Income-Level Heterogeneity in the Sustainable Development–Human Development Nexus: Evidence from Machine Learning

Abstract

Human development is increasingly expected to reflect progress in health, education, living conditions, and sustainability. Yet evidence on how specific Sustainable Development Indicators (SDIs) relate to such progress remains limited, especially in studies that jointly consider cross-income heterogeneity, high-dimensional indicators, and nonlinear relationships. This study examines the SDI–HDI relationship across low-, lower-middle-, upper-middle-, and high-income countries using 408 World Bank SDG indicators and UNDP HDI series for 1990–2020. An interpretable Random Forest framework, combined with SHAP rankings and Partial Dependence Plots, identifies the most influential predictors and marginal associations with HDI. The model shows strong predictive performance across income groups and marked heterogeneity in the predictors associated with HDI. In low-income countries, HDI is mainly associated with early-life health conditions and human capital; in lower-middle-income countries, electrification and service access become more prominent; and in upper-middle- and high-income groups, digital connectivity, higher education, and institutional factors gain importance. Mortality-related indicators are consistently associated with lower predicted HDI, whereas literacy, electricity access, and internet use are associated with higher HDI. These results highlight how AI-based analytical tools can support sustainable economic development by identifying income-specific development priorities and structural constraints. They also suggest that disparities in health, education, infrastructure, and digital connectivity may influence the conditions under which entrepreneurial opportunities emerge or remain constrained across development stages. Overall, the SDI–HDI relationship is nonlinear and income-specific, supporting more differentiated, data-driven development strategies.

1. Introduction

Economic development (ED) has long been viewed from a growth-accounting perspective, proxied by economy-wide measures of production and income, gross domestic product (GDP) and gross national income (GNI), alongside indicators of capital accumulation and efficiency such as capital formation and labour productivity. These variables provide the foundations for growth models, including the Solow–Swan model and the Harrod–Domar model. Yet economic development cannot be reduced to aggregate expansion; its relevance lies in whether economic progress delivers sustained improvements in living standards and human well-being [1].

Sustainable development (SD) is increasingly pursued in contexts where firms and governments face intertwined constraints, including climate risks, rapid urbanization, and socio-cultural resistance that complicate long-run policy design [2,3]. These pressures are not neutral; they can widen disparities and reinforce inequality, especially in developing countries [4]. Within this context, economic development is better understood as more than aggregate growth: it refers to durable improvements in living standards and social welfare, often supported by structural transformation toward higher-productivity activities [5]. Several contributions therefore argue that development strategies should integrate sustainability considerations from early stages [6]. At the same time, the green-growth debate highlights a central tension, as some authors contend that “green” growth cannot be achieved without confronting the environmental burdens associated with economic activity [7].

Evidence linking sustainable development (SD) to human development through Sustainable Development Indicators (SDIs) remains fragmented and relatively limited, despite the centrality of this relationship in policy debates [8,9,10]. A key limitation lies in the continued reliance on conventional statistical and econometric specifications, which often impose restrictive functional forms. In particular, linearity assumptions may be too restrictive to capture the nonlinear and interactive mechanisms through which SDIs relate to development outcomes [11]. Reflecting these limitations, some studies have explored alternative modelling strategies for development-related questions [12,13]. In parallel, machine learning (ML) has attracted growing interest because it can learn complex structures directly from data, including nonlinearities and interactions, while often offering stronger predictive performance in high-dimensional settings [14,15,16,17,18,19].

Despite the growing literature on sustainable development and development outcomes, the specific relationship between SDIs and the Human Development Index (HDI) remains insufficiently examined in three important respects. First, much of the existing evidence relies on relatively narrow subsets of indicators or aggregate sustainability proxies, which limits the ability to identify which specific dimensions of sustainable development are most closely associated with human development. Second, prior studies rarely assess whether these relationships differ systematically across countries at different stages of development, despite the likelihood that SDI priorities are income-contingent. Third, conventional econometric approaches are often less suited to capturing the nonlinearities and interaction effects that may characterize the relationship between a high-dimensional SDI space and HDI. This study addresses these limitations jointly by combining a broad SDI dataset, a comparative income-group design, and an interpretable machine-learning framework capable of identifying both variable importance and marginal predictive patterns.

