Source-linked AI summary
Learning from waste: Machine Learning for health risk prediction and computer vision-based sorting in Ghana
Hilda Adwubi Osei, Catherine Tenewaa Osei, Desdemona Yaa Asobayire
TL;DR
Improper waste disposal poses public-health and environmental concerns, while prior evidence in Atonsu documented disposal–illness links only descriptively. This study applies survey-based illness prediction and camera-based waste classification, finding measurable illness-type signal and strong benchmark sorting performance.
Problem
Prior Atonsu evidence linked household waste disposal with illness descriptively, leaving its quantitative relationship insufficiently validated amid substantial sanitation and waste-management burdens.
Method
The study combines a Random Forest classifier for survey-based illness categories with a MobileNetV2 image classifier for automated waste sorting.
Results
The survey model achieved a macro F1 score of 0.63 among 69 ill respondents, while the image model achieved 88.2% test accuracy and macro F1 = 0.87.
Takeaways & Limitations
The findings quantitatively support a disposal–illness relationship and indicate potential for camera-based waste sorting in resource-constrained environments.
Takeaways & Limitations
The cross-sectional survey supports association but not causal inference, and illness categories were self-reported rather than clinically confirmed.
Abstract
from arXiv · showhide
The inappropriate disposal of solid waste remains a significant public health and environmental concern worldwide, including in Ghana. Poor sanitation and improper waste management practices contribute to substantial economic costs and avoidable deaths annually. In 2022, a field study in Atonsu, Kumasi, Ghana, reported a community-perceived relationship between household waste disposal and illness patterns, but only through descriptive analysis without quantitative validation. This study extends that investigation using two data-driven approaches. First, a Random Forest classifier was developed to predict illness categories using waste disposal practices and demographic survey data. On a held-out group of respondents who reported illness (N=69), the model obtained a macro F1 score of 0.63, with disposal method emerging as the most important substantive predictor of illness type. Second, a MobileNetV2 image classification model enabled automated waste sorting via visual recognition, achieving 88.2% accuracy and a macro F1 score of 0.87 on the test set (N=415). The vision-based approach offers an affordable, camera-driven alternative to complex multi-sensor systems, making it highly suitable for resource-constrained settings. Taken together, the findings provide quantitative evidence for a community health relationship previously documented only qualitatively. They demonstrate the potential for automated waste-sorting in low-resource environments. Importantly, the results illustrate that technological performance alone does not guarantee public health improvements; effective institutional support and implementation are equally necessary.
1 Introduction
The introduction frames unmanaged waste as a growing global and Ghanaian public-health problem, then presents two data-driven models extending a qualitative investigation in Atonsu, Kumasi. The study predicts illness categories from disposal and demographic survey data and evaluates automated waste sorting through image classification.
- Global waste burden: 2.56 billion tonnes of municipal solid waste were generated globally in 2022, with projections reaching 3.86 billion tonnes by 2050.The introduction attributes this growth largely to urban expansion, population growth, and changing consumption behavior.
- Global waste burden: 30% of global waste was mismanaged in 2022, especially in low- and middle-income countries, with the share projected to decline to around 20% by 2050.Mismanagement includes open dumping, informal burning, and unmanaged landfilling where infrastructure has not kept pace with waste generation.
- Ghanaian public-health burden: GH¢420 million annually, equivalent to 1.6 percent of national GDP, was the estimated economic cost of poor sanitation in Ghana.The cited 2012 study attributed 74 percent of this cost to premature mortality and estimated 19,000 annual deaths from diarrheal disease.
- Study rationale and contributions: The study extends a 2022 Atonsu investigation by developing a Random Forest classifier that predicts illness categories from disposal methods and demographic variables.The earlier investigation associated improper waste disposal with recurrent illnesses including cholera, malaria, typhoid, diarrhea, and colds and flu.
