A Machine Learning Approach to Prediction and Multimorbidity Risk Factor Identification in a low- and middle-income country

A Machine Learning Approach to Prediction and Multimorbidity Risk Factor Identification in a low- and middle-income country | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search A Machine Learning Approach to Prediction and Multimorbidity Risk Factor Identification in a low- and middle-income country Olalekan A. Uthman , Matthew Hazell , Muhammed Mubashir Babatunde Uthman , Kolawole W Wahab , Ponnusamy Saravanan , Paramjit Gill , Andre Pascal Kengne doi: https://doi.org/10.1101/2025.10.13.25337900 Olalekan A. Uthman 1 Warwick Applied Health, Warwick Centre for Global Health, Warwick Medical School, University of Warwick , Coventry, CV4 7AL, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: Olalekan.Uthman{at}warwick.ac.uk Matthew Hazell 2 Warwick Applied Health, Warwick Medical School, University of Warwick , Coventry, CV4 7AL, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Muhammed Mubashir Babatunde Uthman 3 Department of Epidemiology and Community Health. College of Health Sciences, University of Ilorin, University of Ilorin Teaching Hospital. Ilorin , Kwara State, Nigeria Find this author on Google Scholar Find this author on PubMed Search for this author on this site Kolawole W Wahab 4 Department of Medicine, University of Ilorin , Ilorin 240001, Nigeria Find this author on Google Scholar Find this author on PubMed Search for this author on this site Ponnusamy Saravanan 1 Warwick Applied Health, Warwick Centre for Global Health, Warwick Medical School, University of Warwick , Coventry, CV4 7AL, United Kingdom 2 Warwick Applied Health, Warwick Medical School, University of Warwick , Coventry, CV4 7AL, United Kingdom 5 Department of Diabetes, Endocrinology and Metabolism, George Eliot Hospital , Nuneaton, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site Paramjit Gill 1 Warwick Applied Health, Warwick Centre for Global Health, Warwick Medical School, University of Warwick , Coventry, CV4 7AL, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Andre Pascal Kengne 6 Non-Communicable Diseases Research Unit, South African Medical Research Council , Francie Van Zijl Dr, Parow Valley, Cape Town, 7501, South Africa 7 Department of Biological and Environmental Sciences, Faculty of Natural Sciences, Walter Sisulu University , Mthatha, South Africa Find this author on Google Scholar Find this author on PubMed Search for this author on this site Abstract Full Text Info/History Metrics Data/Code Preview PDF Abstract Importance Multimorbidity, the coexistence of multiple chronic conditions, is a growing public health challenge, particularly in low- and middle-income countries like South Africa. Identifying individuals at high risk of multimorbidity is crucial for developing targeted interventions and allocating healthcare resources effectively. Objective To investigate the predictive performance of various machine learning models in identifying individuals at risk of multimorbidity in South Africa and to identify the most influential predictors of multimorbidity, considering both individual-level and contextual factors. Design, Setting, and Participants This cross-sectional study utilized data from the South Africa Demographic and Health Survey (SADHS) 2016, a nationally representative household survey. The study included 5,342 participants aged 18 years and older, of which 2,107 (33.9%) had multimorbidity, defined as the presence of two or more chronic conditions. Main Outcomes and Measures The primary outcome was the presence of multimorbidity. Machine learning models, including gradient boosting classifier, linear discriminant analysis, ada boost classifier, logistic regression, ridge classifier, catboost classifier, random forest classifier, light gradient boosting machine, extra trees classifier, naive bayes, quadratic discriminant analysis, extreme gradient boosting, k neighbors classifier, dummy classifier, decision tree classifier, svm - linear kernel, were developed and evaluated using a repeated train-test split approach. Model performance was assessed using accuracy, area under the receiver operating characteristic curve (AUC), recall, precision, F1 score, Cohen’s Kappa, and Matthews Correlation Coefficient (MCC). Shapley Additive Explanations (SHAP) were used to identify the most influential predictors of multimorbidity. Results The Gradient Boosting Classifier achieved the highest predictive performance, with an AUC of 0.7809, accuracy of 0.7478, and F1 score of 0.5798. Age, no medication use, sex, poor health perception, and community illiteracy rate were identified as the most influential predictors of multimorbidity. Individual-level factors had a more substantial impact on the likelihood of multimorbidity compared to community-level factors. However, higher community illiteracy rates and regional unemployment rates were associated with an increased likelihood of multimorbidity, highlighting the importance of contextual factors. The fairness and demographic bias assessment revealed that the Gradient Boosting Classifier maintained a high level of fairness across different regions, wealth index categories, age groups, and genders. Conclusion and Relevance Machine learning algorithms, particularly the