Stroke Risk Assessment Tool

Advanced Cerebrovascular Disease Risk Prediction System Based on Machine Learning Models

About This Tool

This comprehensive stroke risk assessment tool integrates multiple validated clinical scoring systems with advanced machine learning algorithms to provide accurate cardiovascular risk evaluation. Our system combines vessel stenosis analysis (X_score), Framingham risk factors, Gensini coronary artery scoring, and CT-FFR measurements to deliver precise risk stratification for clinical decision-making.

Designed for healthcare professionals, this tool offers four distinct machine learning assessment models tailored for different clinical scenarios. Each model has been trained on validated datasets and optimized for clinical accuracy and practical applicability, supporting evidence-based patient care and risk management strategies.

AI-assisted analysis synthesizes your inputs (e.g., X_score, Framingham, Gensini, CT-FFR) with the model’s predicted risk to generate a one-click, structured clinical interpretation and follow-up recommendations; the API key is stored locally.

ML Algorithm

Advanced machine learning with optimized hyperparameters for optimal prediction accuracy

Clinically Validated

Based on established scoring systems and validated training datasets

Multi-Model

Four specialized models for different feature combinations

Real-time Prediction

Instant risk calculation with detailed interpretation

Key Features

Multi-Model Assessment

Choose from four specialized models: X_score only, X_score + Framingham, Framingham + Gensini, or comprehensive assessment including CT-FFR. Each model uses optimized parameters and distance metrics.

Machine Learning Powered

Advanced machine learning algorithm with standardized feature scaling, optimized distance calculation, and validated hyperparameters for optimal performance.

Real-time Risk Prediction

Instant risk probability calculation using trained models with comprehensive training data for accurate risk assessment and clinical decision support.

Machine Learning Algorithm & Implementation

1. Machine Learning Model Selection & Validation

After comprehensive evaluation of multiple machine learning models, an optimized algorithm was selected that demonstrated superior performance across all evaluation metrics.

Cross-Validation Protocol

Validation Method: 5-fold stratified cross-validation with 3 repetitions
Performance Metrics: AUC, Sensitivity, Specificity, Accuracy, Brier Score
Feature Scaling: Z-score standardization (mean=0, std=1)
Random Seed: Fixed at seed for reproducibility

2. Training Data Overview

Four datasets were used to train the machine learning models, each optimized for specific feature combinations with comprehensive clinical data.

Model	Features	Training Samples	Distance Metric	Positive Rate
Model 1 (X)	X_score	591	Minkowski (p=1.0)	75.8%
Model 2 (X+F)	X_score + Framingham	591	Minkowski (p=2.5)	75.8%
Model 3 (F+G)	Framingham + Gensini	591	Minkowski (p=2.5)	75.8%
Model 4 (F+G+CT)	Framingham + Gensini + CT-FFR	591	Minkowski (p=2.5)	75.8%

3. Detailed Model Parameters

Each model has been optimized with specific hyperparameters, feature weights, class weights, and calibration parameters for optimal performance.

Model 1: X_score Only

Parameter	Value	Description
k	23	Number of nearest neighbors
Distance (p)	1.0	Minkowski distance parameter
Kernel	Rectangular	Adaptive bandwidth h = d_k
Features	X_score	Input feature dimensions
Scaling Center	33.0968	X_score standardization mean
Scaling Scale	46.7473	X_score standardization std
Class Weight (Pos)	0.6596	Positive class weight
Class Weight (Neg)	2.0664	Negative class weight
Feature Weight	1.0	X_score feature importance
Calibration α	-1.0445	Platt calibration parameter
Calibration β	4.6656	Platt calibration parameter

