🤖 machine learning

Learning from waste: Machine Learning for health risk prediction and computer vision-based sorting in Ghana

This study demonstrates that machine learning can quantitatively validate the link between waste disposal practices and health risks in Ghana while simultaneously offering an affordable, camera-based solution for automated waste sorting in resource-constrained settings, though it emphasizes that technological success must be paired with institutional support to achieve public health improvements.

Original authors: Hilda Adwubi Osei, Catherine Tenewaa Osei, Desdemona Yaa Asobayire

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hilda Adwubi Osei, Catherine Tenewaa Osei, Desdemona Yaa Asobayire

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In many growing cities across the world, the way people throw away their trash is changing the air they breathe and the water they drink, often with serious consequences for their health. When waste is left to rot in open spaces, burned in the streets, or dumped into rivers, it creates breeding grounds for disease and pollution. Scientists have long known that these bad habits are linked to sickness, but usually, they can only describe the connection in general terms, noting that more trash in a stream often means more cases of stomach illness in the neighborhood. What has been missing is a way to measure exactly how a specific way of throwing things away predicts a specific kind of sickness for a specific person, and a practical, affordable way to fix the sorting problem before the waste even reaches the dump. This is where the work of researchers in Ghana steps in, using modern computer tools to turn messy, real-world observations into clear, actionable data.

In the community of Atonsu, near the city of Kumasi, a team of researchers began by listening to the people who live there. In a 2022 field study, residents explained that they saw a clear pattern: when they burned their trash, they got headaches and heart trouble; when they dumped it in bushes, they faced malaria and snake bites; and when they threw it into streams, they suffered from cholera and diarrhea. While these stories were powerful, they were just descriptions. The researchers wanted to know if a computer could learn from these stories to predict illness based on how waste was handled. They gathered survey data from hundreds of residents, asking them about their disposal habits and their health. They then taught a computer program, known as a random forest classifier, to look for patterns in this information. The program learned to sort people into groups based on the type of sickness they reported and how they threw away their trash. When tested on a group of people who were actually sick, the computer was able to correctly guess the type of illness about 63 percent of the time. This might not sound perfect, but it is a significant step forward. It proves that the link between trash and sickness is not just a feeling or a coincidence; it is a measurable pattern that a machine can learn. The computer found that the method of disposal was the most important clue, far more than the age or job of the person reporting the illness.

While the first part of the study looked at the health consequences, the second part tackled the solution: how to sort the trash efficiently without spending a fortune on expensive machinery. The original idea for fixing Atonsu's waste problem involved a complex machine with many different sensors, like infrared lights and metal detectors, to separate different types of trash. However, such machines are costly and difficult to maintain in places with limited resources. The researchers asked if a simple camera could do the same job. They built a system using a camera and a type of artificial intelligence called a convolutional neural network, which is designed to recognize images. They trained this system using a large collection of photos of different kinds of waste, such as cardboard, glass, metal, paper, plastic, and general trash. The result was striking. The camera-based system correctly identified the type of waste in nearly 88 percent of the cases it was tested on. This means that a simple camera, which could be attached to a conveyor belt or even a smartphone, can sort trash almost as well as the much more complicated multi-sensor machines, but at a fraction of the cost and complexity.

The study brings these two discoveries together to paint a clearer picture of the problem and a potential path forward. The computer model that predicts illness confirms that the community's observations were correct: the way you throw away your trash directly influences what kind of sickness you might get. The camera model shows that sorting that trash does not require a high-tech factory; it can be done with affordable, accessible technology. However, the researchers are careful to note that having a working model is not the same as having a solved problem. The ability to predict sickness or sort trash does not automatically fix the situation. The study points out that the real barriers in Ghana are often not technical but institutional, involving issues like funding, enforcement of laws, and the willingness of local authorities to act. The technology is ready to help, but it needs the support of people and policies to turn these predictions and automated sorts into better health and a cleaner environment. The work demonstrates that even in resource-constrained settings, modern data tools can validate local knowledge and offer practical, low-cost solutions, provided there is a clear plan to put them into action.

Technical Summary: Learning from Waste – Machine Learning for Health Risk Prediction and Computer Vision–Based Sorting in Ghana

Problem Statement
Inappropriate solid waste disposal in Ghana, particularly in rapidly urbanizing communities like Atonsu, Kumasi, poses significant public health and environmental risks. A 2022 field investigation in Atonsu established a descriptive, community-perceived link between specific household waste disposal methods (open burning, bush dumping, stream dumping, and regulated bin use) and distinct patterns of illness (e.g., malaria/snake bites, cholera/diarrhea, respiratory/cardiac issues). However, this relationship remained qualitative and anecdotal, lacking quantitative modeling to validate predictive power. Furthermore, while the original investigation proposed a complex, multi-sensor mechanical sorting system to address waste management, such infrastructure is often cost-prohibitive for resource-constrained settings. This study addresses two gaps: (1) the need to quantitatively model the relationship between disposal practices and illness categories using survey data, and (2) the need for a more affordable, deployable alternative to multi-sensor mechanical sorting.

