the 11 most important clues.
• The Champion: They tested 10 different types of AI "detectives." One of them, called CatBoost, was the clear winner. It was like finding a detective who could see patterns the others missed.
  • The Score: This CatBoost model was incredibly accurate, scoring a 95% on its practice test and 93% on its real-world test. That's like a student getting an A+ on a difficult exam.

4. Making the "Black Box" Transparent

Usually, AI is a "black box"—you put data in, and a result comes out, but you don't know why. The researchers used a special tool called SHAP (which sounds like "shapley," a math concept) to open the box and show the detective's notebook.

What did the notebook say?
The AI explained that the top three reasons it predicted pneumonia were:

1. NIHSS_7d: How bad the stroke was after 7 days (the longer the brain struggles, the higher the risk).
2. SIRI_24h: The "smoke signal" score at 24 hours (high inflammation = high risk).
3. WBC_24h: The white blood cell count at 24 hours (the body's army is already fighting a battle).

It's like the AI saying: "I'm worried about this patient because their brain is still struggling, and their blood is screaming that there's a massive inflammation battle happening right now."

5. Why This Matters

This study is a game-changer because it moves medicine from reacting to preventing.

• Before: Doctors wait until a patient has a fever or a cough to treat pneumonia. By then, the infection is already established.
• Now: With this "Crystal Ball," doctors can look at a patient's blood test 24 hours after the stroke, run it through the model, and say, "This patient is high-risk."
• The Result: They can then give extra care, monitor the lungs more closely, or adjust treatment before the pneumonia takes hold.

The Bottom Line

The researchers built a high-tech, easy-to-understand warning system. By combining simple blood tests (the smoke signals) with a smart computer program (the digital detective), they can spot the invisible threat of pneumonia early. This helps doctors save more lives and ensure that the "rescue mission" for stroke patients isn't sabotaged by a secondary infection.

In short: They taught a computer to read the body's early warning signs, turning a guessing game into a precise science.

Technical Summary

1. Problem Statement

Stroke-associated pneumonia (SAP) is a severe complication in patients with Acute Ischemic Stroke (AIS) who undergo bridging therapy (intravenous thrombolysis followed by mechanical thrombectomy). SAP significantly worsens prognosis, increases mortality, and prolongs hospital stays.

• Current Limitations: Existing clinical prediction tools (e.g., A2DS2, AIS-APS) rely heavily on subjective clinical symptoms, which can delay early identification.
• The Gap: While inflammatory biomarkers (like NLR, PLR, SII, SIRI) show promise, their complex interactions with clinical factors are often not captured by traditional statistical models. Furthermore, there is a lack of interpretable machine learning (ML) models specifically designed for this high-risk bridging therapy population to guide personalized antibiotic prophylaxis and treatment.

2. Methodology

The study employed a single-center, retrospective observational design involving 135 AIS patients treated at Xinxiang Central Hospital between January 2019 and December 2023.

Data Collection & Preprocessing

• Cohort: 135 patients (70 developed SAP, 65 did not).
• Variables: Collected demographic data, medical history, clinical scores (NIHSS, mRS), and laboratory indicators at three timepoints: Admission, 24 hours, and 48 hours.
• Key Biomarkers: Neutrophil-to-Lymphocyte Ratio (NLR), Platelet-to-Lymphocyte Ratio (PLR), Systemic Immune-Inflammation Index (SII), and Systemic Inflammatory Response Index (SIRI).
• Exclusion: Patients with pre-existing infections, autoimmune disorders, malignancies, or missing blood data were excluded.

Feature Selection

• Univariate Analysis: Identified variables with $P < 0.05$.
• LASSO Regression: Used to reduce dimensionality and select the most predictive features while preventing overfitting.
• Result: 11 key variables were selected from an initial pool of 63, including: Age, NIHSS_7d, WBC_24h, SII_24h, SIRI_24h, NLR_48h, Fasting Blood Glucose (FBG), D-dimer, Gastric Tube placement (G-tube), NIHSS_preMT, and NIHSS_24h_postIVT.

