A System For Snake Bite Imaging And Classification Pending Patent Approval

Snakebite envenomation is a serious public health challenge, particularly in rural areas with limited access to timely medical care. Accurate snake species identification is critical for administering the correct antivenom, as misidentification can lead to ineffective treatment. This project developed a predictive model that integrates epidemiological data, machine learning, and bite mark imaging to accurately identify the snake species responsible for envenomation.

Data Acquisition

Data for this project was acquired through a collaboration with Monosha Biotech, a serpentarium and venom lab. Monosha Biotech provided expert knowledge on the probabilities associated with different features, which were organized into an Associated Probability Matrix. This matrix represented the likelihood of a specific snake species being responsible for a bite based on factors like location and symptoms.

Bite Mark Imaging

Images were preprocessed through auto-orientation, resizing, and contrast adjustments to enhance clarity, while adaptive equalization highlighted subtle details. Augmentation techniques, like flips and rotations, were applied to introduce real-world variability and strengthen the model's generalization. The YOLOv8 deep learning model was used to detect bite marks and assess envenomation severity, contributing to the prediction of the snake species involved.

Synthetic Data Generation

Due to the scarcity of real-world snakebite data, synthetic data was generated to bridge this gap. This dataset simulated epidemiological factors, including location, time of day, and symptoms, with a cross-covariance matrix capturing dependencies to maintain real-world correlations. By replicating these patterns, the synthetic data expanded the model's training scenarios, enhancing its predictive accuracy and robustness.

Model Training

The predictive model was trained using both real and synthetic data, incorporating key features like location, time of bite, symptoms, and bite mark images. After optimizing the model, it achieved strong performance metrics, demonstrating its effectiveness in accurately predicting the snake species responsible for a bite. This approach significantly improved predictive accuracy, combining epidemiological data and imaging for real-world applications.

Results

The results of the predictive model showcased its effectiveness in accurately identifying snake species based on a combination of epidemiological data and bite mark imaging. By integrating real and synthetic data, the model was able to predict the responsible species with a high degree of accuracy, making it a valuable tool for improving snakebite management. The model was tested and validated through many real-world case studies demonstrating its practical application and robustness.

Case Study 1

The model accurately predicted a Russell’s viper bite, aligning with clinical observations, despite the absence of typical envenomation symptoms like swelling and pain.

Clinical assessment revealed an absence of typical symptoms of venomous snakebite. The patient reported no pain, burning, or discoloration at the bite site, and there was no local swelling or blistering. Systemic symptoms of envenomation, such as drowsiness, respiratory difficulty, abdominal pain, ptosis, swallowing issues, chest and throat burning, hemorrhage, or paralysis, were also absent.

Case Study 2

A bite showing no local symptoms but distinct bite marks was accurately identified as caused by a Spectacled cobra, consistent with the ground truth.

Upon clinical assessment, two distinct bite marks were observed on the hand, but no local symptoms of envenomation were present. The absence of pain, burning, swelling, or skin discoloration suggested a lack of venom-induced effects. Systemically, the patient showed no signs of envenomation, with no drowsiness, respiratory issues, abdominal pain, ptosis, swallowing difficulty, chest or throat burning, hemorrhage, or paralysis. The incident occurred during heavy rain

Page updated

Google Sites

Report abuse