While machine-learning approaches have already been applied in parts of the sustainability and development literature, their use has rarely been directed toward a comparative examination of the SDI–HDI relationship across countries at different stages of development within a broad and interpretable indicator framework. The contribution of the present study therefore lies not simply in the use of machine learning itself, but in the joint integration of three elements that remain insufficiently combined in the existing literature: a high-dimensional SDI dataset covering 408 indicators, a comparative cross-income-group design, and an interpretable modelling strategy combining Random Forest, SHAP rankings, and Partial Dependence Plots. This framing allows the study to move beyond broad sustainability mapping and to provide more differentiated evidence on how SDIs are associated with HDI across development stages.

Accordingly, the central objective of this study is to identify how a broad set of Sustainable Development Indicators is associated with HDI across income groups, and whether these associations vary systematically across development stages within a non-linear and interpretable analytical framework.

Beyond their relevance for macro-level development analysis, SDIs may also have indirect implications for entrepreneurship and SME development, since entrepreneurial activity depends on foundational conditions such as education, infrastructure, health services, digital connectivity, and institutional quality [20]. In this sense, examining the SDI–HDI relationship is also useful for understanding the broader policy environment within which entrepreneurship and small-business development evolve across countries at different stages of development [21]. In the present study, however, this dimension is treated as a secondary policy implication of the findings rather than as a separate empirical focus [22].

To address these gaps, this study examines the SDI–human development relationship using a Random Forest framework across four income groups: low-, lower-middle-, upper-middle-, and high-income countries. The analysis identifies the SDIs with the strongest predictive relevance for development outcomes and assesses the direction of their association with human development. By translating a high-dimensional set of SDIs into a prioritized and interpretable evidence base, the study supports more effective performance management and helps refine policy targeting and implementation strategies.

The empirical analysis draws on two major data sources: the World Bank’s Sustainable Development Goals (SDGs) database and the United Nations Development Programme’s (UNDP) human development metrics, thereby enabling a multidimensional assessment spanning economic, social, and environmental dimensions. These data are used to address two questions: which Sustainable Development Indicators (SDIs) are most strongly associated with HDI, and do the most influential SDIs exhibit predominantly positive or negative associations with human development across countries?

The remainder of the paper is structured as follows. Section 2 reviews the theoretical, empirical, and bibliometric literature relevant to sustainable development and human development. Section 3 presents the dataset, data collection process, and methodological framework. Section 4 reports the machine learning analysis across income groups. Section 5 discusses the findings in relation to prior research. Section 6 concludes.

2. Theoretical, Empirical, and Bibliometric Review

This section reviews the literature on the relationship between sustainable development and human development from bibliometric, theoretical, and empirical perspectives. To maintain focus on the empirical contribution, the bibliometric component is used only as a concise positioning device. Its purpose is to document the broader expansion of the field, identify the limited visibility of the SDI–HDI relationship in comparative cross-income settings, and motivate the research gap addressed in this study. The section then turns to the theoretical and empirical literature most directly relevant to the conceptual framing, research questions, and methodological choices of the paper.

2.1. Bibliometric Positioning of the Research Gap

This subsection provides a concise bibliometric positioning of the research gap related to sustainable development, human development, SDGs, and machine learning. Based on the Scopus database, the analysis helps identify broad publication trends and highlights that the SDI–HDI relationship in a comparative cross-income setting remains relatively underexplored within the wider literature.

Figure 1 indicates that the literature has expanded markedly in recent years, especially after 2022, and that journal articles account for the dominant share of contributions. This pattern helps position the present study within a rapidly growing research area.

Figure 2 shows that the literature is mainly organized around broad themes such as sustainable development, SDGs, artificial intelligence, and machine learning. By contrast, themes more directly aligned with the present study, particularly the SDI–HDI relationship and comparative cross-income analysis, do not appear as central focal points. In this sense, the figure highlights that themes more directly aligned with the present study remain relatively peripheral within the broader literature.

Taken together, the bibliometric evidence is used here to position the present study within the broader literature and to inform the formulation of its research questions. It suggests that, despite the rapid expansion of the field, the SDI–HDI relationship remains relatively underexplored in comparative cross-income settings. This supports the motivation for examining which SDIs are most strongly associated with HDI and whether these associations vary systematically across development stages.