Gradient Boosting Classifier, can accurately predict multimorbidity in the South African context. The findings emphasize the importance of considering both individual-level and contextual factors in understanding the determinants of multimorbidity. Background Multimorbidity, defined as the co-occurrence of multiple long-term conditions within an individual, has emerged as a significant global health challenge, disproportionately affecting low- and middle-income countries (LMICs) ( Afshar et al., 2015 ). In South Africa, a country grappling with a double burden of communicable and non-communicable diseases, the prevalence of multimorbidity is on the rise ( Weimann et al., 2016 ). The complex interplay of socioeconomic determinants, including neighbourhood-level factors, has been recognized as a crucial driver of this growing burden ( Hurst et al., 2015 ). The impact of multimorbidity on individuals, healthcare systems, and societies in LMICs is profound. Patients with multiple chronic conditions often experience decreased quality of life, functional limitations, and increased healthcare utilization ( Arokiasamy et al., 2015 ). Healthcare systems in LMICs, already strained by resource constraints, face the challenge of providing comprehensive and coordinated care for individuals with complex health needs ( Oni et al., 2015 ). Moreover, the economic burden of multimorbidity, both in terms of direct healthcare costs and indirect costs such as lost productivity, poses a significant threat to the socioeconomic development of these nations ( Wang et al., 2014 ). Recognising the urgent need to address the growing burden of multimorbidity in LMICs, researchers have sought to identify the underlying determinants and predictors of this phenomenon. While individual-level factors such as age, gender, and lifestyle behaviours have been extensively studied, the role of neighbourhood-level socioeconomic factors in shaping the risk of multimorbidity has garnered increasing attention ( Alaba & Chola, 2013 ). Neighbourhood socioeconomic status (SES), encompassing factors such as income levels, educational attainment, and employment rates, has been linked to the development and progression of chronic diseases ( Weimann et al., 2016 ). Recent advancements in machine learning have opened up new avenues for exploring the complex relationships between neighbourhood SES and multimorbidity in LMICs ( Panch et al., 2018 ). By leveraging large datasets and advanced analytical techniques, machine learning models can uncover intricate patterns and predictors that traditional statistical methods may overlook ( Prosperi et al., 2020 ). This approach holds immense potential for identifying neighbourhood-level socioeconomic determinants of multimorbidity, enabling the development of targeted interventions and policies to mitigate the burden of this condition in resource-limited settings. Although the influence of neighbourhood socioeconomic status (SES) on health outcomes is increasingly acknowledged, to the best of our knowledge, machine learning approaches have not been employed to examine its predictive utility for multimorbidity within low- and middle-income country (LMIC) settings such as South Africa. This study aims to address this gap by harnessing the power of machine learning to elucidate the relationship between neighbourhood socioeconomic indicators and multimorbidity. Methods Study Design and Data Sources This study employed a cross-sectional design using data from the South Africa Demographic and Health Survey (SADHS) 2016. The SADHS 2016 is a nationally representative household survey that provides comprehensive information on various demographic and health indicators ( NDoH et al., 2019 ). The DHS programme has assisted in conducting over 350 nationally representative household surveys across 90 countries since 1984, making it an essential data source for policy-making, monitoring, and evaluation in many LMICs ( Corsi et al., 2012 ; USAID, 2023). The SADHS 2016 utilized a stratified two-stage sample design ( NDoH et al., 2019 ). In the first stage, 750 primary sampling units (PSUs) were selected from 26 sampling strata based on urban, traditional, and rural areas within each of the nine South African provinces. In the second stage, a fixed number of 20 dwelling units (DUs) were randomly selected from each PSU. This design allows for the estimation of key variables at the national, provincial, and area (urban, rural, traditional) levels. Data collection took place from 27 June 2016 to 4 November 2016. All DUs were eligible for the primary modules on women, fertility, and children, while half were subsampled for modules on men and adult health. The adult health module included self-reported chronic conditions and biomarker collection for anthropometry, anaemia, hypertension, HbA1c levels for diabetes, and HIV for participants aged over 15. Study Population Men and women aged 18 and over were included in this study if they contributed data to the SADHS 2016 adult health modules. Individuals under 18 were excluded to ensure comparability with the literature. Participants were also excluded if they were missing information on the multimorbidity outcome. Variables Outcome Variable Multimorbidity was defined as the presence of two or more co-existing chronic conditions ( Johnston et al., 2019 ). Twelve current chronic diseases were considered: tuberculosis, hypertension, stroke, high blood cholesterol, anaemia, chronic bronchitis, diabetes, asthma, cancer, heart disease, human immunodeficiency virus (HIV), and chronic pain. These conditions were recorded through self-reports and biomarker measurements. Participants with missing information on a disease (self-report or biomarker) were coded as missing rather than not having the disease to reduce misclassification bias. Explanatory Variables Explanatory variables were selected based on literature review and a-priori reasoning. Socioeconomic variables included household wealth, education level, occupational status, health insurance, and marital status. Individual-level health variables included BMI, dietary health, sugary drink intake, smoking status, alcohol consumption, and occupational smoke exposure. Access to old and new media were included as proxies for health information access. Age, sex, and ethnicity were also included. Further details on variable derivation and coding are provided in Table S1 (Supplementary file). Neighbourhoods were defined as clusters of households serving as PSUs within the DHS. Neighbourhood and province-level explanatory variables included poverty, rurality, unemployment level, and illiteracy (community-level only). These were categorized as low, medium, or high based on the proportion of individuals in the most deprived group for each neighbourhood and province, calculated using the larger adult-health sample (N=10,336) to include more contextual information. Data Preprocessing and Analysis Data preprocessing involved cleaning, recoding variables, and handling missing data as described in the Variables section. Participants with missing information on the multimorbidity outcome were excluded from the analysis. Descriptive statistics, including means, standard deviations, and proportions, were calculated for all variables, stratified by multimorbidity status. These summary statistics provided an overview of the characteristics of the study population and potential differences between those with and without multimorbidity. Machine Learning Model Development A repeated train-test split approach was employed for model development and evaluation. The dataset was randomly divided into training (80%) and test (20%) sets, ensuring that data points from the same patient were exclusively included in either the training or test set to prevent data leakage. Various machine learning algorithms, including Gradient Boosting Classifier, Linear Discriminant Analysis, AdaBoost Classifier, Logistic Regression, Ridge Classifier, CatBoost Classifier, Random Forest Classifier, Light Gradient Boosting Machine, Extra Trees Classifier, Naive Bayes, Quadratic Discriminant Analysis, Extreme Gradient Boosting, K-Nearest Neighbors Classifier, Dummy Classifier, Decision Tree Classifier, and Support Vector Machine with Linear Kernel, were applied. Hyperparameters for each model were optimized using random search with 10-fold cross-validation on the training set. The process of splitting, hyperparameter optimization, model training, and evaluation was repeated 1,000 times to ensure the robustness of the results and to account for potential variations due to random splitting. The performance metrics, including accuracy, AUC, recall, precision, F1 score, Cohen’s Kappa, and Matthews Correlation Coefficient (MCC), were calculated and averaged across all iterations. Model Interpretation To gain insights into the factors contributing to the prediction of multimorbidity, Shapley Additive Explanations (SHAP) were employed. SHAP values were calculated for the best-performing model, which was the Gradient Boosting Classifier with the highest AUC on the test set across all iterations. The SHAP values provided a unified measure of feature importance by quantifying the contribution of each predictor variable to the model output. The SHAP summary plot displayed the average impact of each feature on the model output, with positive values indicating an increase in the likelihood of multimorbidity and negative values indicating a decrease. Additionally, SHAP dependence plots were generated to visualize the impact of individual-level and community-level features on the model output. These plots showed the relationship between the feature values and the SHAP values, providing a more detailed understanding of how each feature influenced the prediction of multimorbidity. The interpretation of these plots offered insights into the complex interplay between individual and contextual factors in shaping the risk of multiple chronic conditions. The SHAP analysis was conducted using Python (version 3.10.4) and the SHAP library. Evaluation The performance of the machine learning models was evaluated using various metrics, including accuracy, AUC, recall, precision, F1 score, Cohen’s Kappa, and MCC. The primary evaluation metric was the AUC, which measures the model’s ability to discriminate between individuals with and without multimorbidity. The AUC ranges from 0 to 1, with a value of 0.5 indicating a random classifier and a value of 1 indicating a perfect classifier. The other metrics provided additional insights into the model’s performance, such as the balance between precision and recall (F1 