Multimodal Analysis of Factors Associated with Respiratory Symptoms in Colombian Informal Waste Pickers: A Study Based on Statistical Models and Machine Learning Algorithm.

Multimodal Analysis of Factors Associated with Respiratory Symptoms in Colombian Informal Waste Pickers: A Study Based on Statistical Models and Machine Learning Algorithm Yenny Andrea Rozo Silva, BSc, MSc 1 , Ana Delgado-García, BA 2 , Leidy Isabel Calderón Sierra, BSc, MSc 1 , and Raúl Aguilar-Elena, PhD 3 Abstract Objective: This study aimed to assess the prevalence of respiratory symptoms among informal waste pickers in Colombia and identify the contributing demographic, environmental, and occupational risk factors. Methods: A cross-sectional study was conducted with 179 informal waste pickers from four Colombian cities. Respiratory symptoms and demographic data were collected through questionnaires. Statistical methods including logistic regression, principal component analysis (PCA), Random Forest modelling, and K-means clustering were applied to identify predictors and patterns related to respiratory health. Results: The most frequently reported symptoms were cough (46.8%), phlegm (18.4%), and shortness of breath (19.6%). Logistic regression identified age as a significant predictor of respiratory symptoms, while Random Forest analysis highlighted cough as the strongest predictor, followed by age, race, and education level. K-means clustering revealed three groups, with older workers showing the highest prevalence of symptoms. PCA indicated that respiratory symptoms and demographic factors explained significant variance in health outcomes. Conclusions: Informal waste pickers in Colombia are at elevated risk for respiratory health issues, particularly older workers exposed to prolonged occupational hazards. Targeted interventions, including improved use of protective measures and policies addressing informal work conditions, are needed to mitigate these risks and improve workers’ health and safety. Keywords Informal waste pickers, respiratory symptoms, occupational health, environmental exposure, logistic regression, principal component analysis, Random Forest, K-means clustering, protective measures Introduction Environmental pollution caused by waste accumulation is a significant global challenge, leading to air, soil, and water contamination with severe public health implications. In response, governments worldwide have introduced policies to manage waste sustainably. (Abubakar et al., 2022; Pathak et al., 2024). The Rio Declaration’s Agenda 21 underscores reuse and recycling as key strategies for addressing solid waste. (United Nations Conference on Environment & Development. Agenda 21. Río de Janeiro, Brasil;, 1992. Disponible En: Https://Sustainabledevelopment.Un.Org/Outcomedocuments/ Agenda21, n.d.), aligning with the 2030 Agenda for Sustainable Development (Goal 12), which promotes waste reduction, recycling, and sustainable consumption (Walsh et al., 2022). Recycling plays a crucial role in mitigating environmental damage by conserving resources, reducing greenhouse gas emissions, and extending landfill lifespans. In Colombia, the recycling sector significantly contributes to waste management by recovering materials such as paper, plastic, and metals, lowering waste disposal costs and alleviating environmental burdens. (Abruzzese & Bandura, 2017). However, despite 1 Faculty of Business Sciences, Safe, Healthy, and Sustainable Work Environments Research Group (GIALSSS), Minuto de Dios University Corporation - UNIMINUTO, Bogotá, Colombia 2 IA Research Group BISITE, Universidad de Salamanca, Salamanca, Spain 3 Faculty of Social and Legal Sciences, Occupational Risk Prevention and Occupational Health and Safety Research Group (GPRL), Valencian International University - VIU, Valencia, Spain Corresponding Author: Raúl Aguilar-Elena, Prof. Industrial Hygiene, Occupational Risk Prevention and Occupational Health and Safety Research Group (GPRL). Valencian International University. (VIU), 0034666501202, C/ Pintor Sorolla, 21, 46002-Valencia, Spain. Email: [email protected] International: Research, Clinical, or Policy Hispanic Health Care International 1–15 © The Author(s) 2025 Article reuse guidelines: sagepub.com/journals-permissions DOI: 10.1177/15404153251342734 journals.sagepub.com/home/hci these advancements, waste management remains a major issue in Latin America. More than 40 million people lack access to basic urban waste collection, and approximately 90% of waste ends up in landfills. Additionally, informal waste collection remains prevalent, with around 4 million people in the region depending on this activity for their livelihoods (Jagun et al., 2022). However, studies in high-income countries have also documented significant occupational hazards among formal waste workers, including exposure to bioaerosols, infectious agents, and hazardous materials, leading to higher rates of occupational injuries and illnesses compared to