2.2. Theoretical and Empirical Literature Review

At the core of contemporary development debates lies the tension between welfare enhancement and environmental limits. Sustainable development (SD) addresses this tension by embedding an explicitly intergenerational lens into the evaluation of progress, requiring that today’s gains remain compatible with tomorrow’s welfare [23]. In this perspective, economic development strategies built on pollution-intensive expansion, ecosystem degradation, or the depletion of natural resources may deliver short-run improvements while undermining their long-run sustainability, and thus should be revisited [24,25]. Accordingly, integrating responsible production and consumption, alongside strengthened waste management and circular-economy practices, becomes a core condition for development pathways that aim to sustain welfare while containing environmental externalities.

A common starting point in sustainable development research is the triple bottom line framework, which treats sustainability as a multidimensional construct spanning economic prosperity, social welfare, and environmental protection [26]. This framing implicitly places economic development within sustainable development rather than alongside it, while also highlighting the conceptual difficulty of evaluating these dimensions in isolation without introducing distortion. Ref. [27] argues that their interdependence is sufficiently strong that separating them may misrepresent how sustainability and development outcomes are produced.

Beyond the triple bottom line, the sustainable development literature has increasingly relied on complementary frameworks that make ecological limits more explicit. The planetary boundaries framework defines nine biophysical thresholds within which human activity should remain to preserve environmental stability [28]. Along similar lines [29], proposes a doughnut model that conceptualizes sustainable development as meeting essential social foundations while respecting ecological ceilings. Building on these constraint-oriented perspectives, a growing body of research examines how structural forces, particularly urbanization, globalization, and conflict, interact with social and environmental indicators to shape patterns of environmental degradation. For instance [30], find that globalization and urbanization significantly exacerbate environmental harm, and they call on policy-makers to rethink growth strategies that are primarily driven by globalization, especially in conflict-prone contexts.

The United Nations’ 2030 Agenda has become the central reference for translating sustainable development into measurable policy commitments, articulating progress through the 17 Sustainable Development Goals and 169 targets. However, several authors question the operational feasibility of this architecture. Ref. [31] note that gaps in conceptual definition and measurement practice can undermine implementation, while [32] argue that the agenda’s scope, reflected in the large number of goals and targets, reduces its usability for governments by making adoption and routine monitoring demanding in practice.

Academic debates on economic development increasingly revolve around the role of sustainability in shaping growth trajectories. A strand of the literature maintains that development can be achieved without significant environmental or social compromise when sustainable practices are effectively integrated into production systems and policy frameworks [6,24,25]. In contrast, other scholars argue that trade-offs between economic, environmental, and social objectives are structurally embedded in development processes and therefore unavoidable [7]. In line with this perspective [33], emphasizes that development strategies should avoid one-size-fits-all approaches and instead reflect country-specific conditions. As a result, nations pursuing economic growth are generally required to confront and manage the environmental and social repercussions associated with extraction, production, and waste management.

Ref. [34] notes that sustainable development (SD) has been broadly institutionalized across developed countries, where it is consistently reflected in official discourse, national strategies, and legislative instruments. Yet, as underscored by this study, empirical research linking sustainable development indicators (SDIs) to human development still relies disproportionately on evidence drawn from advanced economies, which constrains the generalizability of findings to developing contexts. At the same time, increasing recognition of the strong interdependence between economic activity, social dynamics, and environmental pressures calls for greater analytical caution when assessing development trajectories [35]. Although many developing countries have made relatively limited contributions to global environmental degradation, they are increasingly engaged in sustainability debates because long-run prosperity depends on responsible resource governance and social equity. Their central challenge is to reconcile urgent development imperatives with environmental constraints shaped partly by the legacy and spillovers of high-emission development pathways in high-income economies. In such settings, SD is often difficult to define and even harder to operationalize. Ref. [36] therefore propose treating SD as a pragmatic framework for simultaneously addressing poverty, climate change, and rapid urbanization, emphasizing the need to prioritize poverty reduction and manage accelerated urban growth. Complementing this perspective [37], call on high-emission countries to reduce their carbon footprints and suggest that emissions in sub-Saharan Africa are likely to rise substantially as development accelerates.