Model 2: X_score + Framingham

Parameter	Value	Description
k	23	Number of nearest neighbors
Distance (p)	2.5	Minkowski distance parameter
Kernel	Epanechnikov	Adaptive bandwidth h = d_k
Features	X_score, Framingham_Score	Input feature dimensions
X_score Center/Scale	33.0968 / 46.7473	X_score standardization
Framingham Center/Scale	18.9543 / 6.335	Framingham standardization
Class Weight (Pos/Neg)	0.6596 / 2.0664	Class balancing weights
X_score Weight	0.9002	X_score feature importance
Framingham Weight	0.0998	Framingham feature importance
Calibration α/β	-1.1161 / 4.7379	Platt calibration parameters

Model 3: Framingham + Gensini

Parameter	Value	Description
k	23	Number of nearest neighbors
Distance (p)	2.5	Minkowski distance parameter
Kernel	Rectangular	Adaptive bandwidth h = d_k
Features	Framingham_Score, Gensini_score	Input feature dimensions
Framingham Center/Scale	18.9543 / 6.335	Framingham standardization
Gensini Center/Scale	21.1328 / 23.0726	Gensini standardization
Class Weight (Pos/Neg)	0.6596 / 2.0664	Class balancing weights
Framingham Weight	0.5974	Framingham feature importance
Gensini Weight	0.4026	Gensini feature importance
Calibration α/β	-0.5576 / 3.357	Platt calibration parameters

Model 4: Framingham + Gensini + CT-FFR

Parameter	Value	Description
k	23	Number of nearest neighbors
Distance (p)	2.5	Minkowski distance parameter
Kernel	Rectangular	Adaptive bandwidth h = d_k
Features	Framingham_Score, Gensini_score, CT-FFR	Input feature dimensions
Framingham Center/Scale	18.9543 / 6.335	Framingham standardization
Gensini Center/Scale	21.1328 / 23.0726	Gensini standardization
CT-FFR Center/Scale	0.7453 / 0.1284	CT-FFR standardization
Class Weight (Pos/Neg)	0.6596 / 2.0664	Class balancing weights
Framingham Weight	0.5837	Framingham feature importance
Gensini Weight	0.3113	Gensini feature importance
CT-FFR Weight	0.105	CT-FFR feature importance
Calibration α/β	-0.5453 / 3.325	Platt calibration parameters

4. Machine Learning Algorithm Implementation

// Machine Learning Risk Prediction Algorithm

function calculateRisk(inputFeatures, modelKey) {

 // 1. Load model parameters and training data

 const model = models\[modelKey];

 const {trainX, trainY} = trainingData\[modelKey];

 // 2. Standardize input features

 const normalizedFeatures = standardizeFeatures(inputFeatures, model);

 // 3. Calculate weighted distances to training samples

 const distances = calculateWeightedDistances(normalizedFeatures, trainX, model);

 // 4. Select k nearest neighbors

 const nearestNeighbors = selectNearestNeighbors(distances, trainY, model.k);

 // 5. Apply kernel weighting (bandwidth = distance of k-th neighbor)

 const h = maxDistanceAmongK(nearestNeighbors);      // h = d_k

 const prob = kernelWeightedMean(

  nearestNeighbors, h,

  model.kernel /* 'epanechnikov' | 'rectangular' */,

  model.class_weights

 );

 // 6. Apply Platt calibration (GLM-fitted)

 const p = 1 / (1 + Math.exp(-(model.calibration.alpha + model.calibration.beta * Math.min(1-1e-6, Math.max(1e-6, prob)))) );

 return p;

}

Training Data Integration

Data Integration:

Real-time loading of complete training datasets from CSV files
Automatic feature standardization using training set parameters
Optimized kernel weighting and nearest neighbor selection for accurate prediction
Platt scaling for probability calibration

Risk Stratification

Risk Level	Probability Range	Clinical Action
Low Risk	< 30%	Continue regular monitoring, 6-12 month follow-up
Moderate Risk	30% - 60%	Enhanced monitoring, lifestyle modifications, 3-6 month follow-up
High Risk	> 60%	Urgent evaluation, aggressive intervention, close monitoring

5. X_score Scoring

X_score is calculated from stenosis severity in different vascular segments. Values are non-negative with no upper bound (≥0) and come directly from imaging.