other industrial sectors. (Eriksen et al., 2023; Le et al., 2023; Salambanga et al., 2022; Sara et al., 2022; Tehrani et al., 2024). The lack of proper waste disposal systems contributes to environmental crises that affect ecosystems, human health, and urban infrastructure. The accumulation of waste pollutes land, air, and water, intensifying the risk of respiratory diseases and other health complications (Salambanga et al., 2022). In response, many countries have adopted policies to regulate waste generation and encourage sustainable waste management (Abubakar et al., 2022; Pathak et al., 2024). The Rio Declaration and the UN’s 2030 Agenda emphasize the importance of recycling and reuse in minimizing waste accumulation (United Nations Conference on Environment & Development. Agenda 21. Río de Janeiro, Brasil;, 1992. Disponible En: Https://Sustainabledevelopment.Un.Org/Outcomedocuments/ Agenda21, n.d.; Walsh et al., 2022). Despite these efforts, Latin America’s waste management sector still faces considerable challenges. Informal waste collection provides a source of income for many economically disadvantaged individuals, yet it remains unregulated and often unsafe. A study conducted in Peru found that informal waste pickers operate in hazardous environments with limited access to protective equipment, occupational health services, and basic labour rights (Jiménez-de-Aliaga et al., 2020). This highlights the urgent need for formalizing waste-picking activities and implementing policies that protect the health and wellbeing of these workers. Approximately 4 million people across Latin America engage in informal waste picking, working under precarious conditions without social protection. (Lopez-Yamunaqué & Iannacone, 2023). Recognizing their role in the recycling chain, the United Nations Environment Programme has referred to these workers as “invisible environmentalists”. However, their informal status exposes them to occupational hazards, unsafe conditions, and job insecurity. (Schenck et al., 2019). In Colombia, each person generates approximately 1 kg of urban solid waste per day, yet formal recycling programs recover only a fraction of this material. Although legal frameworks acknowledge the contributions of waste pickers and provide financial support, their working conditions remain precarious (Lopez-Yamunaqué & Iannacone, 2023). To address this issue, Colombia has implemented regulations such as Decrees 2676 of 2000 and 4741 of 2005, aligning national waste management practices with global agreements like the Basel Convention. Bogotá, in particular, has introduced waste management plans to address hazardous waste, yet the city generated over 100,000 tons of hazardous waste in 2008 (Moreno-Bergaño et al., n.d.). An estimated 50,000 families in Bogotá depend on informal waste picking for income, with more than 24,000 active waste pickers. These individuals face serious physical risks due to direct exposure to hazardous materials, lack of protective equipment, and social stigmatization (Coletto & Carbonai, 2023). Studies have consistently highlighted the occupational health risks faced by waste pickers. (Melaku & Tiruneh, 2020). Prolonged exposure to airborne contaminants, toxic chemicals, and decomposing organic matter is associated with chronic respiratory conditions, including persistent coughing, wheezing, and excessive phlegm production Research conducted in Ghana, Delhi, South Africa and Latin America has demonstrated a strong correlation between waste-picking activities and respiratory diseases, reinforcing the urgent need for health interventions and protective measures in this sector. (Darboe et al., 2015; Jiménez-de-Aliaga et al., 2020; Laskaris et al., 2024; Ray et al., 2005). Moreover, indoor air pollution in waste sorting and storage areas further compounds health risks. Pollutants such as volatile organic compounds (VOCs) and fine particulate matter (PM2.5) can aggravate respiratory conditions, especially in poorly ventilated spaces. (Kwarteng et al., 2022; Wikuats et al., 2020). While this study primarily focuses on outdoor waste collection, it is important to acknowledge the broader health risks posed by indoor air pollutants in similar work environments (Mannan & Al-Ghamdi, 2021; Niza et al., 2024). This research aims to analyze the relationship between biological risks and the incidence of acute respiratory infections among informal waste pickers in Bogotá and surrounding municipalities in the Sabana Centro Province of Cundinamarca. By evaluating the demographic, occupational, and environmental factors contributing to respiratory symptoms, this study seeks to inform occupational health policies and propose interventions to protect the respiratory health of this vulnerable workforce. Materials and Methods This study used a descriptive cross-sectional design to examine respiratory symptoms and related factors among informal waste pickers in Bogotá, Tocancipá, Cajicá, and Funza (Colombia). The goal was