score), the agreement between predicted and actual labels (Cohen’s Kappa), and the correlation between predictions and observations (MCC). Fairness and Demographic Bias Assessment To ensure the machine learning model’s effectiveness and equitable application across diverse populations, a fairness and demographic bias assessment was integrated into the study. During the model development phase, key demographic factors, such as age, gender, region, and wealth index, were considered to capture the wide-ranging realities and experiences of the patient population. The model’s performance was evaluated across these demographic groups using various metrics, including accuracy, precision, recall, F1 score, Selection Rate, Cohen’s Kappa, and Matthews Correlation Coefficient (MCC). Results Descriptive Statistics As presented in Table 1 , the study included a total of 5,342 participants, of which 2382 (44.6%) were classified as having multimorbidity, defined as the presence of two or more chronic conditions. The mean age of participants with multimorbidity was nearly 14 years higher than those without multimorbidity. Access to old and new media, health insurance coverage, and poverty status were found to be similar between the two groups. View this table: View inline View popup Table 1: Summary characteristics of included participants Characteristics of the study population by multimorbidity status. Model Performance Comparison The performance metrics of the various machine learning models employed are summarized in Table 2 . The Gradient Boosting Classifier demonstrated the highest accuracy (0.7478), AUC (0.7809), and F1 score (0.5798) among all the models, indicating its superior ability to predict multimorbidity in this population. Linear Discriminant Analysis and AdaBoost Classifier also exhibited strong performance, with AUC scores of 0.7777 and 0.7775, respectively. View this table: View inline View popup Table 2: Model predictions comparison The Dummy Classifier, which served as a baseline model, achieved an accuracy of 0.6609 and an AUC of 0.5, confirming that all the employed models performed substantially better than random guessing. This finding underscores the value of using machine learning techniques to identify individuals at risk of multimorbidity based on their sociodemographic and health-related characteristics. The confusion matrix ( Figure 1 ) provides a more detailed breakdown of the Gradient Boosting Classifier’s performance. The model correctly identified 55.3% of the true positives (individuals with multimorbidity) and 89.2% of the true negatives (individuals without multimorbidity), demonstrating its ability to accurately classify a significant proportion of the study participants. However, the model misclassified 10.8% of the individuals with multimorbidity as not having the condition (false negatives) and 44.7% of the individuals without multimorbidity as having the condition (false positives), highlighting the need for further refinement of the model to improve its predictive accuracy. Download figure Open in new tab Figure 1: Summary metrics for best performing model Feature Importance The SHAP (Shapley Additive Explanations) values, depicted in Figure 2 , provide insights into the average impact of the predictor variables on the model output. Age emerged as the most influential factor, followed by no medication use, female gender, poor health perception, and community illiteracy rate. These findings suggest that individual-level factors, such as age, gender, and health status, exert a more substantial influence on the likelihood of multimorbidity compared to community-level factors, such as illiteracy and poverty rates. Download figure Open in new tab Figure 2: Variable importance Individual-Level Feature Impact A more granular analysis of the impact of individual-level features on the model output is provided in Figure 3 . Higher values of age, female gender, poor health perception, no medication use, and overweight/obesity status were found to be associated with an increased likelihood of multimorbidity. These findings are consistent with previous research highlighting the role of aging, gender disparities, and modifiable risk factors, such as obesity and medication non-adherence, in the development of multiple chronic conditions. Conversely, younger age, being a male, no medication, good health perception were associated with a decreased likelihood of multimorbidity. Download figure Open in new tab Figure 3: Shapley Addictive explanation (SHAP) force plot for two selected patients Fairness and Demographic Bias Assessment As shown in Figure 4A , the assessment of the model’s performance across different regions (Eastern Cape, Free State, Gauteng, KwaZulu-Natal, Limpopo, Mpumalanga, North West, Northern Cape, and Western Cape) revealed consistent results for all metrics, suggesting that the model is fair and unbiased with respect to the region feature. Similarly, the model maintained high levels of fairness across different wealth index categories (poorest to richest) for most metrics, with only slight variations in the Kappa and MCC metrics, which were marginally higher for the richest category. Download figure Open in new tab Figure 4: Fairness and Demographic Bias Assessment results Figure 4B compares the model’s performance