AI Powers Smarter Healthcare Waste Management in Nepal with Color-Coded Precision

Researchers from DWaste (USA) and Lambton College (Canada) developed a deep learning system to automatically classify healthcare waste in Nepal, aligning results with the country’s color-coded bin guidelines. Their study found YOLOv5-s the most accurate model (95.1%), offering a promising tool to improve waste segregation and reduce health risks.

Nepal is facing a growing mountain of medical waste, with over 16,600 healthcare facilities across the country generating hazardous and non-hazardous refuse daily. The risks of poor handling are severe: improper segregation spreads infections, chemical residues seep into the environment, and waste handlers face constant exposure to dangerous materials. To address this, researchers Suman Kunwar of DWaste, USA, and Prabesh Rai of Lambton College, Canada, have turned to deep learning as a potential remedy. Their study benchmarks several state-of-the-art computer vision models to automatically classify medical waste in line with Nepal’s official bin-color guidelines. The goal is to reduce human error, enforce compliance with national standards, and ultimately protect both health workers and surrounding communities.

The Logic of Color-Coded Waste

Nepal’s national guidelines categorize healthcare waste into general and hazardous categories, each assigned to a distinct color-coded bin. Green bins receive biodegradable matter like fruit peels or food scraps, while blue bins are assigned to non-biodegradable plastics, bottles, and papers. Infectious waste, including gloves, gauze, and IV sets, should be placed in red bins, whereas yellow bins are reserved for chemicals, solvents, and disinfectants. Other specialized categories, from pharmaceuticals and cytotoxic drugs to pathological tissue and radioactive items, are similarly assigned their own colors. While the system is clear on paper, enforcement is often weak. Studies in government hospitals of Madhesh Province revealed that many waste handlers lacked knowledge of the standards despite having access to proper facilities. This shortfall leaves handlers and residents vulnerable, and it is precisely this gap that Kunwar and Rai hope to bridge with automation.

Building and Training the Models

The researchers relied on two complementary datasets to cover Nepal’s wide array of waste categories. The first, Medical Waste Dataset 4.0 from Italy, contained high-resolution images of gloves, gauze, shoe covers, urine bags, and test tubes. The second, a pharmaceutical and biomedical dataset from Thailand, contributed images of body tissue, syringes, tweezers, organic waste, and packaging. Redundant classes were removed to prevent overlap, and underrepresented categories were bolstered through augmentation techniques such as flipping and contrast adjustments. The combined dataset was then divided into five folds, with 80 percent dedicated to training and 20 percent to validation, ensuring rigorous cross-validation.

With this foundation, the team tested five prominent deep learning models: ResNeXt-50, EfficientNet-B0, MobileNetV3-S, YOLOv8-n, and YOLOv5-s. The first three used ImageNet-pretrained weights with custom classification heads, while the YOLO models were trained from scratch. Training took place on NVIDIA Tesla T4 GPUs over 30 epochs for each fold. To evaluate performance, the models were judged on accuracy, precision, recall, F1-score, and inference speed, measures that together capture both effectiveness and real-world viability.

The Clear Victory of YOLO Models

The results of the experiments were striking. ResNeXt-50 fared the worst, managing an accuracy of just 74.5 percent. MobileNetV3-S performed better at 91 percent accuracy, but with high inference time. EfficientNet-B0 produced 93.2 percent accuracy and excellent precision, yet it suffered from slow inference, taking nearly 445 milliseconds per image. The real stars were the YOLO models. YOLOv8-n achieved 94.7 percent accuracy at lightning speed, averaging just 9.3 milliseconds per inference, while YOLOv5-s recorded the highest accuracy overall at 95.1 percent with slightly slower inference times of around 11 milliseconds.

To confirm the significance of these differences, the researchers applied ANOVA and Tukey’s post-hoc tests, which showed the performance gaps were statistically robust. Visual comparisons further highlighted how YOLO models consistently outperformed rivals across all major metrics. Given its strong balance between accuracy and speed, YOLOv5-s was ultimately selected as the most practical model for deployment.

From Laboratory to Real-World Application

The researchers did not stop at benchmarking. They deployed the YOLOv5-s model as a publicly accessible application on Hugging Face, enabling users to upload images of medical waste for instant classification. A demonstration shows the system correctly identifying a syringe as infectious waste and mapping it to the red bin, proving its potential for hospital settings. This leap from experiment to deployment represents a bold attempt to bridge the lab-to-field gap.

Yet the study is candid about its limitations. Key waste categories from Nepal’s system, cytotoxic, radioactive, soiled, and liquid chemical, were absent from the dataset, leaving blind spots in classification. Moreover, much of the training data originated outside Nepal, where images were collected under controlled conditions. By contrast, waste in Nepali hospitals is often cluttered, occluded, or mixed in ways that could challenge the model. The dataset was also skewed toward common items like gloves and gauze. The authors stress the urgent need for locally collected data to strengthen model reliability and ensure it can operate effectively in real-world Nepali conditions.

Despite these challenges, the research demonstrates that AI-powered classification of health care waste is not just feasible but also practical. Automating segregation could significantly reduce infections, protect handlers, and raise compliance with Nepal’s color-coded disposal system. While such technology cannot replace the need for infrastructure upgrades and staff training, it can serve as a powerful assistive tool in a country still struggling with waste management fundamentals. By aligning artificial intelligence with public health policy, Kunwar and Rai’s work opens the door to a safer and more sustainable future for Nepal’s healthcare sector.