Model Development & Evaluation

• Algorithms: Ten ML models were trained and compared: Logistic Regression (LR), SVM, Random Forest (RF), Gradient Boosting (GBM), Neural Network (NN), XGBoost, KNN, AdaBoost, LightGBM, and CatBoost.
• Splitting: Data was stratified into a Training Set (70%, n=95) and a Test Set (30%, n=40).
• Metrics: Performance was evaluated using Area Under the Curve (AUC-ROC), Accuracy, Precision, Recall, F1-Score, and Decision Curve Analysis (DCA).
• Interpretability: The SHAP (Shapley Additive Explanations) framework was applied to the best-performing model to quantify feature contributions and visualize interactions, addressing the "black-box" nature of ML.

3. Key Results

Biomarker Analysis

• Significant Differences: The SAP group showed significantly higher levels of 24h and 48h NLR, 24h SII, and 24h/48h SIRI compared to the non-SAP group ($P < 0.05$).
• Non-Significant: No significant difference was found in PLR between the two groups, suggesting PLR may be less predictive in this specific bridging therapy population.
• Clinical Correlates: SAP patients had higher NIHSS scores (indicating severe neurological deficit), higher fasting blood glucose, elevated D-dimer, and a higher rate of gastric tube placement.

Model Performance

• Best Model: The CatBoost model demonstrated superior performance compared to all other algorithms.
  • Training Set AUC: 0.952
  • Test Set AUC: 0.932
• Comparison: CatBoost outperformed other strong contenders like SVM (Test AUC 0.922) and Random Forest (Test AUC 0.940, though RF showed signs of overfitting with a perfect training score of 1.0).
• Decision Curve Analysis (DCA): Confirmed that the CatBoost model provided a higher net clinical benefit across a wide range of threshold probabilities compared to other models.

Interpretability (SHAP Analysis)

• Top Predictors: The SHAP summary plot identified the three most influential factors for SAP prediction:
  1. NIHSS_7d (7-day National Institutes of Health Stroke Scale score)
  2. SIRI_24h (24-hour Systemic Inflammatory Response Index)
  3. WBC_24h (24-hour White Blood Cell count)
• Feature Interaction: SHAP dependence plots revealed that the combination of advanced age and elevated NIHSS_24h_postIVT significantly increased the predicted risk of SAP.
• Individual Prediction: SHAP waterfall and force plots successfully visualized how specific patient features (e.g., high pre-MT NIHSS) pushed the probability toward the "SAP" class, while others acted as protective factors.

4. Key Contributions

1. Novel Population Focus: This is the first study to construct an interpretable ML model specifically for AIS patients undergoing bridging therapy (IVT + MT), a high-risk subgroup often excluded from general SAP studies.
2. Dynamic Biomarker Integration: Unlike static models, this study utilized longitudinal data (Admission, 24h, 48h) to capture the dynamic evolution of inflammatory responses (NLR, SII, SIRI), proving that 24h and 48h markers are critical for early risk stratification.
3. Interpretable AI: By integrating SHAP with the CatBoost algorithm, the study overcomes the "black-box" limitation of ML, providing clinicians with transparent, feature-level explanations for risk predictions.
4. Feature Selection Rigor: The use of LASSO regression ensured that the final model relied on a parsimonious set of 11 clinically relevant variables, enhancing its potential for real-world deployment.

5. Significance and Clinical Implications

• Early Intervention: The model enables clinicians to identify high-risk SAP patients early (within 24-48 hours) based on objective biomarkers and clinical scores, rather than waiting for clinical symptoms to manifest.
• Personalized Medicine: Early identification allows for targeted prophylactic strategies (e.g., strict aspiration precautions, early mobilization, or judicious antibiotic use) for high-risk individuals, potentially reducing the incidence of SAP and improving functional outcomes (mRS).
• Resource Optimization: By distinguishing high-risk from low-risk patients, hospitals can better allocate intensive care resources and avoid the overuse of prophylactic antibiotics in low-risk patients, mitigating antibiotic resistance.
• Future Direction: While the model shows high accuracy, the authors acknowledge the need for external validation in multicenter, larger-scale studies to confirm generalizability before widespread clinical adoption.

Conclusion: The study successfully demonstrates that an interpretable CatBoost model, driven by dynamic inflammatory biomarkers (specifically NLR, SII, and SIRI at 24/48h) and clinical severity scores, offers a highly accurate tool for stratifying SAP risk in AIS patients undergoing bridging therapy.