Evidence from diverse contexts continues to suggest that growth may be associated with environmental pressure, particularly at early stages of development. In the Baltic region, Ref. [38] document a strong association between GDP growth and environmental degradation, in line with the Environmental Kuznets Curve intuition that pollution can rise alongside income at lower development levels. However, although their analysis links several macroeconomic indicators (e.g., employment and exchange rates) to GDP across development stages, it remains largely confined to short-run economic–environmental performance and gives limited attention to social sustainability dimensions such as inequality and access to public goods, which are central to a comprehensive SD perspective. Ref. [39] introduces the carbon intensity of poverty reduction (CIPR) and shows a non-linear pattern whereby economic growth initially reduces CIPR before increasing it beyond a threshold, yet the study does not incorporate sector-level mechanisms, policy channels, or technological change that could potentially ease the growth–environment trade-off. Focusing on emerging economies [40], report that GDP is positively associated with poverty and negatively associated with gender equity, and they highlight the need for more nuanced empirical strategies, particularly for the GDP–hunger relationship to better capture complex causal linkages among these variables.

Standard assessments of “economic progress” have traditionally relied on national accounting aggregates, such as GDP, GNI, and gross national product (GNP) as key indicators. Yet, recent scholarship cautions that these aggregates largely describe the volume of economic output and may fail to reflect how prosperity translates into lived well-being and living standards [41,42]. This limitation has motivated the use of alternative indices, including the Multidimensional Poverty Index (MPI) and the Gender Development Index (GDI), which emphasize distributional and social aspects of development. Even so, these indices often provide a fragmented view, since they tend to isolate a single pillar of sustainable development rather than capturing its broader multidimensional nature.

In response to the limitations of output-based measures, the United Nations Development Programme (UNDP) introduced the Human Development Index (HDI) in 1990 as a composite indicator designed to capture broader dimensions of development [43]. Compared with GDP-centered measures, the HDI is widely viewed as more comprehensive because it incorporates key capabilities, most notably education and health, and therefore places human well-being and quality of life at the core of assessment [43]. As a consequence of its multidimensional construction, country rankings based on the HDI often diverge markedly from rankings derived from GDP alone.

Ref. [44] argue that standard progress indicators provide a biased view of sustainability because they emphasize short-run growth and insufficiently incorporate environmental constraints. Their findings indicate that most indicators are more closely associated with non-renewable resource use, whereas associations with renewable resources are generally weak and, in some cases, inverse. The implication is that SD assessment requires complementary, context-specific measures rather than reliance on a single composite indicator.

Recent applications of machine learning in development-related research suggest clear analytical advantages, especially when the objective is prediction rather than causal identification. In contrast to conventional econometric models, machine-learning approaches can accommodate nonlinearities, interactions, and large predictor spaces more flexibly, which makes them attractive for complex development datasets. At the same time, this literature also highlights an important methodological distinction: strong predictive performance does not by itself establish causal mechanisms or policy effects. In this sense, machine learning is especially useful for pattern detection, ranking predictive relevance, and uncovering heterogeneous associations, whereas causal inference requires additional identification strategies that go beyond predictive modelling alone.

High-dimensional indicator models also raise specific challenges that are particularly relevant in development analysis. When a large number of indicators are considered simultaneously, redundancy, measurement heterogeneity, overfitting risk, and instability in variable-importance rankings may complicate interpretation. These issues are especially important when indicators span multiple domains—economic, social, environmental, and institutional—and when the same predictive model is used across heterogeneous country contexts. Accordingly, the value of machine-learning approaches in such settings lies not only in their predictive capacity, but also in the extent to which they are combined with transparent validation procedures and interpretable tools that help distinguish robust predictive patterns from potentially misleading model-driven patterns.

Existing work on sustainable development often remains at an aggregate level, leaving the indicator-specific connections between Sustainable Development Indicators (SDIs) and human development insufficiently documented across heterogeneous income contexts. Moreover, the literature has only partially capitalized on large-scale datasets that could help uncover complex SDI–HDI patterns, while often leaving unclear how predictive modelling should be interpreted relative to causal inference. To move this line of inquiry forward, the present study adopts a prediction-oriented and interpretable machine-learning framework, combining extensive World Bank and UNDP data with Random Forest, SHAP rankings, and Partial Dependence Plots to examine how SDIs are associated with HDI across four income groups: low-, lower-middle-, upper-middle-, and high-income countries.