- Study rationale and contributions: 0.63 was the Random Forest model’s macro F1 score across four illness categories among respondents who reported illness.The model used survey data collected across Atonsu’s eight sub-areas.
2 Related Work
Prior work links improper waste disposal with community health outcomes and increasingly applies machine learning to sanitation and environmental data. Waste-sorting research has shifted from sensor-based mechanical systems toward camera-based vision models, motivating a more deployable approach for resource-constrained settings.
- Health and environmental relationships: Prior Sub-Saharan African research consistently links improper solid waste disposal with community health outcomes, mainly through descriptive and qualitative methods.A systematic review synthesized 27 studies across eight Sub-Saharan African countries, including Ghana, into contamination pathways involving soil, air, and water.
- Machine learning for health and sanitation: Machine learning has been applied to survey-based health and environmental data across Sub-Saharan Africa, typically at regional scale.One study compared Random Forest, Decision Tree, XGBoost, Logistic Regression, and an artificial neural network using DHS data from 34 countries and over 500,000 households.
- Automated waste sorting: Waste-sorting automation has developed along two paths: sensor-based mechanical systems and vision-based classification.West African engineering projects include Arduino-controlled systems using infrared and inductive sensors to separate metals from non-metals on conveyor systems.
- Vision-based classification: 88% test accuracy and a macro F1 score of 0.87 were achieved by this study’s MobileNetV2 model on a proper 70/15/15 train/validation/test split.The result was comparable to mid-tier published TrashNet results and above the cited from-scratch or non-transfer-learning baselines.
- Study motivation: Vision-based classification can provide practically useful waste-sorting accuracy without the physical sensor infrastructure required by mechanical designs.This motivates replacing the proposed multi-sensor mechanical sorter with a camera-based classifier for resource-constrained community settings.
3 Study Area and Background
The study is situated in Atonsu, Kumasi, where recurring illness patterns and widespread concern about improper waste management motivated investigation. Earlier community findings were descriptive and correlational, prompting individual-level predictive modeling.
- Problem motivation: Recurring hospital visits revealed disproportionately frequent reports of cholera, malaria, typhoid, diarrhea, colds and flu, and snake bites among Atonsu residents.This observation led to the hypothesis that household waste disposal practices might relate to illness patterns.
- Community background: 80 percent of approximately 1,000 residents identified improper waste management as a significant community problem.Residents reported open burning, dumping in nearby streams or water bodies, and waste collection through Zoomlion among dominant disposal practices.
- Study rationale: The earlier four-case framework identified an association between disposal method and health outcome but did not quantify its strength, consistency, or individual-level predictive value.This limitation motivated the survey-based supervised classification model developed later using disposal and demographic data.
4 Methodology
The study uses separate field-survey and public-image datasets to develop illness-category prediction and automated waste-classification models. The methodology includes survey-data cleaning, stratified evaluation of candidate classifiers, and corrected MobileNetV2 transfer learning on TrashNet.
- Datasets: Two datasets underpin the study: an original field survey and a publicly available image dataset used as-is.The datasets support the survey predictive model and the image classification model, respectively.
- Survey data: Survey-data cleaning, including consent filtering and missing-value handling, produced 470 cleaned respondent records for modeling.The pipeline also removed duplicate respondent records and rows missing the target illness variable.
- Image model: TrashNet provided approximately 2,500 labeled images across six categories and was used in its original, unmodified form.The image model used MobileNetV2 transfer learning with a corrected preprocessing pipeline and a 70/15/15 train/validation/test split.
- Survey model: Three classifiers—logistic regression, decision tree, and random forest—were evaluated with stratified cross-validation and an 80/20 split, leaving 94 respondents as an untouched test set.The training partition contained N = 376 records from the full N = 470 dataset.
- Survey model: 0.697 cross-validated macro F1 was achieved by the best Random Forest configuration during model selection on the N = 376 training partition.The tuned Random Forest used class_weight=None, max_depth=5, min_samples_leaf=10, and n_estimators=100.