across different wealth index categories, ranging from poorest to richest. The samples, precision, accuracy, recall, F1, and Selection Rate metrics are nearly identical across all wealth index categories, indicating a high level of fairness. However, there are slight variations in the Kappa metric, with the richest category having a marginally higher value compared to the other categories. This suggests that the model has a slightly better class agreement for the richest wealth index category. The MCC metric is also marginally higher for the richest category, indicating a slightly better overall performance. While these differences are small, they should be taken into account when interpreting the model’s results for each wealth index category. Further analysis could help identify the underlying reasons for these variations and ensure complete fairness across different socioeconomic groups. Figure 4C analyses the model’s performance across four age categories: Adult, Middle-Aged, Older, and Young Adult. The model maintains high levels of fairness across all age groups for the samples, precision, accuracy, F1, and Selection Rate metrics. However, there are some variations in the Kappa, recall, and MCC metrics. The Young Adult and Adult categories have slightly higher Kappa and recall values compared to the Middle-Aged and Older categories. This indicates that the model might be slightly better at correctly identifying positive cases for younger age groups. The differences in MCC suggest that the model’s overall performance is somewhat better for younger age categories. While these variations are not drastic, they should be considered when applying the model to different age groups. Figure 4D compares the model’s performance between the female and male genders. The samples, precision, accuracy, recall, F1, and Selection Rate metrics are nearly identical for both genders, indicating a high level of fairness. However, the Kappa metric is slightly higher for females, suggesting that the model has a better class agreement for the female gender. The MCC metric is also marginally higher for females, indicating a slightly better overall performance. While these differences are small, they should be taken into account when interpreting the model’s results for each gender. Further analysis could help identify the underlying reasons for these variations and ensure complete gender fairness. Discussion Main Findings This study aimed to investigate the predictive performance of various machine learning models in identifying individuals at risk of multimorbidity in South Africa. The results revealed that the Gradient Boosting Classifier achieved the highest predictive performance, with an AUC of 0.7809, accuracy of 0.7478, and F1 score of 0.5798. The SHAP analysis identified age, no medication use, female gender, poor health perception, and community illiteracy rate as the most influential predictors of multimorbidity. Individual-level factors were found to have a more substantial impact on the likelihood of multimorbidity compared to community-level factors. However, the study also highlighted the importance of considering contextual factors, as higher community illiteracy rates and regional unemployment rates were associated with an increased likelihood of multimorbidity. The fairness and demographic bias assessment revealed that the Gradient Boosting Classifier maintained a high level of fairness across different regions, wealth index categories, age groups, and genders. However, slight variations were observed in certain metrics for specific demographic categories. The model performed consistently across all regions, while the Kappa and MCC metrics were marginally higher for the richest wealth index category. The Young Adult and Adult age categories had slightly higher Kappa and recall values compared to the Middle-Aged and Older categories, and the Kappa and MCC metrics were slightly higher for females. These findings suggest that an accurate classification algorithm, such as the Gradient Boosting Classifier, could serve as a clinically useful tool to advance the prevention and control of multimorbidity by empowering healthcare providers to prescribe preventative measures to at-risk patients and take earlier action to order diagnostic tests ( Alonso-Morán et al., 2015 ). By identifying individuals at high risk of developing multiple chronic conditions, healthcare systems can allocate resources more effectively and implement targeted interventions to improve health outcomes and reduce the burden of multimorbidity ( Arokiasamy et al., 2015 ). The study findings underscore the importance of considering both individual-level factors, such as age, gender, and health status, and community-level factors, such as illiteracy and unemployment rates, in understanding the determinants of multimorbidity in the South African context. Higher values of community illiteracy rate and regional unemployment rate were found to be associated with an increased likelihood of multimorbidity, underscoring the potential influence of socioeconomic factors on health outcomes at the community level. These findings highlight the need for interventions that address the social determinants of health, such as education and employment, to reduce the burden of multimorbidity in disadvantaged