Health Care Waste Classification Using Deep Learning Aligned with Nepal's Bin Color Guidelines
Suman Kunwar, Prabesh Rai · 2025
Read Health Care Waste Classification Using Deep Learning Aligned with Nepal's Bin Color Guidelines by Suman Kunwar, Prabesh Rai (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-07450.txt.
Introduction
Suman Kunwar and Prabesh Rai benchmark learning models that sort health care waste by photo. Their paper appeared in 2025. Nepal's hospitals are growing, and poor waste sorting spreads infection and endangers waste handlers. The country sorts waste into bins by color under national rules. The team trained five well-known image models on a combined waste data set, testing each with repeated cross-checks.
One model, YOLOv5-s, sorted best at about 95 percent accuracy. Another was faster by milliseconds; a third was accurate but slowest. Statistical tests confirmed the gaps were real. The winning model was put on the web, with its labels mapped to Nepal's bin colors. The authors note the data set is small and call for larger, more local collections.
Catalog record
| CREATOR | Suman Kunwar, Prabesh Rai |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 12 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-07450
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Health Care Waste Classification Using Deep Learning Aligned with Nepal's Bin Color Guidelines Suman Kunwar, Prabesh Rai Abstract The increasing number of Health Care facilities in Nepal has added up the challenges on managing health care waste (HCW). Improper segregation and disposal of HCW leads to contamination, spreading of infectious diseases and risk for waste handlers. This study benchmarks the state of the art waste classification models: ResNeXt-50, EfficientNet-B0, MobileNetV3-S, YOLOv8-n and YOLOv5-s using stratified 5-fold cross-validation technique on combined HCW data. YOLOv5-s achieved the highest accuracy (95.06%) but fell short with the YOLOv8-n model in inference speed with few milliseconds. The EfficientNet-B0 showed promising results of 93.22% accuracy but took the highest inference time. Following a repetitive ANOVA test to confirm the statistical significance, the best performing model (YOLOv5-s) was deployed to the web with bin color mapped using Nepal’s HCW management standards. Further work is suggested to address data limitation and ensure localized context. Keywords: health care waste, deep learning, waste classification, computer vision, Nepal health care waste standards 1 Introduction Managing HCW is a pressing challenge fueled by increasing waste generation by health care facilities. Improper segregation and disposal of HCW results in the spread of infectious diseases, contamination and puts risk towards waste handlers. Traditionally, HCW are done manually which is often labor-intensive, prone to human errors and poses a threat to waste handlers. Numerous researchers have demonstrated the feasibility of AI powered solutions for automated waste sorting and its management. Studies have applied deep learning (DL) ResNeXt-50 [ 1 ] and EfficientNet_B7 [ 2 ] models, with transfer learning to get higher accuracy results in HCW classifications, even with small datasets. The intersection of IoT with models like YOLOv5 [ 3 ] , YOLOv8 [ 4 ] and MobileNetV3 [ 5 ] has shown promising results on automated sorting of medical waste showcasing real world implementation. In the study, Zhou et al. used ResNeXt-50 and achieved 97.2% accuracy on an eight-class private dataset, while other fine tune transfer learning on EfficientNet-B7, reaching an accuracy of 99% on six class dataset. Automated sorting with the help of sensors, robotic arms and edge devices has also shown promising results. These approaches have reported with 97.93% accuracy highlighting the real-world deployment in healthcare environments [ 6 ] . Moktar et al. used YOLO combined with IoT for the detection of HCW and achieved a mAP of 98% using 6,7860 images. With the promising results, challenges like small dataset size, poor image quality and controlled environment testing persists raising concerns [ 7 ] on the scalability, infrastructure feasibility and adoption with existing waste management systems. These put forward the needs of expansion of dataset to generalize the detection, needs of real time deployment with real world settings and integration of AI with IoT for automation. In Nepal, there are 16,611 health care facilities [ 8 ] serving people that follows to the guidelines set by National Health Care Waste Management Standards and Operating Procedures to manage