- Image model: 88% test accuracy and a 0.87 macro F1 score were achieved by the fine-tuned MobileNetV2 model across six TrashNet material categories.Training used a frozen-base phase followed by partial unfreezing and fine-tuning at a reduced learning rate.
5 Results
The Random Forest model showed moderate performance in predicting illness categories, with disposal-method features providing the main substantive signal after illness-status indicators. MobileNetV2 achieved strong waste-material classification performance, although errors remained between visually similar materials and heterogeneous trash.
- Random Forest illness prediction: The Random Forest achieved 0.72 overall accuracy and a 0.70 macro F1 score across five categories on the held-out test set (N = 94).Case-level performance varied, with Case A reaching F1 0.63 and Case B reaching F1 0.53.
- Random Forest illness prediction: Among respondents reporting illness, accuracy fell to 0.62 and macro F1 to 0.63 on the clinically relevant subset (N = 69).This subset covered Cases A through D and had a weighted average F1 of 0.62.
- Random Forest illness prediction: Illness-status indicators accounted for approximately 60% of Random Forest feature importance, while disposal-method indicators were the most important remaining substantive features.The illness-status features primarily reflected the near-deterministic “None reported” class rather than illness type.
- MobileNetV2 waste sorting: MobileNetV2 achieved 88.2% test accuracy and a 0.87 macro F1 score on 415 waste images.The weighted-average F1 score was 0.88, while trash was the weakest-performing class with F1 0.77.
- MobileNetV2 waste sorting: The main vision-model errors involved glass–metal and paper–cardboard/metal confusion, while heterogeneous trash showed the weakest and most even performance.These patterns were consistent with similarities in reflectivity, shape, color, and texture under varying lighting conditions.
6 Discussion
The discussion evaluates whether disposal practices contain quantifiable illness-prediction signal and whether camera-based sorting is a practical alternative to the proposed multi-sensor system. It also emphasizes careful interpretation, deployment tradeoffs, and the institutional follow-through needed for intervention.
- Discussion overview: The study addresses whether the community-reported disposal–illness relationship contains genuine predictive signal and whether camera-based sorting can achieve practically useful accuracy.These are the two related but distinct questions motivating the discussion.
- Survey model interpretation: The final Random Forest’s dominant predictor is fell_sick, not disposal method, because it near-deterministically separates respondents reporting illness from those reporting none.The discussion characterizes this as expected and largely tautological rather than substantive.
- Automated sorting: The vision approach is evaluated as an alternative to Concept 2, a multi-sensor conveyor system combining infrared, inductive, capacitive, and optical glass detection.Concept 2 was selected as the strongest of three candidate designs in the original weighted decision matrix.
- Limitations: The TrashNet-trained vision classifier has not been validated on mixed, degraded, or context-specific waste from Atonsu.TrashNet contains individually staged items photographed under comparatively controlled conditions.
- Implementation: Moving from working models to community intervention requires more than data collection and modeling, because reviewed monitoring programs often lacked reporting or corrective action.The discussion uses this literature as a cautionary benchmark for implementation in Atonsu or comparable communities.
7 Limitations
The study’s limitations primarily concern the survey model’s cross-sectional design and the image model’s mismatch between controlled training data and Ghana’s real-world waste stream.
- Survey predictive model: The cross-sectional survey captures disposal practices and reported illnesses at one time point, supporting association but not causal inference.Its predictive relationship between disposal method and illness type is therefore not evidence that disposal causes a given illness.
- Survey predictive model: The survey model’s limitations are concentrated in the study’s cross-sectional predictive analysis.The model uses disposal practices and reported illnesses measured at a single point in time.
- Image classification model: The image model faces a domain gap because TrashNet contains individually staged waste items photographed in relatively controlled lighting and background conditions.These training and evaluation conditions differ substantially from the actual waste stream in Atonsu.
