-
-
NeuroShield - Harnessing AI to Safeguard Minds and Data
Inspiration
NeuroShield was inspired by the ongoing need for early, accessible detection tools for neurological disorders like Parkinson’s Disease and stroke. These conditions often manifest subtle motor anomalies such as tremors or changes in gait long before diagnosis. Faced with stories from both the medical and patient communities, we aimed to create a companion technology that could non-invasively detect such signs from ordinary video data using computer vision and deep learning
What it does
NeuroShield provides a real-time gait and movement analysis platform. Patients or clinicians can upload short video segments, which are then processed by deep learning models (YOLOv8 for object detection and custom CNN/RNN for sequence analysis). The system detects anomalies such as tremors, imbalances, and atypical gait patterns flagging potential neurological concerns and providing visualized feedback for further assessment
How I built it
Frontend: The app is built using Streamlit for interactive patient and clinician use, featuring a custom green-themed UI for visual comfort and clarity.
Backend: We adapted the YOLOv8 model, retrained on custom annotated video frames, enabling robust detection of hands, feet, and overall posture. We also developed CNN/RNN modules for temporal feature extraction and anomaly classification.
Workflow: Raw videos are uploaded, frames are extracted and annotated (using CVAT/LabelImg), and labeled data is used to train YOLO and our sequence models. Inference pipelines process videos end-to-end, reporting results through the frontend.
Integration: Python (3.10), PyTorch, ultralytics, OpenCV, and pandas are all tightly integrated for reliable training and deployment.
Challenges I ran into
Python environment errors: Frequent DLL/installation issues (especially with PyTorch on Windows), missing package errors, and broken dependencies required careful venv and PATH management.
Resource constraints: Many objects encountered memory-related errors when training large models on CPU—necessitating batch size, image size, and system configuration optimizations.
Annotation workflow: Balancing annotation completeness vs. developer time was tricky; video-level labeling directly with tools like CVAT expedited data prep.
Model loading: Safe weight loading in PyTorch 2.6+ required extra context management to avoid unpickling errors.
Deployment quirks: Streamlit integration, CSS theming, and real-time video processing each presented unique debugging challenges.
Accomplishments that I am proud of
1) Created an end-to-end pipeline that reliably accepts user video and produces validated medical anomaly feedback.
2)Overcame substantial dependency, resource, and deployment obstacles to achieve working desktop inference in Python.
3)Developed a reusable, modular codebase and workflow that can be adapted to other medical anomaly detection domains.
4)Provided meaningful feedback and visualization to both clinician and patient users, enhancing digital accessibility in health screenings.
What I have learned
1)The intersection of computer vision, time-series neural networks, and medical data can be powerful—but demands careful engineering and validation.
2)Managing cross-platform Python dependencies, virtual environments, and system configuration is crucial for reproducibility.
3)Data annotation for medical applications benefits greatly from specialized tools and an iterative workflow.
4)Deploying AI in healthcare means balancing technical feasibility, user experience, and ethical considerations at every step.
What's next for NeuroShield - An AI Gait Companion
Expand anomaly detection: Integrate more advanced gait analytics (e.g., spatiotemporal joint tracking, deep ensemble models).
Clinical validation: Collaborate with neurologists to benchmark and tune detection models against gold-standard clinical datasets.
Explainability: Improve interpretability via annotated visual feedback and explainable AI methods.
Scalability: Move toward cloud-based processing and multi-device compatibility for broader outreach.
Mathematical improvements: Further optimize detection accuracy with improved models.
Log in or sign up for Devpost to join the conversation.