A camera-based artificial-intelligence model classified waste photographs with 88.2% accuracy in a test of 415 held-out images, but the images came from the public TrashNet benchmark rather than waste collected in Atonsu. The same study found a much more tentative signal when it used household survey information to predict reported illness categories. The contrast is central: the sorting model has stronger technical evidence under controlled conditions, while the health-related finding is modest and limited.
Two questions, two datasets
The preprint asks two related questions: whether a community-reported relationship between waste disposal and illness can be quantified with supervised learning, an approach in which a model learns patterns from labelled examples, and whether a camera system could offer a practical alternative to multi-sensor waste sorting. It combines an Atonsu survey with the original public TrashNet dataset, which contains approximately 2,500 labelled images across six waste categories.
The survey was stratified across Atonsu's eight sub-areas, with bin users deliberately oversampled. After cleaning, the modelling set contained 470 records: 376 were used for training and 94 were kept as an untouched test set. For the image arm, MobileNetV2 used a 70/15/15 split for training, validation and testing, first keeping its base fixed and then fine-tuning it; model choices were made from validation data before one final test evaluation.
The health prediction was weaker than the headline score
Researchers tuned the survey candidates with five-fold stratified cross-validation. Macro F1, a score that gives each category equal weight, was 0.697 for Random Forest, compared with 0.678 for logistic regression. The edge was small relative to fold-to-fold variability, but Random Forest was selected for its performance and interpretability.
On the 94-record full held-out survey test set, Random Forest reached 72% accuracy and a macro F1 of 0.70 across five categories, including "None reported". That result needs care: the "None reported" group was nearly determined by the illness-reporting field, so the full score partly benefited from separating people who reported no illness from the rest.
When the analysis was restricted to the 69 respondents who reported an illness, performance across Cases A–D was 62% accuracy, with a macro F1 of 0.63 and a weighted-average F1 of 0.62. The illness-only result relied on self-reported responses and a small cross-sectional subset, so the authors treat it as cautious, partial support for a health-outcomes signal rather than a causal finding.
Feature importance also put the result in perspective. Two indicators recording whether someone had "felt sick" accounted for approximately 60% of total feature importance, mainly because they separated the "None reported" category. Among the remaining predictors, indicators for burning waste, bush or open dumping, streams and bins ranked above demographic features. That ranking identifies patterns the model used; it does not show that any disposal method caused illness.
The camera result is a benchmark, not a field trial
The image model produced the stronger numerical result. On 415 held-out TrashNet test images, fine-tuned MobileNetV2 reached 88.2% accuracy and a macro F1 of 0.87. The authors interpret this as technical validation of a camera-based engineering substitution, not as evidence that the system is field-ready.
But the boundary around that number is important. TrashNet performance was not validated on Atonsu images, and 88.2% accuracy with a macro F1 of 0.87 is an upper-bound benchmark result under favourable, controlled conditions. It shows how the model performed on this test set; it does not establish the same performance for local waste.
A cautious reading
The authors present the two models as complementary: the survey model offers cautious, partial support for a health-outcomes signal, while the image model offers more confident technical validation of an engineering substitution. The distinction matters because the survey measures disposal practice and reported illness at one time point, so it supports association rather than causal inference.
Taken together, the findings support a cautious reading: the survey model provides partial predictive evidence, while the image result remains a benchmark result pending validation on Atonsu waste.
The manuscript is a preprint, arXiv:2608.25759v1, dated 26 August 2026. The code, notebooks, generated figures and results are publicly available in the authors' stated GitHub repository. Raw respondent-level survey data are not shared because linked household-location information could compromise privacy, although a de-identified aggregate may be available on reasonable request.
Paper data and sources
Original title: Learning from waste: Machine Learning for health risk prediction and computer vision-based sorting in Ghana
Authors: Hilda Adwubi Osei, Catherine Tenewaa Osei, Desdemona Yaa Asobayire
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text