- Image classification model: Atonsu’s actual waste stream is frequently mixed and degraded by open burning, unlike the isolated items represented in TrashNet.This difference limits direct generalization from the dataset to the intended deployment context.
8 Conclusion and Recommendations
The study quantitatively extended the 2022 community-reported disposal–illness framework using a Random Forest model and identified disposal method as a learnable predictor of illness type. Its recommendations refine public-health messaging and call for real-world image-classifier validation beyond controlled TrashNet conditions.
- Contributions: 0.63 macro F1 on 69 held-out sick respondents showed that disposal method is a learnable predictor of illness type.This extends the 2022 study’s case-based associations with quantitative evidence.
- Recommendations: Disposal method was the leading substantive predictor of illness type among sick respondents, supporting more targeted public-health messaging in Atonsu.The recommendation builds on earlier calls for citizen education about improper waste disposal’s health risks.
- Future Work: On-site validation should test the image classifier on mixed, burned, water-exposed waste embedded in real Atonsu disposal contexts.This would address the domain gap and assess whether controlled TrashNet accuracy transfers to real-world conditions.
9 Author Contributions
H.A. Osei led and directed the study, while the authors contributed across conceptualization, investigation, technical development, analysis, writing, and related work. All authors read and approved the final manuscript.
- Leadership: H.A. Osei originated and directed the 2022 field investigation and its extension into the present study.She was also the lead and corresponding author.
- Author Contributions: Hilda Adwubi Osei handled conceptualization, methodology, software, formal analysis, investigation, data curation, writing, visualization, and project administration.Her writing contributions covered the original draft and review and editing.
- Author Contributions: Catherine Tenewaa Osei contributed to health outcomes framing, investigation, and review and editing.Her conceptualization role specifically addressed health outcomes framing.
- Author Contributions: Desdemona Yaa Asobayire investigated mechanical and sensor-based waste management systems and related work, and contributed to review and editing.Her investigation covered both waste management systems and the Related Work section.
- Manuscript Approval: All authors read and approved the final manuscript.This represents the authors’ collective final-manuscript approval.
A Survey Questionnaire
The reproduced questionnaire documents the instrument used in the 2022 Atonsu Bukuro field investigation. It covers resident characteristics, waste-management practices, perceived impacts, illness, and proposed solutions.
- Instrument purpose: The original questionnaire instrument is reproduced in full to support reference and reproducibility.It was administered during the 2022 field investigation described in Section 3.
- Instrument purpose: The questionnaire investigates matters arising from improper waste management among residents of Atonsu Bukuro.Respondents were asked whether improper waste management is a local problem and to explain their reasons.
- Respondent characteristics: Respondents report sex, age, educational background, residency, and occupation.Education options include junior high school, senior high school, tertiary, and illiterate.
- Waste and health: Questions assess waste-disposal methods and whether respondents consider those methods harmful.The instrument also asks whether harmful disposal methods have affected respondents.
- Waste and health: The instrument asks about sickness, frequent complaints, solutions for improper waste management, environmentally safe control methods, and useful project comments.These items capture illness experiences and respondents’ preferred interventions.
B Additional Figures
The appendix provides supporting figures from model development, including feature importances and held-out-test confusion matrices for the survey and image classification models.
- Appendix purpose: The appendix contains supporting figures referenced in Section 5 but not fully reproduced in the main text.These figures were generated during model development.
- Survey model: The final Random Forest survey model’s top 15 feature importances are dominated by fell_sick_Yes and fell_sick_No, followed by four disposal-method indicators.The disposal-method indicators appear in descending order after the two illness-status features.
- Survey model: N = 94 respondents comprise the full held-out test set for the survey predictive model’s five-category confusion matrix.The categories include “None reported.”
- Image classification model: N = 415 images comprise the held-out test set for the image classification model’s six-category TrashNet confusion matrix.The matrix covers all six TrashNet waste categories.