3. Methodology

A quantitative empirical approach was adopted, drawing on two prominent data sources: the SDGs database of the World Bank (2025) and the United Nations Development Programme’s (UNDP) all-composite indices and components time series dataset (UNDP, 2025). To account for cross-country heterogeneity in development patterns, the empirical assessment was conducted across four income groups defined by the World Bank, namely low-income, lower-middle-income, upper-middle-income, and high-income countries. The stages followed in the research methodology include dataset selection and collection, and data analysis methods covering data pre-processing, algorithm selection and training, feature selection, and model evaluation.

3.1. Dataset Selection and Collection

Robust empirical analysis in machine learning relies on the careful selection of data sources and on their ability to provide consistent measurements across countries and over time. Accordingly, SDIs were obtained from the World Bank Sustainable Development Goals dataset, while HDI series were sourced from the UNDP “all-composite indices and components time series” dataset. The dataset descriptions are as follows:

3.1.1. World Bank (2025) Data Collection

Sustainable Development Indicators were operationalized using the World Bank SDG database, which offers a comprehensive set of 408 indicators capturing economic, social, and environmental conditions at the country level. Rather than serving solely as descriptive measures, these indicators provide a structured empirical basis for examining development dynamics across heterogeneous economies. In this study, the complete set of SDIs was employed to characterize sustainable development profiles, and their association with human development was examined through a Random Forest (RF) modelling framework.

3.1.2. UNDP (2025) Data Collection

Human development was measured using the Human Development Index (HDI) obtained from the UNDP time-series database. As a composite measure, the HDI consolidates health, education, and living-standard dimensions into a single index, offering a development benchmark that is not confined to output-based performance. This choice supports a human-centred interpretation of development progress rather than an exclusive focus on economic scale.

As regards the reliability and coverage of the underlying data, prior studies emphasize that World Bank and UNDP datasets are widely used, publicly accessible, and appropriate for mitigating potential measurement and selection biases in cross-country research [45,46]. Related evidence further recognizes these institutions for their comprehensive global coverage [47,48]. Relying on such reputable sources therefore enhances transparency and reproducibility, facilitating replication and extension of the analysis.

3.2. Data Analysis Method

To examine the relationship between SDIs and HDI, this study applied the Random Forest (RF) algorithm, which is widely used for supervised learning tasks due to its strong predictive performance and robustness to complex, nonlinear relationships. Prior research has argued that machine learning approaches can outperform conventional modelling frameworks in prediction-oriented settings, often delivering higher accuracy and allowing more flexible uncertainty quantification. In this regard [49], highlights the superior ability of ML models to quantify predictive uncertainty relative to traditional methods. Related evidence also reports higher predictive correctness and accuracy for ML techniques compared with alternative approaches, including standard econometric models [50,51].

More specifically [19], emphasize the distinct contribution of machine learning to economic analysis by improving predictive performance and supporting evidence-based decision-making. In particular, they note that Random Forest (RF) is well suited to settings characterized by complex and nonlinear associations that may be insufficiently captured by conventional regression specifications. In line with this argument, RF was selected to accommodate nonlinearities, higher-order interactions, and complex relationships between high-dimensional predictors and development outcomes. As emphasized by [52], standard regression models may become restrictive in the presence of such complexity, whereas RF enables automated estimation under minimal modelling assumptions and delivers transparent assessments of variable importance. These properties make RF appropriate for evaluating how heterogeneous SDIs relate to human development outcomes across the income-group sample.

3.3. Data Pre-Processing

The raw dataset contained missing observations across countries, indicators, and years. To preserve the cross-country and longitudinal coverage of the dataset, missing values were handled using iterative imputation, which exploits multivariate relationships among indicators to generate refined estimates. Outliers were identified using the interquartile range (IQR) criterion and treated through winsorization in order to reduce the influence of extreme values while preserving the structure of the data. Because the indicators were expressed in heterogeneous units, all explanatory variables were standardized using z-score normalization to ensure comparability across features and to support stable model estimation.

3.4. Algorithm Training and Feature Selection

Random Forest (RF) constructs multiple decision trees through a two-step randomization process. The first randomization relies on bootstrap sampling or “bagging”, in which new training sets are generated by selecting subjects with replacement, leaving a portion of observations as an “out-of-bag” (OOB) sample for internal validation [53]. The second randomization is applied at each tree’s decision nodes, where a random subset of predictor variables is evaluated to identify the best split [54]. For regression settings, the results are averaged across trees [55], which helps reduce overfitting while capturing complex patterns and interactions in the data.

