Development of a Deep Learning Based Framework for Classification of Indian Venomous Snakes Integrated with Explainable Artificial Intelligence for primary and emergency care providers

This paper presents a clinically oriented deep learning framework utilizing a high-performing ResNeXt-50 model and Grad-CAM++ interpretability to accurately classify venomous versus non-venomous Indian snakes, aiming to assist primary care providers in rural settings with timely triage and treatment decisions.

The Gist

Imagine you are in a rural village in India, and someone has just been bitten by a snake. The clock is ticking. The most important question right now isn't "What kind of snake is this?" but rather, "Is this snake dangerous, or can we just watch and wait?"

In many remote areas, there are no herpetologists (snake experts) or doctors nearby to answer that question. People often guess based on local myths, which can lead to two terrible outcomes:

1. Panic: They treat a harmless snake like a killer, wasting precious and expensive medicine (antivenom) that could save someone else.
2. Tragedy: They ignore a deadly snake, thinking it's harmless, and the victim gets too sick to be saved.

This paper describes a new "Digital Snake Detective" designed to solve this problem using Artificial Intelligence (AI).

The Problem: The "Big Four" Myth

For a long time, doctors in India focused only on the "Big Four" deadly snakes. But nature is messy! There are other dangerous snakes (like the Hump-nosed pit viper) that the standard medicine doesn't even work on, and harmless snakes that look exactly like the dangerous ones. It's like trying to identify a criminal in a crowd, but you only know the faces of four specific people, while the real criminal is wearing a disguise that looks like one of the innocent bystanders.

The Solution: A Smart Camera App

The researchers built a smart system that acts like a super-powered camera lens. Here is how it works, step-by-step:

1. The Training Camp (Learning from Real Life)

Most AI is trained on perfect, studio-quality photos of snakes taken by nature photographers. But in an emergency, people take photos with shaky hands, bad lighting, and blurry backgrounds.

• The Analogy: Imagine teaching a student to recognize a car. If you only show them photos of shiny new cars in a showroom, they might fail when they see a rusty, muddy car on a dirt road.
• What they did: This team trained their AI on "messy" photos taken by real patients and field workers in emergency situations. They even used photos of dead snakes found on roads (roadkill) to teach the AI what the snakes look like in the wild.

2. The Brainy Detective (The AI Models)

They tested four different "brain" architectures (types of AI models) to see which one was the best detective:

• MobileViT, ConvNeXt, EfficientNet: These are like fast, lightweight detectives.
• ResNeXt-50: This turned out to be the Sherlock Holmes of the group. It was the most accurate at spotting the difference between a venomous and a non-venomous snake.

3. The "Why" Factor (Explainable AI)

One of the biggest fears with AI is that it might be a "black box"—it gives an answer, but you don't know why. What if the AI just guessed "Venomous" because the photo had a green background (grass)?

• The Analogy: Imagine a teacher grading a test. If they just give you an "A" without showing the work, you don't know if you actually learned the lesson.
• The Fix: The researchers used a tool called Grad-CAM++. Think of this as a heat map highlighter. When the AI says "This is venomous," the highlighter shows exactly where it is looking. It lights up the snake's head shape and body patterns, proving it isn't just guessing based on the background. It's looking at the "fingerprint" of the snake.

4. The Safety Net (Human-in-the-Loop)

The system isn't meant to replace doctors; it's meant to help them.

• The Analogy: Think of it like a GPS. The GPS tells you the fastest route, but you (the driver) are still in control. If the GPS says "Turn left," but you see a police car there, you ignore it.
• How it works: The AI makes a prediction, but a human expert (a doctor or a telemedicine specialist) can review it instantly. If the AI is unsure, the human steps in to make the final call.

The Results: A Near-Perfect Score

The "Sherlock Holmes" model (ResNeXt-50) got it right 96.7% of the time.

• It was incredibly good at spotting the dangerous snakes (high sensitivity).
• It rarely made mistakes on the harmless ones (high specificity).
• It achieved a near-perfect score on a test called "ROC-AUC" (0.995), which is like getting an A+ on a very difficult exam.

Why This Matters

This isn't just a cool tech demo; it's a lifesaving tool for the people who need it most.

• For the Village Health Worker: They can take a photo of the snake, get an instant "Dangerous" or "Safe" alert, and know exactly what to do next.
• For the Hospital: It stops them from wasting expensive antivenom on harmless snakes and ensures the right medicine is ready for the dangerous ones.
• For the Future: It creates a digital record of snakebites, helping governments understand where the dangers are and how to stop them.

The Bottom Line

The authors admit their system isn't perfect yet (it needs more data from different regions), but they have built a scalable, smart, and safe assistant. It's like giving every rural clinic a tiny, super-smart herpetologist in their pocket, ready to help save lives when the sun goes down and the snakes come out.

Technical Summary

1. Problem Statement

Snakebite envenoming is a critical global health crisis, particularly in India, which accounts for the highest proportion of snakebite deaths worldwide. Effective clinical management relies on the timely identification of the snake species to determine the appropriate antivenom. However, several barriers exist in peripheral healthcare settings:

• Lack of Expertise: Expert herpetologists are rarely available in rural or emergency settings.
• Clinical Oversimplification: Current protocols often rely on the "Big Four" species, ignoring other medically significant snakes (e.g., Hump-nosed pit vipers, King cobras) that require different treatments or are not covered by standard polyvalent antivenoms.
• Data Limitations: Existing AI studies often use high-quality, curated wildlife images that do not reflect the chaotic, poor-quality, and variable lighting conditions of real-world emergency photography (e.g., patient-taken photos, roadkill, blurred images).
• Binary Triage Gap: There is a lack of clinically oriented tools that perform a binary classification (Venomous vs. Non-venomous) specifically for triage, rather than just species-level identification.