Methodology
The study employs two distinct data-driven approaches:

Survey-Based Health Risk Prediction:
- Data: A structured survey was collected from 470 residents across eight sub-areas of Atonsu using KoboToolbox. The dataset included demographic variables and self-reported disposal methods mapped to four illness categories (plus a "None reported" class).
- Modeling: Three classifiers (Logistic Regression, Decision Tree, Random Forest) were evaluated using 5-fold stratified cross-validation. The Random Forest (RF) classifier was selected as the final model due to its slight performance edge and superior interpretability via feature importance scores.
- Training: The model was trained on a partition of $N=376$ and evaluated on a held-out test set of $N=94$ . A specific subset analysis ( $N=69$ ) focused exclusively on respondents who reported illness to isolate the predictive power of disposal methods on illness types.
Vision-Based Waste Sorting:
- Data: The study utilized the publicly available TrashNet dataset (~2,500 labeled images across six waste categories: cardboard, glass, metal, paper, plastic, and trash). No local images from Atonsu were used for training to avoid data scarcity issues, though this introduces a domain gap.
- Modeling: A MobileNetV2 architecture was employed using transfer learning. The development process involved correcting a preprocessing error where redundant rescaling layers degraded performance. The model underwent a two-phase training: initial training of the classification head with a frozen base, followed by fine-tuning of base layers.
- Evaluation: A strict 70/15/15 train/validation/test split was used. The model was evaluated once on the held-out test set ( $N=415$ ) to prevent overfitting.

Key Results

Health Prediction Model: On the full test set ( $N=94$ $N = 94$ ), the Random Forest model achieved a macro F1 score of 0.70. When restricted to the subset of respondents who reported illness ( $N=69$ $N = 69$ ), the model achieved a macro F1 of 0.63 and an accuracy of 0.62.
- Feature Importance: The "fell_sick" variable (indicating whether illness was reported) was the dominant predictor, separating the "None reported" class from illness categories. Among substantive predictors, disposal method emerged as the most important feature, outweighing demographics (age, education, occupation, sex, sub-area).
- Confusion: Misclassifications were concentrated among the three illness categories associated with unsafe disposal (Cases A, B, and C), suggesting the model distinguishes sick from healthy well but struggles to differentiate specific illness types, likely due to symptom overlap or self-reporting imprecision.
Image Classification Model: The fine-tuned MobileNetV2 model achieved 88.2% test accuracy and a macro F1 score of 0.87 on the TrashNet test set.
- Performance: The model performed strongest on cardboard, glass, and metal. The "trash" category (heterogeneous non-recyclables) was the weakest performer (F1 0.77).
- Error Analysis: Confusion occurred primarily between visually similar materials (e.g., glass/metal, paper/cardboard) or within the heterogeneous "trash" class, consistent with visual similarity rather than random error.

Key Contributions

Quantification of Community Health Links: The study provides the first quantitative evidence supporting the community-reported link between specific waste disposal methods and illness categories in Atonsu. It demonstrates that disposal method is a learnable, substantive predictor of illness type, moving the relationship from descriptive correlation to a testable statistical model.
Technological Substitution for Resource-Constrained Settings: The study proposes and validates a vision-based waste sorting system as a viable, lower-cost alternative to the multi-sensor mechanical designs previously proposed. By achieving high accuracy (88.2%) using a single camera and a lightweight model (MobileNetV2), it offers a pathway for automated sorting that does not require expensive infrared, inductive, or capacitive sensor arrays.
Methodological Transparency: The paper explicitly documents and corrects a preprocessing error in the transfer learning pipeline (redundant rescaling), providing a cautionary example for similar applications using pretrained Keras architectures.

Significance and Claims
The authors claim that their findings offer a "qualified yes" to the feasibility of data-driven interventions in this context, but with important caveats:

Modest Predictive Power: The health prediction model confirms a relationship but does not achieve clinical diagnostic reliability. The authors emphasize that disposal method alone cannot serve as a definitive diagnostic tool for individuals, as the signal is modest (F1 0.63) and confounded by self-reporting limitations.
Technical vs. Institutional Barriers: While the image classifier demonstrates that vision-based sorting is technically feasible and cheaper than sensor-based systems, the authors argue that technological performance alone does not guarantee public health improvements. Citing literature on Ghana's waste governance, they assert that the primary bottlenecks are institutional (budgetary allocation, enforcement of by-laws) rather than technical.
Deployment Reality: The vision-based approach is presented as a "candidate" for deployment, not a proven solution for Atonsu's specific waste stream. The authors acknowledge a significant domain gap: the model was trained on clean, staged images (TrashNet) rather than the mixed, degraded, and context-embedded waste found in Atonsu.

In conclusion, the study demonstrates that community-level environmental challenges can be extended with modern data-driven methods without large-scale infrastructure. However, the authors maintain a modest stance, urging that future work must focus on on-site validation of the vision model, longitudinal studies to establish causality, and the development of institutional pathways to translate these technical tools into actual policy and health outcomes.

More like this