StrokeSense: AI-Powered Patient Stroke Risk Assessment

Abstract

Stroke is a leading cause of death and long-term disability. Early risk stratification enables timely lifestyle modification and clinical intervention. This project builds a supervised learning pipeline to predict individual stroke risk from routinely available attributes (e.g., age, gender, hypertension, heart disease, work/residence type, average glucose level, BMI, smoking status). We implement and compare five algorithms—Logistic Regression, Support Vector Machine, Random Forest, Gaussian Naïve Bayes, and Gradient Boosting Classifier—across a standardized ML workflow: data cleaning, imputation, outlier handling, encoding, scaling, feature selection, training, and evaluation using stratified cross-validation. The best model is deployed through a Streamlit web UI for interactive inference and explanation (feature importance/SHAP-like insights). Predictions, feedback, and anonymized usage logs are stored in MySQL for auditability and model monitoring. Emphasis is placed on clinically relevant metrics (recall/sensitivity for the positive “stroke” class), calibration, and fairness checks across subgroups. The solution is modular, reproducible, and ready for integration into screening workflows as a decision-support tool—not a diagnostic substitute.

Introduction

Stroke risk arises from modifiable and non-modifiable factors. Traditional risk scores can be rigid, while ML can capture nonlinearities and feature interactions. However, medical ML must prioritize sensitivity, interpretability, and data governance. This project addresses these by:

• Establishing a robust pipeline for tabular clinical-like data.
• Comparing a diverse set of models from linear to ensemble methods.
• Prioritizing recall (to reduce missed high-risk cases) while maintaining acceptable precision.
• Calibrating predicted probabilities to improve clinical trust.
• Providing an intuitive Streamlit app for clinicians/researchers and saving predictions to MySQL for continuous monitoring and iterative improvement.

Problem Statement (What we solve)

Given a patient’s demographic and clinical attributes, predict whether they are at risk of stroke (binary classification). We need a deployable, data-driven, and interpretable system that balances performance (especially recall) with transparency and maintainability, suitable for screening support in non-specialist settings.

Existing System

• Manual/score-based assessments (e.g., rule-of-thumb risk charts).
• Fragmented spreadsheets or basic statistical models without calibration.
• Little or no feedback loop for improving models with real-world usage.
• Limited UI for quick triage and no central database for tracking predictions.

Disadvantages of Existing System

• Limited accuracy for heterogeneous populations.
• Poor adaptability to new data distributions.
• Lack of explainability and calibration in many ad-hoc tools.
• No centralized storage, hindering quality control and improvement.

Proposed System

• End-to-end ML pipeline with five candidate algorithms, model selection by stratified CV.
• Bias-aware evaluation (class weights, resampling if needed), probability calibration (Platt/Isotonic).
• Streamlit UI for data entry/batch CSV scoring, feature importance view, and result export.
• MySQL backend to store inputs, predictions, model/version metadata, and feedback.
• Monitoring hooks (drift indicators, confusion trends) for iterative retraining.

Advantages

• Higher recall for positive stroke class via tuned thresholds and class weighting.
• Transparency: coefficients/feature importance, partial dependence, or SHAP-like summaries.
• Reproducibility: consistent preprocessing and versioned models.
• Operationalization: easy-to-use app and persistent database.
• Scalability: modular design supports future models and new features.

Modules

1. Data Ingestion & Validation: CSV upload/API intake; schema/NA/type checks.
2. Preprocessing:
   • Imputation (median for numeric, most frequent for categorical); optional BMI/glucose winsorization.
   • Encoding (One-Hot for nominal, ordinal if applicable), scaling (Standard/MinMax as needed).
3. EDA & Feature Engineering: correlations, mutual information, interaction terms (e.g., age×hypertension), feature selection.
4. Model Training & Selection: train LR, SVM, RF, GNB, GBC with stratified CV; hyperparameter tuning (Grid/Random).
5. Evaluation & Calibration: metrics (Recall, Precision, F1, AUC-ROC/PR), calibration curves, threshold tuning for operating point.
6. Model Persistence: joblib/pickle for pipeline + label encoder + threshold + metadata (version, date, metrics).
7. Streamlit UI: single-record form, batch scoring, explanations, and result download.
8. MySQL Persistence Layer: store predictions, inputs (de-identified), feedback, model version, and audit trails.
9. Monitoring & Retraining Hooks: drift checks, periodic reports, dataset snapshots for retraining.

Algorithms / Models

• Logistic Regression (LR): Baseline interpretable linear model; supports class_weight and calibration; useful for odds interpretation.
• Support Vector Machine (SVM): Nonlinear boundaries with RBF; strong performance on tabular data; requires scaling; probability via calibration.
• Random Forest (RF): Bagged trees; robust to noise/outliers; handles nonlinear interactions; built-in feature importance; good default.
• Gaussian Naïve Bayes (GNB): Simple, fast, works well with conditional independence; good baseline for probabilistic outputs.
• Gradient Boosting Classifier (GBC, e.g., XGBoost/LightGBM/Sklearn GBC): Powerful ensembles capturing complex patterns; often top performer; supports class imbalance and calibrated probabilities.

Model Selection Strategy:

• Stratified 5–10-fold CV on training set; compare AUC-PR (focus on minority class), Recall@selected-Precision, F1, and Brier score (calibration).
• Choose winner by primary metric = Recall (or AUC-PR) with calibration check, then set operating threshold to desired sensitivity/precision tradeoff.
• Validate on hold-out/bootstrapped test set; run subgroup analysis (sex/age bins, comorbidity).

Software & Hardware Requirements

Software

• Language: Python 3.10+
• Libraries: scikit-learn, pandas, numpy, imbalanced-learn (optional), joblib, matplotlib/plotly, shap (optional), sqlAlchemy/mysql-connector-python
• App: Streamlit for UI
• Database: MySQL

Hardware (typical dev/deploy)

• Dev Machine: 8–16 GB RAM, 4+ CPU cores, 10+ GB storage
• Server (small pilot): 2–4 vCPU, 8 GB RAM, SSD; no GPU required
• Scaling: Horizontal scale or containerization (Docker) for higher traffic

Conclusion

This project delivers a practical, explainable stroke risk screening tool leveraging five complementary ML algorithms, rigorous evaluation, and a user-friendly Streamlit interface backed by MySQL. By emphasizing recall, calibration, and transparency, it supports early identification of at-risk individuals and provides a foundation for continuous improvement through monitored deployments.

Future Enhancement

• Explainability: Full SHAP integration with per-prediction breakdown and global summaries.
• Data Expansion: Incorporate labs, vitals, wearables, or longitudinal EHR features.
• AutoML & Ensembles: Stacking/blending winners to boost AUC-PR and stability.
• MLOps: CI/CD, model registry, drift alerts, scheduled retraining, and canary releases.
• Fairness Audits: Regular disparate impact and equalized odds monitoring.
• Clinician Feedback Loop: Label corrections, outcome capture for real-world calibration.
• Security & Privacy: Role-based access, encryption at rest/in transit, de-identification pipelines.