The Random Forest (RF) model was trained on the pre-processed dataset comprising 408 SDIs, covering countries across four World Bank income groups (low-income, lower-middle-income, upper-middle-income, and high-income). To reduce the risk of data leakage arising from repeated country-year observations, the dataset was partitioned at the country level rather than through a purely random observation-level split. Accordingly, all country-year observations belonging to a given country were assigned either to the training set or to the test set, but not to both. The final data partition retained 80% of countries for training and 20% for testing. During training, the RF algorithm characterized the SDI–HDI relationship by building an ensemble of decision trees from randomized data subsets, thereby capturing linear and nonlinear relationships while improving generalization through variance reduction. HDI predictions were generated by averaging outputs across trees, enhancing stability. Finally, RF feature-importance rankings highlighted the SDIs most influential for HDI across income groups, providing deeper insights beyond overall model performance.

Hyperparameter selection was conducted using randomized search with grouped cross-validation on the training set, with country identifiers used as grouping units. To ensure comparability across income-group models, the final Random Forest configuration was kept identical for all four groups. The retained hyperparameters were n_estimators = 800, max_depth = 20, max_features = “sqrt”, min_samples_split = 2, and min_samples_leaf = 1. This common specification was used for the low-, lower-middle-, upper-middle-, and high-income country groups.

For model validation, grouped cross-validation was conducted on the training set using country identifiers as grouping units, so that observations from the same country did not appear simultaneously in both training and validation folds. This grouped validation strategy was used to account for within-country dependence in the country-year structure of the dataset and to reduce the risk that predictive performance was overstated by leakage across repeated observations from the same country. In addition, the Random Forest framework benefits from an internal validation logic through out-of-bag observations generated during bootstrap aggregation. Taken together, these validation procedures strengthen the robustness of the empirical results and reduce the risk that model performance is driven by a single data partition. The present analysis was not designed as a formal baseline-model comparison exercise, as its main objective was to examine and interpret the SDI–HDI relationship within a Random Forest framework suited to high-dimensional predictors and nonlinear associations.

3.5. Model Evaluation

Model evaluation relied on several widely used metrics, namely Mean Squared Error (MSE), Root Mean Squared Error (RMSE), R-squared (R²), and Mean Absolute Error (MAE). MSE measures the average squared difference between predicted and actual values [56], while RMSE, as the square root of MSE, provides a more interpretable error measure. R² indicates the extent to which the model explains variation in the outcome variable, whereas MAE computes the average absolute difference between predicted and observed values, offering a robust indicator of model accuracy [57,58]. These evaluation metrics were essential for assessing RF performance and verifying that the model adequately represented predictive patterns in the SDI–HDI relationship. Accordingly, their combined use enabled a comprehensive assessment of predictive capability and reinforced the robustness of the results.

In studies using ML models such as RF, model performance is commonly assessed through statistical metrics that capture predictive accuracy and stability across datasets. In this context, R² indicates the proportion of variation in the target variable accounted for by the model, while MSE, RMSE, and MAE summarize prediction error from complementary perspectives [59,60,61,62]. Higher R² values, especially those approaching 1, together with low error values, generally indicate stronger predictive performance and better generalization across data partitions. In the present study, these metrics are interpreted strictly in predictive terms and not as evidence of causality. Consistent with the results reported in Section 4.2, the model outputs therefore support accurate and generalizable interpretation of the predictive relationship between SDIs and HDI across the income-group setting.

4. Machine Learning Analysis Across Income Groups

4.1. Research Context

The research population covers all countries worldwide. To account for structural heterogeneity across development stages, this study adopts an income-stratified design based on the World Bank classification and examines countries grouped into four categories: low-income, lower-middle-income, upper-middle-income, and high-income. This stratification is appropriate because development constraints and policy trade-offs, as well as the sustainability dimensions most closely associated with development performance, are expected to differ systematically across income levels.

Empirically, the analysis relies on harmonized international indicators over 1990–2020, drawing 408 Sustainable Development Indicators from the World Bank SDGs database and measuring human development using the Human Development Index (HDI) from UNDP. Comparing income groups provides a standardized basis to assess how the model-based relative importance of SDIs varies across income groups, while limiting reliance on a small set of country cases and reducing the influence of country-specific idiosyncrasies.