X_score = Σ(Vessel Stenosis Value × Vessel Weight Coefficient)

Vessel Segment	Unilateral Multiplier	Bilateral Multiplier	Calculation Formula
Subclavian Artery (SA)	-	1	(L_SA + R_SA) × 1
Common Carotid Artery (CCA)	8	12	Single: value × 8, Both: (L_CCA + R_CCA) × 12
Internal Carotid Artery (ICA)	9	12	Single: value × 9, Both: (L_ICA + R_ICA) × 12
Basilar Artery (BAA)	25	-	BAA × 25
Vertebral Artery (VA)	19	22	Single: value × 19, Both: (L_VA + R_VA) × 22
Anterior Cerebral Artery (ACA)	8	9	Single: value × 8, Both: (L_ACA + R_ACA) × 9
Middle Cerebral Artery (MCA)	11	13	Single: value × 11, Both: (L_MCA + R_MCA) × 13
Posterior Cerebral Artery (PCA)	33	37	Single: value × 33, Both: (L_PCA + R_PCA) × 37

6. Framingham Score Calculation

Risk score based on Age, Systolic Blood Pressure (SBP) (both untreated and treated, if provided), and cardiovascular risk factors, with gender-specific tables.

Framingham Score = Age Points + SBP (Untreated) Points + SBP (Treated) Points + Risk Factor Points

Note: If only one SBP is provided, only that one is scored; values beyond highest brackets are capped at the top points.

Male Scoring Table

Age (years)	Points	SBP Untreated (mmHg)	Points	SBP Treated (mmHg)	Points
≤56	0	97–105	0	97–105	0
57–59	1	106–115	1	106–112	1
60–62	2	116–125	2	113–117	2
63–65	3	126–135	3	118–123	3
66–68	4	136–145	4	124–129	4
69–72	5	146–155	5	130–135	5
73–75	6	156–165	6	136–142	6
76–78	7	166–175	7	143–150	7
79–81	8	176–185	8	151–161	8
82–84	9	186–195	9	162–176	9
≥85	10	196–205	10	177–205	10

Additional Risk Factors (Male)

Risk Factor	Points
Diabetes	2
Smoking	3
Cardiovascular Disease	4
Atrial Fibrillation	4
Left Ventricular Hypertrophy	5

Female Scoring Table

Age (years)	Points
≤56	0
57–59	1
60–62	2
63–64	3
65–67	4
68–70	5
71–73	6
74–76	7
77–78	8
79–81	9
82–84	10
≥85	10

SBP Untreated (mmHg)	Points
95–106	1
107–118	2
119–130	3
131–143	4
144–155	5
156–167	6
168–180	7
181–192	8
193–204	9
205–216	10

SBP Treated (mmHg)	Points
95–106	1
103–113	2
114–119	3
120–125	4
126–131	5
132–139	6
140–148	7
149–160	8
161–204	9
205–216	10

Additional Risk Factors (Female)

Risk Factor	Points
Diabetes	3
Smoking	3
Cardiovascular Disease	2
Atrial Fibrillation	4
Left Ventricular Hypertrophy	5

7. Gensini Score System

Coronary artery disease severity assessment based on stenosis degree and vessel importance coefficients.

Gensini Score = Σ(Stenosis Points × Vessel Coefficient)

Stenosis Degree	Points	Vessel Segment	Coefficient
<25%	1	Left Main	5.0
25–49%	2	LAD Proximal	2.5
50–74%	4	LAD Mid	1.5
75–89%	8	LAD Distal	1.0
90–98%	16	First Diagonal	1.0
≥99%	32	Second Diagonal	0.5
-	-	LCX Proximal	2.5
-	-	LCX Distal/PDA	1.0
-	-	RCA (all segments)	1.0
-	-	Posterolateral	0.5