2. Methodology

Dataset and Acquisition

• Source: Images were sourced from the VENOMS Registry (a prospective registry at Kasturba Medical College, Manipal), field expeditions, roadkill encounters, and emergency department submissions.
• Composition: The dataset includes 20 medically important Indian snake species.
  • Venomous: 1,110 images (including species like Russell's Viper, Common Krait, Hump-nosed Pit Viper, King Cobra).
  • Non-venomous: 384 images (including Rat Snake, Indian Rock Python, etc.).
• Characteristics: The dataset is intentionally "noisy," containing images with motion blur, poor lighting, partial visibility, and complex backgrounds to simulate real-world emergency conditions.

Data Preprocessing and Augmentation

To ensure model robustness and prevent overfitting to specific backgrounds or lighting:

• Resizing: All images resized to $320 \times 320$ pixels.
• Augmentation: Applied geometric and photometric transformations including:
  • Random rotation ($\pm45^\circ$).
  • Horizontal (50%) and vertical (30%) flipping.
  • Affine transformations (shear $\pm10^\circ$) and perspective distortion.
  • Random resized cropping and color jittering (brightness, contrast, saturation, hue).

Model Architecture and Training

Four state-of-the-art deep learning architectures were trained under identical hyperparameters for a fair comparison:

1. MobileViT-S (Lightweight Vision Transformer)
2. ConvNeXt-Tiny
3. EfficientNet-V2-S
4. ResNeXt-50 (32$\times$4d)

The models were trained to perform binary classification (Venomous vs. Non-venomous).

Explainability and Deployment

• Interpretability: Grad-CAM++ was employed to generate heatmaps, verifying that the model focused on anatomically relevant features (head shape, body contours, scales) rather than background artifacts.
• Human-in-the-Loop: A web-based interface was developed where AI predictions can be confirmed or overridden by medical experts, ensuring safety-critical oversight.

3. Key Contributions

1. Clinically Oriented Binary Classification: Unlike previous studies focusing on multi-species identification, this framework targets the immediate clinical need: distinguishing venomous from non-venomous snakes to guide triage and antivenom decisions.
2. Real-World Data Focus: The model is trained on "messy" real-world emergency images rather than idealized wildlife photography, enhancing its applicability in field settings.
3. Explainable AI (XAI) Integration: The use of Grad-CAM++ provides visual evidence that the model learns biological features, increasing trust among medical practitioners.
4. Human-AI Collaboration: The system integrates a verification layer where experts can validate AI outputs, addressing ethical concerns regarding AI autonomy in healthcare.

4. Results

Performance Metrics

Among the four architectures, ResNeXt-50 (32$\times$4d) demonstrated superior performance:

• Test Accuracy: 96.73%
• Sensitivity (Recall): 93.46% (Crucial for minimizing false negatives in venomous cases).
• Specificity: 100% (No false positives for non-venomous snakes).
• F1-Score: 96.73%
• ROC-AUC: 0.9950
• PR-AUC: 0.9959

Comparison:

• MobileViT-S: Lower sensitivity (85.98%), risking missed venomous cases.
• ConvNeXt-Tiny & EfficientNet-V2-S: Showed good performance but were slightly outperformed by ResNeXt-50 in stability and overall metrics.
• ResNeXt-50 achieved the lowest number of false negatives (7 out of test set), making it the safest choice for clinical screening.

Interpretability Findings

Grad-CAM++ visualizations confirmed that the ResNeXt-50 model attended to anatomically relevant regions, specifically the head morphology, body contours, and scale patterns. This indicates the model learned biologically meaningful features rather than relying on background noise.

5. Significance and Implications

• Clinical Triage Support: The system serves as a scalable decision-support tool for rural healthcare workers, paramedics, and telemedicine platforms (e.g., the VENOMS Helpline) to facilitate rapid risk stratification.
• Rational Antivenom Use: By accurately distinguishing venomous from non-venomous bites, the tool can reduce the unnecessary administration of antivenom (preventing anaphylaxis) and ensure timely treatment for actual envenomations.
• Registry Enhancement: Integration with snakebite registries (like VENOMS) can improve data fidelity by providing objective, AI-verified species classification, reducing reporting bias caused by patient recall.
• Ethical Framework: The study emphasizes that AI should be advisory, not prescriptive. The "human-in-the-loop" design ensures that final clinical decisions remain under expert supervision, adhering to principles of non-maleficence and autonomy.

Limitations & Future Work:
The study acknowledges limitations regarding the dataset size and its origin from a single institution. Future work aims to expand the dataset with multi-center validation, incorporate species-level identification for specific antivenom selection, and optimize models for edge devices (smartphones) in low-connectivity rural areas.

Conclusion:
This research successfully demonstrates that a deep learning framework, trained on real-world emergency data and augmented with explainability and human oversight, can provide a reliable, high-accuracy tool for the binary classification of Indian snakes, potentially saving lives by improving early triage in resource-limited settings.