4.2. Empirical Analysis

The datasets obtained from the World Bank and UNDP were analyzed using the Python programming language (version 3.14.5). The analysis process started with data cleaning and data transformation to convert the raw files into a usable format and to ensure that the data were ready for model training and further prediction. The data preprocessing stage addressed several key issues, including data normalization, so that features with different scales were standardized to support effective model learning. At this stage, the datasets were imported and cleaned, and the shape and data types were checked to confirm consistency across the dataset. In addition, missing values were identified for each country across different years and were recorded as null values for the purpose of analysis.

Inspection of the “Country Name” field indicated that the sample includes countries classified into four income groups: low-income (27 countries), lower-middle-income (55 countries), upper-middle-income (56 countries), and high-income (79 countries). The “Series Name” field contained 409 unique values, corresponding to 408 explanatory features and one target variable. The columns labeled “1990” through “2020” span 31 years of available data. Following these preparation steps, results are presented in two parts: (1) descriptive analysis and (2) ML model outputs.

4.2.1. Descriptive Analysis

To establish a descriptive baseline for Human Development Index (HDI) dynamics across income groups, HDI data spanning a 31-year period (1990–2020) were examined for high-income, upper-middle-income, lower-middle-income, and low-income economies. Figure 3 depicts the corresponding HDI trends, highlighting the relative development trajectories of these groups over time.

As shown in Figure 3, the high-income group consistently records the highest HDI levels throughout the period, exhibiting a steady upward trajectory with only a slight softening at the end of the series. In contrast, the low-income group remains the lowest-performing category across all years, although it shows gradual improvement over time. The upper-middle-income and lower-middle-income groups follow intermediate paths, both displaying sustained increases, with the upper-middle-income group remaining clearly above the lower-middle-income group. Overall, the figure indicates broad-based HDI gains across all income groups, while also highlighting a persistent income-related development gap that remains substantial over the full 1990–2020 horizon.

To complement the HDI-based perspective with an economic benchmark, Figure 4 reports GDP per capita trends over 1990–2020 for four income groups: high-income, upper-middle-income, lower-middle-income, and low-income economies. GDP per capita is commonly used as an indicator of average material living standards, computed as total GDP divided by population, and thus provides a standardized measure of economic prosperity across groups.

As shown in Figure 4, the high-income group consistently exhibits the highest GDP per capita throughout the period and follows a generally upward trajectory, with short-lived slowdowns in some years. The upper-middle-income group also shows sustained gains, remaining markedly above the lower-middle-income and low-income groups, which display lower levels despite gradual improvement over time. When considered alongside the HDI trends in Figure 3, the persistence of wide GDP-per capita gaps suggests that improvements in average income have not eliminated cross-group disparities in human development, pointing to structural differences in development capacity and the effectiveness of translating economic gains into broad-based welfare outcomes (e.g., education, health, and infrastructure).

4.2.2. Machine Learning Model Outputs

Following the descriptive analysis, the next step involved implementing the Random Forest (RF) algorithm to predict human development, measured by the Human Development Index (HDI), using the Sustainable Development Indicators (SDIs) as independent variables. The RF model was trained on the SDI datasets for each income group (low-, lower-middle-, upper-middle-, and high-income), with the objective of identifying which indicators exert the strongest predictive relevance for HDI across heterogeneous development contexts. As discussed in Section 3.5, several evaluation metrics were employed to assess the accuracy and predictive performance of the RF model, and the resulting model performance is reported in

Table 1. Model Evaluation (Random Forest) by income group.

Income Group	Mean Absolute Error (MAE) for RF	Test Mean Squared Error (MSE) Score	Test R-Squared (R2) Score	Test Root Mean Squared Error (RMSE) Score
Low income	0.00828	0.00014	0.99082	0.01206
Lower-middle income	0.00785	0.00013	0.98740	0.01177
Upper-middle income	0.00744	0.00012	0.97184	0.01120
High income	0.00627	0.00008	0.98230	0.00926

.

The RF algorithm used in this research demonstrated strong predictive accuracy for human development, measured by the Human Development Index (HDI), across the four income groups (low-, lower-middle-, upper-middle-, and high-income). The evaluation metrics provided a comprehensive view of model performance, indicating both high predictive power and low prediction errors. As reported in Table 1, the RF model achieved high test R² values across all groups, ranging from 0.97184 (upper-middle income) to 0.99082 (low income), suggesting that SDI-based predictors explain a substantial share of the variation in HDI within each income category. In addition, the relatively small MAE (0.00627–0.00828) and RMSE (0.00926–0.01206) values further confirm the strong predictive performance of the RF model across heterogeneous development contexts.