communities. Interestingly, higher values of age category and female gender were associated with a decreased likelihood of multimorbidity at the community level, in contrast to the individual-level findings. This observation may reflect the complex interplay between individual and contextual factors in shaping the risk of multiple chronic conditions and warrants further investigation to unravel the underlying mechanisms and potential effect modifiers. These insights can inform the development of targeted interventions and policies aimed at preventing and managing multimorbidity, with a focus on addressing modifiable risk factors at the individual level and social determinants of health at the community level. Further research is needed to validate these findings in other populations and settings, and to explore the potential of integrating machine learning techniques into clinical decision support systems to improve the early detection and management of multimorbidity. Comparison with Previous Studies The findings of this study are consistent with previous research highlighting the role of individual-level factors, such as age, gender, and health status, in the development of multimorbidity ( Barnett et al., 2012 ; Marengoni et al., 2011 ). The higher prevalence of multimorbidity among females observed in this study aligns with the findings of a systematic review by Violan et al. (2014) , which reported a consistent association between female gender and multimorbidity across various settings. The identification of poor health perception as a significant predictor of multimorbidity is supported by a study by Mavaddat et al. (2014) , which found that self-rated health was strongly associated with the presence of multiple chronic conditions. The importance of contextual factors, such as community illiteracy rates and regional unemployment rates, in predicting multimorbidity is consistent with the growing body of evidence on the social determinants of health ( Marmot et al., 2008 ). A study by Nielsen et al. (2019) found that living in socioeconomically deprived areas was associated with an increased risk of multimorbidity, independent of individual-level factors. The findings of this study extend this line of research by demonstrating the potential of machine learning techniques to capture the complex interplay between individual and contextual factors in shaping the risk of multimorbidity. Implications for Practice and Future Research The results of this study have important implications for clinical practice and public health policy. The accurate identification of individuals at high risk of multimorbidity using machine learning algorithms can facilitate the development of targeted interventions and care coordination strategies to prevent the onset and progression of multiple chronic conditions ( Nicholson et al., 2019 ). Healthcare providers can use these tools to prioritize preventive care, such as lifestyle modifications and screening programs, for at-risk patients, potentially reducing the burden of multimorbidity and improving health outcomes ( Salisbury et al., 2018 ). Furthermore, the study highlights the need for a comprehensive approach to addressing the social determinants of health to reduce the burden of multimorbidity in disadvantaged communities. Public health policies and interventions should focus on improving access to education, employment opportunities, and healthcare services in these communities to mitigate the impact of socioeconomic disparities on health outcomes ( Marmot et al., 2008 ). Future research should explore the effectiveness of community-level interventions in reducing the prevalence of multimorbidity and investigate the potential of integrating machine learning algorithms into clinical decision support systems to optimize the prevention and management of multiple chronic conditions ( Nicholson et al., 2019 ). Study Strengths and Limitations One of the main strengths of this study is the use of a large, nationally representative dataset, which enhances the generalizability of the findings to the South African population. The application of various machine learning algorithms and the rigorous evaluation of their predictive performance using multiple metrics provide a comprehensive assessment of the potential utility of these tools in identifying individuals at risk of multimorbidity. The use of SHAP values for model interpretation offers a transparent and interpretable approach to understanding the contribution of individual predictors to the model output, facilitating the translation of these findings into clinical practice ( Lundberg & Lee, 2017 ). However, the study also has several limitations that should be considered when interpreting the results. First, the cross-sectional design of the study precludes the establishment of causal relationships between the predictor variables and multimorbidity. Future research should employ longitudinal designs to investigate the temporal dynamics of these associations and to validate the predictive performance of the machine learning models over time. Second, the study relied on self-reported data for some of the predictor variables, such as health perception and medication use, which may be subject to recall bias and social desirability bias. Future studies should incorporate objective