to determine the prevalence of respiratory symptoms in this vulnerable group, considering various sociodemographic and occupational factors. Participants were selected via non-probabilistic convenience sampling. Initially, 335 informal waste pickers aged 18–80 with at least three months of recycling experience were identified. Ultimately, 179 met the inclusion criteria. Exclusions were due to refusal or unavailability. All participants were fully informed and gave written consent, and the Ethics Committee of UNIMINUTO approved the study. Paper-based questionnaires were used, focusing on sociodemographic data and respiratory symptoms using the ATS-DLD-78-A1 questionnaire. (Ferris, 1978). This tool 2Hispanic Health Care International assessed age, sex, medical history, lifestyle habits, and symptoms like cough and wheezing. Additional questions on personal protective equipment (PPE) use were also included (Aguilar-Elena et al., 2016). Data were analysed using R software, with descriptive statistics summarizing key variables. Custom functions generated summary tables, with visualizations like bar charts and histograms(Comtois, 2020) and relevant packages such as tidyverse(Wickham et al., 2019), caret (Kuhnn et al., 2023), randomForest (Breiman et al., 2002), and factoextra (Kassambara & Mundt, 2020). Missing data were excluded, and a binary variable was created from the RESPIRATORY_SYMPTOMS field. Categorical variables were converted into factors for regression models, and the dataset was split for training (66%) and testing (33%). We employed multiple statistical techniques to explore factors associated with respiratory symptoms. Logistic regression and linear regression were used to assess the relationships between sociodemographic and occupational variables and respiratory symptoms. Principal Component Analysis (PCA) was conducted to reduce dimensionality and identify key patterns in the data. Random Forest models provided insights into variable importance, highlighting key predictors of respiratory symptoms. K-means clustering was applied to classify participants based on their characteristics and symptoms. Finally, a neural network analysis was performed but faced challenges in identifying true positive cases due to class imbalance. To address this issue, different resampling techniques were explored, including oversampling of the minority class, under sampling of the majority class, and the application of synthetic data generation techniques such as SMOTE (Synthetic Minority Over-sampling Technique). Despite these adjustments, the model still exhibited poor generalization, with a tendency to misclassify positive cases due to the highly skewed class distribution. Consequently, given the limited predictive reliability and increased model complexity, the neural network analysis was excluded from the conclusions. These combined methods offered a comprehensive approach to understanding respiratory health issues in Colombian informal waste pickers. Results Descriptive Statistics The study involved (n =179) informal waste pickers from Bogotá, Tocancipá, Cajicá, and Funza. The descriptive statistics for the key variables are summarised in Table 1. The age of participants ranged from 18 to 80 years, with a mean age of 41.92 years (SD =15.36). The demographic breakdown revealed a varied representation across several categories. Notably, the majority of participants identified as male (approximately 65%), while the female population constituted about 35%. The educational levels varied, with a significant portion of respondents having completed primary education, and a smaller fraction achieving secondary or higher education. Correlation Analysis A correlation analysis was conducted to examine the relationships between key variables and respiratory symptoms (RESPIRATORY_SYMPTOMS). This analysis helped identify which symptoms and demographic factors were most strongly associated with respiratory health issues among informal waste pickers. The results highlighted several important correlations: •COUGH exhibited the strongest positive correlation with RESPIRATORY_SYMPTOMS (r =0.737), indicating that individuals experiencing respiratory symptoms were highly likely to report coughing as well. •Other respiratory symptoms such as COLD (r =0.429), WHEEZING (r =0.354), and PHLEGM (r =0.253) also showed moderate positive correlations, suggesting a strong interrelation among these symptoms. •Demographic factors displayed weaker correlations with respiratory symptoms: AGE (r =0.082) showed a low positive correlation, while GENDER (r =-0.008) indicated a near-zero relationship. •Additional weak correlations were found with RACE (r =0.092) and MEDICARE (r =0.102), suggesting a limited influence of these factors on respiratory symptoms within this sample. Furthermore, a Variance Inflation Factor (VIF) analysis was conducted to assess multicollinearity among variables. Results showed that multicollinearity was not a significant concern, as