Research &
Clinical Methodology
Automated stroke pre-diagnostic screening driven by real-time computer vision and machine learning. GROUNDED in clinical literature, validated prehospital scales, and a synthetic clinical dataset.
Clinical Screening & Early Warning
Bridging the gap in prehospital emergency settings with objective, instant computer-vision assessments.
Every Minute Matters: Why Early Screening Saves Lives
When a stroke occurs, every single second counts. Restoring blood flow and delivering medical treatment within the critical "Golden Hour" (the first few hours) can significantly prevent permanent brain damage and completely restore patient health.
However, due to a lack of proper medical knowledge, people often fail to notice subtle early signs of a stroke (such as facial droop) before it is too late. NeuroScan was designed to bridge this gap, providing a fast, accessible, browser-based screening tool that flags early facial motor deficit markers immediately.
সহজ বাংলায় বার্তা
যদি কোনো ব্যক্তি স্ট্রোকে আক্রান্ত হন, তবে অল্প সময়ের মধ্যে (Golden Hour) দ্রুত মেডিকেল চিকিৎসা দেওয়া গেলে তাকে পুরোপুরি সুস্থ করা সম্ভব।
কিন্তু সাধারণ মানুষের সঠিক চিকিৎসা জ্ঞান না থাকার কারণে অনেকেই বুঝতে পারেন না যে স্ট্রোকে আক্রান্ত হওয়ার কোনো সম্ভাবনা তৈরি হয়েছে কি না। আর এই সমস্যার সমাধান করতেই আমাদের এই সিস্টেমে সহজ এবং দ্রুত স্ক্রিনিংয়ের ব্যবস্থা করা হয়েছে।
Timeline of Scientific Evidence
Chronological research grounding the validity and clinical necessity of computer-vision stroke screening.
Cincinnati Prehospital Stroke Scale (CPSS)
Clinically establishes the standard prehospital stroke checklist. Proves facial droop/droop asymmetry carries twice the diagnostic weight of eye signs.
View DOI →Asymmetry Metric Formulation
Formulates left-right landmark euclidean distance ratios for mouth, eyebrow, and eyelid features. Directly adapted into NeuroScan's core geometry library.
View Preprint →Human vs. ML Facial Weakness
Reveals human paramedics fail to detect facial weakness in 17% of acute stroke cases, highlighting the need for objective computer-aided ML tools.
View Publication →Facial Point Graphs for Stroke
Demonstrates that coordinate mappings on landmark deltas bypass variables like skin tone and poor lighting, preventing demographics-based bias in AI detection.
ResearchGate →Deep Learning in FER Diagnosis
Surveys stroke facial recognition models. Notes that Toronto Neuroface (TNF) contains only 14 patients, proving why synthetic data training is mathematically necessary.
View Publication →Synthetic Training & Real Validation
How we bypass regulatory clinical image shortages using clinical simulation and real-world testing sets.
Clinically-Inspired Synthetic Data (N=3,000)
Due to HIPAA privacy constraints and a severe scarcity of public patient facial datasets (for instance, the Toronto Neuroface dataset contains only 14 patients), training deep/statistical classifiers directly on raw patient photos is clinically and statistically limited. We generated 3,000 synthetic patient vectors using range distributions mapped from Cincinnati Prehospital Stroke Scale (CPSS) criteria and Sunnybrook Facial Grading metrics to ensure robust generalizability.
100% HIPAA & Privacy CompliantAnnotated Stroke Dataset (Kaggle)
Consists of 2,490 stroke and 5,000 normal real-world images curated by Abdussalam Elhanashy. Used strictly as the test set to evaluate real-world classifier generalization, validating that the models trained on synthetic vectors translate accurately to real patient photo landmarks.
Kaggle Dataset →Facial Droop & Palsy (Kaggle)
Consists of 1,024 images of confirmed facial palsy (curated by Mehta). Used to calibrate the baseline manual MediaPipe excursion ratios (e.g., eye-close threshold 0.35, eye-open 1.25) to isolate neurological palsy features from natural asymmetries.
Kaggle Dataset →How stroke_model.pkl is Built
Simulate
Simulate 3,000 patient parameters based on CPSS & Sunnybrook grading.
Extract
Extract 6 key landmark asymmetry feature vectors.
Scale
Apply StandardScaler and export as stroke_scaler.pkl.
Train
Fit Random Forest and XGBoost classifiers on synthetic coordinates.