measures of health status and medication adherence to improve the accuracy of the predictive models. Conclusion In conclusion, this study demonstrates the potential of machine learning algorithms, particularly the Gradient Boosting Classifier, in predicting multimorbidity in the South African context. The findings highlight the importance of considering both individual-level factors and contextual factors in understanding the determinants of multimorbidity. The accurate identification of individuals at high risk of developing multiple chronic conditions using these tools can inform the development of targeted interventions and care coordination strategies to prevent the onset and progression of multimorbidity, ultimately improving health outcomes and reducing the burden on healthcare systems. Future research should focus on validating these findings in other populations and settings, ideally in longitudinal studies, exploring the effectiveness of community-level interventions in reducing the prevalence of multimorbidity, and investigating the potential of integrating machine learning algorithms into clinical decision support systems to optimize the prevention and management of multiple chronic conditions, in randomised controlled trials. Data Availability All data produced in the present work are contained in the manuscript References 1. ↵ Alonso-Morán , E. , Nuño-Solinis , R. , Onder , G. , & Tonnara , G. ( 2015 ). Multimorbidity in risk stratification tools to predict negative outcomes in adult population . European Journal of Internal Medicine , 26 ( 3 ), 182 – 189 . OpenUrl PubMed 2. Arokiasamy , P. , Uttamacharya Jain , K. , Biritwum , R. B. , Yawson , A. E. , Wu , F. , … & Kowal , P. ( 2015 ). The impact of multimorbidity on adult physical and mental health in low-and middle-income countries: what does the study on global ageing and adult health (SAGE) reveal? . BMC Medicine , 13 ( 1 ), 1 – 16 . OpenUrl PubMed 3. ↵ Barnett , K. , Mercer , S. W. , Norbury , M. , Watt , G. , Wyke , S. , & Guthrie , B. ( 2012 ). Epidemiology of multimorbidity and implications for health care, research, and medical education: a cross-sectional study . The Lancet , 380 ( 9836 ), 37 – 43 . OpenUrl CrossRef 4. ↵ Lundberg , S. M. , & Lee , S. I. ( 2017 ). A unified approach to interpreting model predictions . Advances in Neural Information Processing Systems , 30 , 4765 – 4774 . OpenUrl 5. ↵ Marengoni , A. , Angleman , S. , Melis , R. , Mangialasche , F. , Karp , A. , Garmen , A. , … & Fratiglioni , L. ( 2011 ). Aging with multimorbidity: a systematic review of the literature . Ageing Research Reviews , 10 ( 4 ), 430 – 439 . OpenUrl CrossRef PubMed Web of Science 6. ↵ Marmot , M. , Friel , S. , Bell , R. , Houweling , T. A. , Taylor , S. , & Commission on Social Determinants of Health . ( 2008 ). Closing the gap in a generation: health equity through action on the social determinants of health . The Lancet , 372 ( 9650 ), 1661 – 1669 . OpenUrl CrossRef 7. ↵ Mavaddat , N. , Valderas , J. M. , van der Linde , R. , Khaw , K. T. , & Kinmonth , A. L. ( 2014 ). Association of self-rated health with multimorbidity, chronic disease and psychosocial factors in a large middle-aged and older cohort from general practice: a cross-sectional study . BMC Family Practice , 15 ( 1 ), 1 – 11 . OpenUrl PubMed 8. ↵ Nicholson , K. , Makovski , T. T. , Griffith , L. E. , Raina , P. , Stranges , S. , & van den Akker , M. ( 2019 ). Multimorbidity and comorbidity revisited: refining the concepts for international health research . Journal of Clinical Epidemiology , 105 , 142 – 146 . OpenUrl CrossRef PubMed 9. ↵ Nielsen , C. R. , Halling , A. , & Andersen-Ranberg , K. ( 2019 ). Disparities in multimorbidity across Europe: Findings from the SHARE Survey . European Geriatric Medicine , 10 ( 2 ), 289 – 300 . OpenUrl 10. ↵ Salisbury , C. , Man , M. S. , Bower , P. , Guthrie , B. , Chaplin , K. , Gaunt , D. M. , … & Mercer , S. W. ( 2018 ). Management of multimorbidity using a patient-centred care model: a pragmatic cluster-randomised trial of the 3D approach . The Lancet , 392 ( 10141 ), 41 – 50 . OpenUrl 11. ↵ Violan , C. , Foguet-Boreu , Q. , Flores-Mateo , G. , Salisbury , C. , Blom , J. , Freitag , M. , … & Valderas , J. M. ( 2014 ). Prevalence, determinants and patterns of multimorbidity in primary care: a systematic review of observational studies . PloS One , 9 ( 7 ), e102149 . OpenUrl CrossRef PubMed 12. ↵ Corsi , D. J. , Neuman , M. , Finlay , J. E. , & Subramanian , S. V. ( 2012 ). Demographic and health surveys: a profile . International Journal of Epidemiology , 41 ( 6 ), 1602 – 1613 . OpenUrl CrossRef PubMed Web of Science 13. ↵ Johnston , M. C. , Crilly , M. , Black , C. , Prescott , G. J. , & Mercer , S. W. ( 2019 ). Defining and measuring multimorbidity: a systematic review of systematic reviews . European Journal of Public Health , 29 ( 1 ), 182 – 189 . OpenUrl CrossRef PubMed 14. ↵ National Department of Health (NDoH), Statistics South Africa (Stats SA), South African Medical Research Council (SAMRC), & ICF . ( 2019 ). South Africa Demographic and Health Survey 2016 . Pretoria, South Africa, and Rockville, Maryland, USA: NDoH, Stats SA, SAMRC, and ICF . 