most VIF values were below 2. The exception was COUGH, which had a VIF of 1.87, indicating some potential for collinearity but within acceptable limits, ensuring that the regression analysis was not substantially affected. (Figure 1. Correlation Matrix for Respiratory Symptoms) Regression Analysis A logistic regression analysis was conducted to examine the association between various factors and the presence of respiratory symptoms (RESPIRATORY_SYMPTOMS) among informal waste workers (Table 2. Logistic Regression Results Predicting Respiratory Symptoms). Initially, all relevant variables were included; however, convergence issues due to multicollinearity were identified. A Variance Inflation Factor (VIF) analysis revealed significant collinearity with EDUCATION_LEVEL (VIF > 5), leading to a refined model retaining variables such as AGE, GENDER, SHORTNESS_BREATH, PNEUMONIA, and BRONCHITIS. In this reduced model, the intercept was statistically significant (p < 0.01), while PNEUMONIA showed a marginal association with respiratory symptoms (p =0.065). Other variables did not reach statistical significance (p > 0.05), indicating weaker associations with respiratory symptoms. To further refine the model, Lasso regression was applied, reducing model error and identifying COUGH as the most significant predictor. Lasso regression penalizes less relevant variables Rozo Silva et al. 3 Table 1. Descriptive Statistics Summarizing. Variable Stats / Values Freqs Graph AGE Media (SD): 41.92 (16.08) – Mediana: 41 Min–Max: 18–80 IQR: 26.5 MEDICARE 1. Contributory Social Security 30 (16.8%) 2. No opinion 50 (27.9%) 3. Subsidiary 99 (55.3%) GENDER 1. Female 71 (39.7%) 2. Male 108 (60.3%) MARITAL STATUS 1. Cohabitation 74 (41.3%) 2. Married 25 (14%) 3. Separated/Divorced 12 (6.7%) 4. Single 61 (34.1%) 5. Widowed 7 (3.9%) RACE 1. Asian 1 (0.6%) 2. Black 12 (6.7%) 3. Other 109 (60.9%) 4. White 57 (31.8%) EDUCATION LEVEL 1. Complete Secundary Education 46 (25.7%) 2. Completed Primary Education 22 (12.3%) 3. Incomplete Primary Education 51 (28.5%) 4. Incomplete Secundary Education 41 (22.9%) 5. No formal education 10 (5.6%) 6. Vocational Education and Training 9 (5%) RESPIRATORY SYMPTOMS 1. No 141 (78.8%) 2. Yes 38 (21.2%) COUGH 1. No 149 (83.2%) 2. Yes 30 (16.8%) PHLEGM 1. No 146 (81.6%) 2. Yes 33 (18.4%) WHEEZING 1. No 146 (81.6%) 2. Yes 33 (18.4%) (continued) 4Hispanic Health Care International Table 1. (Continued) Variable Stats / Values Freqs Graph SHORTNESS BREATH 1. No 175 (97.8%) 2. Yes 4 (2.2%) COLD 1. No 136 (77.7%) 2. Yes 39 (22.3%) PNEUMONIA 1. No 167 (93.8%) 2. Yes 11 (6.2%) RHINITIS 1. No 166 (93.3%) 2. Yes 12 (6.7%) BRONCHITIS 1. No 177 (99.4%) 2. Yes 1 (0.6%) ASTHMA 1. No 166 (93.3%) 2. Yes 12 (6.7%) SMOKER 1. No 95 (53.7%) 2. Yes 82 (46.3%) GLOVES USE 1. No 28 (15.6%) 2. Yes 151 (84.4%) SAFETY GOOGLES USE 1. No 151 (84.4%) 2. Yes 28 (15.6%) (continued) Rozo Silva et al. 5 by shrinking their coefficients toward zero, effectively eliminating those with minimal impact on the predictive performance of the model. This process enhances model interpretability while mitigating overfitting. The criteria for variable inclusion were based on statistical significance and contribution to the overall predictive accuracy, ensuring that only the most relevant factors were retained. The final model demonstrated high accuracy, with a minimal error rate (0.001024) and a classification accuracy of 100%, correctly identifying all 48 negative and 10 positive cases. The model’s predictive performance was further validated using a Receiver Operating Characteristic (ROC) curve, which demonstrated high discriminative ability. The ROC curve indicated an excellent balance between Table 1. (Continued) Variable Stats / Values Freqs Graph FACE MASK USE 1. No 9 (5%) 2. Yes 170 (95%) Figure 1. Correlation matrix about respiratory symptoms. 6Hispanic Health Care International sensitivity and specificity across probability thresholds, confirming the robustness of the model in predicting respiratory symptoms. (Figure 2. ROC Curve of the Logistic Regression Model) PCA Results The Principal Component Analysis (PCA) was selected to reduce the dimensionality of the dataset while preserving the most relevant information regarding respiratory symptoms and occupational health factors among informal waste pickers. Given the complexity of the dataset, which included multiple correlated variables, PCA allowed for the transformation of the data into a lower-dimensional space, facilitating the identification of key patterns and relationships. A Scree Plot analysis revealed an evident elbow point, indicating that the first three principal components (PCs) captured a significant portion of the variance in the dataset. Specifically, PC1 explained 18.4% of the variance, PC2 explained 11.8%, and the first three PCs together accounted for 40.3% of the total variance. Although this proportion is moderate, it reflects the most relevant sources of variability in the dataset. However, since 