Technical Summary: PredictStroke

Problem Statement

Current artificial intelligence solutions for stroke management typically operate in silos, focusing exclusively on either long-term risk prediction using structured tabular data or real-time acute symptom tracking using physiological signals. There is a lack of unified frameworks that bridge preventative analytics with acute triage support. Furthermore, existing multimodal systems often suffer from high computational intensity, lack of real-world deployment considerations (such as mobile integration), and a separation between predictive modeling and acute detection that hinders comprehensive clinical decision-making.

Methodology

The authors present PredictStroke, a modular, two-stage system designed to integrate long-term risk assessment with real-time symptom detection. The system is composed of three distinct components:

1. Stroke Risk Prediction Engine (Preventative Stage)

• Data Source: The Kaggle Stroke Prediction Dataset, containing structured clinical and demographic features (e.g., age, hypertension, heart disease, glucose levels, BMI, gender, smoking status).
• Preprocessing: Missing BMI values were imputed; categorical variables were encoded; numerical features were normalized to prevent distance-based model bias.
• Model Architecture: An ensemble approach utilizing Soft Voting (probability averaging) across three classical machine learning models:
  • K-Nearest Neighbors (KNN): Classifies based on similarity to nearby examples.
  • Random Forest: Captures non-linear feature interactions via multiple decision trees.
  • Support Vector Machine (SVM): Identifies separating boundaries between stroke and non-stroke cases.
• Output: A probability score indicating long-term stroke risk. A threshold of 0.5 was applied to flag elevated risk.

2. Acute Symptom Detection (Triage Stage)

This stage operates on independent feature-engineered pipelines for facial and speech analysis, rather than end-to-end deep learning, to facilitate modularity.

• Facial Drooping Detection:
  • Input: Images processed to extract facial landmarks (eyes, mouth, cheeks, jawline).
  • Feature Engineering: Computed asymmetry metrics, including left-right mouth corner height differences, cheek/jaw landmark position differences, and normalized symmetry ratios.
  • Classifier: An SVM trained to distinguish normal posture from asymmetry using these engineered features.
• Slurred Speech Detection:
  • Input: Audio samples cleaned and normalized to reduce volume and noise variance.
  • Feature Engineering: Extraction of Mel-frequency cepstral coefficients (MFCCs) for spectral shape, pitch variance, and speech rate/pause intervals.
  • Classifier: An SVM trained to classify abnormal speech patterns.
• Symptom Coordination: The system treats facial and speech detectors as independent modules. A final "possible stroke" flag is triggered only if facial drooping, slurred speech, and user-reported arm weakness are all positive.

Key Results

Performance was evaluated using precision, recall, and F1-score to account for class imbalance in healthcare datasets.

• Risk Prediction Engine: Demonstrated robust performance with an F1-score of 0.85, a precision of 0.90, and a recall of 0.80. The ensemble approach proved effective in stabilizing predictions.
• Facial Drooping Module: Achieved an F1-score of 0.65 (Precision: 0.62, Recall: 0.68).
• Slurred Speech Module: Achieved an F1-score of 0.62 (Precision: 0.65, Recall: 0.60).

The authors note that the detection modules achieved moderate performance, attributed to the reliance on handcrafted features and limited labeled datasets compared to the structured clinical data used for risk prediction.

Significance and Claims

The paper positions PredictStroke not as a medical diagnostic tool, but as a conceptually valuable and technically feasible framework for remote triage and early intervention.

• Modularity as a Solution: The primary contribution is the architectural design that allows data modeling pipelines to scale independently. This approach mitigates common bottlenecks in consumer-facing digital health, such as severe data availability issues and complex system integration.
• Bridging the Gap: By combining structured data-based risk modeling with multimodal symptom detection, the system addresses the historical separation between preventative healthcare and acute triage support.
• Design Intent vs. Implementation: While the system was designed with scalability and potential integration into mobile/web platforms in mind, the authors explicitly note that full integration of the detection modules into the web application was not achieved due to development constraints and team composition changes. Despite this limitation, the independent validation of each component confirms the feasibility of the proposed modular architecture.

The authors acknowledge limitations, including the lack of full integration between detection modules and the web interface due to project constraints, the absence of real-time clinical deployment testing, and the reliance on simplified decision rules. However, they assert that the independent validation of each component confirms the feasibility of the proposed modular architecture.