15. Roomaney , R. A. , Chetty , I. , Pillay-van Wyk , V. , & Awotiwon , O. F. ( 2022 ). Prevalence of multimorbidity in South Africa: a systematic review protocol . BMJ Open , 12 ( 2 ), e052295 . OpenUrl Abstract / FREE Full Text 16. United States Agency for International Development (USAID ). ( 2023 ). The DHS Program - Quality information to plan, monitor and improve population, health, and nutrition programs . https://dhsprogram.com/ 17. ↵ Afshar , S. , Roderick , P. J. , Kowal , P. , Dimitrov , B. D. , & Hill , A. G. ( 2015 ). Multimorbidity and the inequalities of global ageing: a cross-sectional study of 28 countries using the World Health Surveys . BMC Public Health , 15 ( 1 ), 776 . OpenUrl CrossRef PubMed 18. ↵ Alaba , O. , & Chola , L. ( 2013 ). The social determinants of multimorbidity in South Africa . International Journal for Equity in Health , 12 ( 1 ), 63 . OpenUrl PubMed 19. ↵ Arokiasamy , P. , Uttamacharya , U. , Jain , K. , Biritwum , R. B. , Yawson , A. E. , Wu , F. , … & Kowal , P. ( 2015 ). The impact of multimorbidity on adult physical and mental health in low-and middle-income countries: what does the study on global ageing and adult health (SAGE) reveal? . BMC Medicine , 13 ( 1 ), 178 . OpenUrl PubMed 20. ↵ Hurst , J. R. , Dickhaus , J. , Maulik , P. K. , Miranda , J. J. , Pastakia , S. D. , Soriano , J. B. , … & Siddharthan , T. ( 2015 ). Global Alliance for Chronic Disease researchers’ statement on multimorbidity . The Lancet Global Health , 3 ( 2 ), e98 – e100 . OpenUrl 21. ↵ Oni , T. , McGrath , N. , BeLue , R. , Roderick , P. , Colagiuri , S. , May , C. R. , & Levitt , N. S. ( 2015 ). Chronic diseases and multi-morbidity--a conceptual modification to the WHO ICCC model for countries in health transition . BMC Public Health , 14 ( 1 ), 575 . OpenUrl 22. ↵ Panch , T. , Szolovits , P. , & Atun , R. ( 2018 ). Artificial intelligence, machine learning and health systems . Journal of Global Health , 8 ( 2 ), 020303 . OpenUrl PubMed 23. ↵ Prosperi , M. , Guo , Y. , Sperrin , M. , Koopman , J. S. , Min , J. S. , He , X. , … & Bian , J. ( 2020 ). Causal inference and counterfactual prediction in machine learning for actionable healthcare . Nature Machine Intelligence , 2 ( 7 ), 369 – 375 . OpenUrl 24. ↵ Wang , H. H. , Wang , J. J. , Wong , S. Y. , Wong , M. C. , Li , F. J. , Wang , P. X. , … & Mercer , S. W. ( 2014 ). Epidemiology of multimorbidity in China and implications for the healthcare system: cross-sectional survey among 162,464 community household residents in southern China . BMC Medicine , 12 ( 1 ), 188 . OpenUrl PubMed 25. ↵ Weimann , A. , Dai , D. , & Oni , T. ( 2016 ). A cross-sectional and spatial analysis of the prevalence of multimorbidity and its association with socioeconomic disadvantage in South Africa: A comparison between 2008 and 2012 . Social Science & Medicine , 163 , 144 – 156 . OpenUrl PubMed View the discussion thread. Back to top Previous Next Posted October 15, 2025. Download PDF Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following A Machine Learning Approach to Prediction and Multimorbidity Risk Factor Identification in a low- and middle-income country Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share A Machine Learning Approach to Prediction and Multimorbidity Risk Factor Identification in a low- and middle-income country Olalekan A. Uthman , Matthew Hazell , Muhammed Mubashir Babatunde Uthman , Kolawole W Wahab , Ponnusamy Saravanan , Paramjit Gill , Andre Pascal Kengne medRxiv 2025.10.13.25337900; doi: https://doi.org/10.1101/2025.10.13.25337900 Share This Article: Copy Citation Tools A Machine Learning Approach to Prediction and Multimorbidity Risk Factor Identification in a low- and middle-income country Olalekan A. Uthman , Matthew Hazell , Muhammed Mubashir Babatunde Uthman , Kolawole W Wahab , Ponnusamy Saravanan , Paramjit Gill , Andre Pascal Kengne medRxiv 2025.10.13.25337900; doi: https://doi.org/10.1101/2025.10.13.25337900 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Public and Global Health Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (300) Cardiovascular Medicine (4436) Dentistry and Oral Medicine (444) Dermatology (382) Emergency Medicine (608) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1509) Epidemiology (15229) Forensic Medicine (30) Gastroenterology (1124) Genetic and Genomic Medicine (6600) Geriatric Medicine (668) Health Economics (997) Health Informatics (4538) Health Policy (1368) Health Systems and Quality Improvement (1613) Hematology (542) HIV/AIDS (1264) Infectious Diseases (except HIV/AIDS) (15916) Intensive Care and Critical Care Medicine (1103) Medical Education (623) Medical Ethics (146) Nephrology (667) Neurology (6599) Nursing (346) Nutrition (998) Obstetrics and Gynecology (1144) Occupational and Environmental Health (957) Oncology (3333) Ophthalmology (974) Orthopedics (369) Otolaryngology (420) Pain Medicine (436) Palliative Medicine (130) Pathology (663) Pediatrics (1693) Pharmacology and Therapeutics (691) Primary Care Research (711) Psychiatry and Clinical Psychology (5447) Public and Global Health (9232) Radiology and Imaging (2198) Rehabilitation Medicine and Physical Therapy (1370) Respiratory Medicine (1196) Rheumatology (593) Sexual and Reproductive Health (712) Sports Medicine (530) Surgery (712) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a00e6d603b0806eb',t:'MTc3OTY0ODAxMA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();