HCW. These guidelines divide waste in General Health Care Waste (General HCW) and Hazardous Health Care Waste (Hazardous HCW). The Hazardous HCW spans in multiple categories and is mapped with various color codes as depicted in Figure 1 . Figure 1: Color codes for waste segregation The mapping of individual items based on waste category, label and color codes are shown in Table 1 . Table 1: Healthcare Waste Categories, Types, Color Codes, and Items Waste Category Type Color Code Items Non-risk HCW Biodegradable Green Organic wastes; peels of vegetables and fruits; rotten or stale foods Non-Biodegradable Blue Paper (A4, duplex, medicine covers, normal papers); plastic (polythene bags, wrappers, packaging plastic materials, cling wraps); bottles (water, soft drinks); vials; saline bottles Risk HCW Infectious Red Cotton, gauze, used gloves, IV sets, blood bags, urine bags, used PPE, diapers, pads, sputum containers, collection tubes, dialyzer Pathological Body organs or tissues; placenta Cytotoxic Unused or expired cytotoxic drugs; chemotherapy agents Pharmaceutical Expired medicines Sharps Needles, blades, scalpels, broken glass, lancets, suture needles Chemical Yellow Discarded chemicals, reagents, solvents, disinfectants Radioactive Black X-ray films; radioisotopes A study by Patel et al. on HCWM practices and knowledge in 10 government district hospitals in Madhesh Pradhesh found that the waste handlers showed low adherence to HCWM guidelines and lacked sufficient knowledge, despite having adequate facilities for managing HCM [ 9 ] . This poses a risk towards waste handlers and residents nearby hospitals as highlighted by Karki et al. study [ 10 ] . Furthermore, some studies suggest that with proper sorting the amount of HCW can be reduced in Nepal’s context [ 11 ] . An example of waste bins from one of the hospitals is shown in Figure 2 . Figure 2: Waste bins at Bhaktapur Cancer Hospital, Nepal The studies demonstrate the effectiveness of DL models on HCW segregation, suggesting that proper sorting can help in waste minimization and de-risk waste handlers. This study, therefore, explores state of the art waste classification models: YOLOv8-n, MobileNetV3-S, ResNeXt-50, EfficientNet-B0 and YOLOv5-s using stratified K-fold cross-validation method, which is preferred for optimizing sample selection and robust prediction results [ 12 ] . The best performing model mapped with Nepal bin color guidelines is deployed for public usage. 2 Materials and Methods The section discusses the dataset used in this study and the methods used to benchmark the models. 2.1 Dataset and Preprocessing We use a combined dataset: Medical Waste Dataset 4.0 [ 13 ] and Pharmaceutical and Biomedical Waste dataset [ 14 ] to cover major categories of Nepal HCW. Medical Waste Dataset 4.0 was acquired from Tuscany Region, using OAK 4.0 camera device and span across: gauze, glove pair (latex, nitrile, surgery), glove single (latex, nitrile, surgery), medical cap, medical glasses, shoe cover pair, shoe cover single, test tube and urine bag. The RGB images are of size 1920 x 1080. The Pharmaceutical and Biomedical Waste dataset is collected by the Engineering UBU and consists of: Body Tissue or Organ, Organic wastes, equipment-packaging (plastic, paper, metal, glass), Syringe needles, Gauze, Gloves, Mask, Syringe and Tweezers. In the Pharmaceutical and Biomedical Waste dataset, the Gloves contain mixed gloves images, a more generalized class. As the separated types of gloves are already present in the Medical Waste Dataset 4.0, we removed this class to remove the bias. The dataset is divided into 5 stratified folds; each fold containing 80% train and 20% validation data, and can be found in Kaggle 1 1 1 https://www.kaggle.com/datasets/sumn2u/medical-waste . The oversample class images were reduced using the median class count value to handle the synthetic bias. For under sampling classes, data augmentation techniques such as flipping and brightness contrast were used. 2.2 Model Training The experiments were conducted using two NVIDIA Tesla T4 GPUs. Models were tested using 5-fold stratified cross validation, with each fold trained for 30 epochs using MobileNetV3-S, ResNeXt-50, EfficientNet-B0 and YOLO models (v5-s/v8-n). The stratified K-fold technique was used to maintain the same percentage of instances for each label [ 15 ] . DL models such as MobileNetV3-S, ResNeXt-50 and EfficientNet-B0, used pre-trained weights trained in ImageNet dataset with frozen base model layers and a custom classification head. In contrast, YOLO models (v5-s/ v8-n) are trained from scratch. The