40.3% of the variance remains relatively low, this suggests that a substantial portion of the variance remains unexplained, warranting further interpretation or exploration of additional components. (Figure 3. Scree Plot for Principal Components) •PC1 (Respiratory Health Dimension): This component was primarily driven by respiratory symptoms, with high loadings for COUGH (0.82), RESPIRATORY_SYMPTOMS (0.78), and WHEEZING (0.71). This suggests that PC1 effectively captures the underlying structure of respiratory health factors. •PC2 (Demographic and Lifestyle Factors): Associated mainly with AGE (0.76) and SMOKER (0.65), indicating that this component reflects demographic influences on respiratory health. •PC3 (Occupational Protective Measures): Contributed by variables such as GLOVES_USE, which played a role in differentiating individuals based on protective behaviours. The biplot representation of the first two PCs illustrated strong clustering among variables like COUGH, RESPIRATORY_SYMPTOMS, and WHEEZING, confirming their interdependence. Additionally, AGE and SMOKER were closely aligned with PC2, reinforcing the impact of aging and smoking behaviour on respiratory health. (Figure 4. PCA Biplot) Random Forest Analysis A Random Forest model was constructed with 500 trees to explore predictors of respiratory symptoms. After preprocessing the data and handling missing values, the model achieved optimal performance on the training set. The confusion matrix showed an out-of-bag error rate of 0%, indicating perfect classification accuracy. The Mean Decrease Gini measure was used to assess variable importance within the model. The most influential predictors of respiratory symptoms were: •COUGH (Mean Decrease Gini =18.73): The most significant predictor, aligning with its clinical relevance. •AGE (Mean Decrease Gini =7.74), RACE (Mean Decrease Gini =1.51), and EDUCATION_LEVEL (Mean Decrease Gini =1.22): These demographic variables also contributed significantly to the predictive performance, highlighting the role of personal characteristics in respiratory health outcomes. The model’s performance metrics were: •Precision =0.6364 •Recall =0.7000 •F1-Score =0.6667 These results confirm that COUGH was the most dominant factor, while demographic factors played a secondary role in predicting respiratory symptoms. (Figure 5. Variable Importance in Prediction Using Mean Decrease in Gini Index in a Random Forest Model) K-means Clustering Results A K-means clustering analysis was conducted to identify patterns related to respiratory symptoms among informal waste pickers (Figure 6. Cluster Plot). The analysis yielded three distinct clusters, revealing important differences in symptom prevalence and demographic characteristics: •Cluster 1: Consisted of 12 participants with an average age of 36.08 years. The average respiratory symptom score was 0.17, with notable occurrences of cough and bronchitis. Table 2. Logistic Regression Results Predicting Respiratory Symptoms. Term Estimate Std. Error Statistic p-value (Intercept) −8.069 3.065 −2.6323625 0.008 GENDER 1.076 0.781 1.3783943 0.168 MEDICARE −0.213 0.516 −0.4127010 0.680 MARITAL STATUS −0.362 0.254 −1.4268026 0.154 RACE 1.219 0.580 2.1002236 0.036 EDUCATION LEVEL −0.370 0.306 −1.2064285 0.228 COUGH 5.634 1.005 5.6075135 0.000 PHLEGM 1.708 0.874 1.9542893 0.051 WHEEZING 1.041 0.894 1.1639656 0.244 COLD 2.214 0.916 2.4178635 0.016 SMOKER 0.673 0.745 0.9034888 0.366 GLOVES_USE 0.321 1.042 0.3082911 0.758 SAFETY GOGGLES USE −0.360 0.891 −0.4040737 0.686 Rozo Silva et al. 7 This cluster represents a younger group with mild symptoms, likely due to exposure to respiratory irritants but with less cumulative impact. •Cluster 2: The largest cluster, including 133 participants, had an average age of 41.63 years. The average respiratory symptom score was only 0.03, indicating minimal symptoms. The low symptom prevalence in this group may reflect effective preventive measures or lower occupational exposure. •Cluster 3: Comprised of 29 participants, this cluster had the highest average age (45.97 years) and the highest respiratory symptom score (1.00). This group exhibited severe Figure 2. ROC curve of the logistic regression model for predicting respiratory symptoms in workers exposed to biological hazards. Figure 3. Scree Plot for principal components. 8Hispanic Health Care International symptoms, including cough, phlegm, and pneumonia, suggesting that age and cumulative exposure contribute to worsening respiratory health. The clustering results emphasize that Cluster 3 is the most vulnerable group, displaying higher respiratory symptoms, likely due to long-term occupational exposure. Cluster 1, Figure 4. Principal Component Analysis (PCA) biplots showing relationships between variables and participants across the first two principal components (PC1 and PC2). Figure 5. Variable importance in prediction using mean decrease in Gini Index in a Random Forest model. Rozo Silva et al. 9