Test
Test models against Elhanashy's Kaggle images; export stroke_model.pkl.
Model Performance & Feature Distribution
Actual data plots extracted from the training notebook showcasing classifier metrics and synthetic feature separation.
Feature Distributions (Normal vs. Stroke)
Probability density distributions of key facial asymmetry parameters. Yellow dashed lines denote clinical thresholds calibrated using synthetic constraints and facial palsy datasets.
Classifier Evaluation (Confusion Matrix & ROC)
Validation performance of Random Forest and XGBoost models on the Kaggle real-world validation set. High sensitivity (Recall) is prioritized to prevent missed stroke presentations.
Facial Feature Importance Ranking
Feature importance weightings showing that **mouth corner drop** (FAST "F" indicator) and nasolabial fold asymmetry hold the highest predictive weight in classifying acute hemiparesis.
The Guided Symmetry Protocol
Every webcam test evaluates specific motor responses mapping to Cranial Nerve VII (Facial Nerve) clinical checklists.
Resting calibration
Maps standard neutral positions of mouth and eyes to construct the reference coordinate matrix.
Active smile
Triggers mouth-droop check, evaluating left/right zygomaticus major movement ratios.
Eye closure
Assesses eyelid closure ratios. Checks for lagophthalmos or asymmetrical eyelid folds.
Eye opening
Evaluates frontalis muscle elevation during wide opening to detect subtle upper eyebrow drop.
Verbal smile
Integrates voluntary smile execution with voice activation to compute final asymmetry probability.
Why a 16-Second Active Stress Test?
Traditional stroke screens rely on simple visual inspection of a resting face. However, mild ischemic events affecting the motor pathway of Cranial Nerve VII (Facial Nerve) may produce no visible asymmetry when facial muscles are relaxed. Resting muscle tone can remain deceptively symmetric.
By prompting the patient through a rapid, 16-second sequence of active facial contractions (voluntary resting calibration, forced smile, tight eyelid closure, and deep brow elevation), the protocol functions as a neuromotor stress test. It forces active corticobulbar nerve transmission, instantly exposing subtle excursion latency and movement discrepancies (Δ) that are completely invisible at rest.
Physiological Mechanics (Cranial Nerve VII)
- Pathway Challenge: Voluntary movements test both the upper and lower motor pathways, stressing zygomaticus major and frontalis contractions.
- Excursion Delta (Δ): Instead of static absolute measurements, the system calculates the delta between left/right coordinates during max dynamic displacement.
- Clinical Reference: Grounded in the Herpich et al. (2022) clinical findings showing automated computer-vision analysis identifies subtle facial paretic deficits that human responders miss in 17% of emergency cases.
Landmark Mapping & API Stack
Real-time landmark coordinates extracted from Google MediaPipe and analyzed securely on our FastAPI server.
Landmark Indices Mapped
Hover over the landmark coordinates below to isolate and view their spatial location on the facial mesh.
| Feature Mapped | Index | Facial Landmark Details |
|---|---|---|
NOSE_TIP | 1 | Central reference landmark coordinate |
FOREHEAD | 10 | Upper face midline boundary coordinate |
CHIN | 152 | Lower jaw midline boundary coordinate |
MOUTH_LEFT | 61 | Left mouth corner (zygomaticus major) |
MOUTH_RIGHT | 291 | Right mouth corner (zygomaticus major) |
LEFT_EYE_TOP | 159 | Left upper eyelid boundary index |
LEFT_EYE_BOTTOM | 145 | Left lower eyelid boundary index |
RIGHT_EYE_TOP | 386 | Right upper eyelid boundary index |
RIGHT_EYE_BOTTOM | 374 | Right lower eyelid boundary index |
LEFT_EYEBROW | 70 | Left frontalis muscle elevation reference |
RIGHT_EYEBROW | 300 | Right frontalis muscle elevation reference |
LEFT_NASOLABIAL | 92 | Left nasolabial fold coordinate (cheek) |
RIGHT_NASOLABIAL | 322 | Right nasolabial fold coordinate (cheek) |
LEFT_FACE_EDGE | 234 | Left lateral head boundary reference |
RIGHT_FACE_EDGE | 454 | Right lateral head boundary reference |
Computer Vision Mapping
MediaPipe Face Mesh extracting spatial coordinates on a patient face, highlighting asymmetry vectors on mouth and eyelid contours.