The predictive accuracy of the model is illustrated in Figure 5, Figure 6, Figure 7 and Figure 8, which show how closely the RF predictions align with the observed HDI values across the four income groups. Overall, the results indicate a strong model fit, with predicted values closely tracking the actual observations for the low-, lower-middle-, upper-middle-, and high-income groups. This predictive performance is further supported by the following key evaluation metrics:

• Mean Absolute Error (MAE): The MAE values ranged between 0.00627 and 0.00828, indicating that the model made only small prediction errors across all income groups. This high level of precision suggests that the RF model effectively captured the key SDI-related factors associated with HDI within each development context.
• Mean Squared Error (MSE): The MSE scores were consistently small, ranging from 0.0000859 to 0.0001456, confirming low overall prediction variance and indicating that the RF predictions remained stable across income groups. These low MSE values point to minimal large deviations between predicted and observed HDI values.
• R-squared (R²): The test R² values ranged from 0.97184 to 0.99082, implying that the RF model explained between 97.18% and 99.08% of the variation in HDI across the four income groups. This strong model fit suggests that the selected SDIs capture a large share of the relevant variation in HDI across income groups.
• Root Mean Squared Error (RMSE): The RMSE values were low, ranging from 0.00926 to 0.01206, further underscoring the model’s accuracy in predicting HDI across income categories. Since RMSE expresses error on the same scale as the outcome variable, these small values indicate that the RF model produced highly accurate predictions with limited error.

Taken together, Figure 5, Figure 6, Figure 7 and Figure 8 confirm that the Random Forest model reproduces the observed HDI values with a high degree of accuracy across all four income groups. In Figure 5, the low-income group shows a close alignment between predicted and observed values, with only minor deviations across the distribution. Figure 6 presents a similarly strong predictive pattern for the lower-middle-income group, with predicted values continuing to track observed HDI closely. In Figure 7, the upper-middle-income group also demonstrates consistent alignment, although slightly wider dispersion is visible in some ranges. Figure 8 shows that predictions remain highly accurate for high-income countries, where HDI values are more concentrated and deviations remain limited. Overall, the consistency across Figure 5, Figure 6, Figure 7 and Figure 8 indicates that the Random Forest model captures the main variation in HDI across heterogeneous development contexts without large systematic errors. Having established that predictive performance remains strong across income groups, the next step is to examine which Sustainable Development Indicators (SDIs) contribute most to this predictive accuracy within each income group.

Figure 9, Figure 10, Figure 11 and Figure 12 present the SHAP-based rankings of the most influential predictors of HDI for each income group. Figure 9 reports the results for low-income countries, while Figure 10, Figure 11 and Figure 12 present the corresponding rankings for lower-middle-, upper-middle-, and high-income groups, respectively.

It is important to note that SHAP-based rankings identify the variables that contribute most to the model’s predictive performance, but they do not in themselves establish causal effects. Accordingly, the following interpretation focuses on predictive importance within the trained model rather than on direct causal mechanisms.

Figure 9 reports the top SHAP feature importance for the low-income group. The ranking indicates that early-life health outcomes dominate the model’s predictive importance ranking: under-five (female, male, and total) and neonatal mortality rates are the most influential predictors of HDI, indicating that these variables are strongly associated with variation in model-predicted HDI in low-income countries. Education-related indicators (youth/adult literacy, primary completion, and tertiary enrolment) follow, highlighting the predictive relevance of human capital-related conditions. Health-system capacity and administrative reach, proxied by health workforce density and the completeness of rural birth registration, also contribute, suggesting that institutional coverage is closely associated with model-predicted HDI. Finally, WASH (water, sanitation and hygiene)-related mortality, GDP per capita (PPP), and the share of wage and salaried workers represent additional predictive features associated with basic infrastructure, income conditions, and labor-market formalization in relation to model-predicted HDI.

Figure 10 presents the top SHAP feature importance for the lower-middle-income group. The ranking shows that early-life health outcomes remain among the most influential predictors of HDI within the model, with neonatal and under-five mortality indicators (male, female, and total) among the most influential features. A notable shift relative to the low-income group is the prominent role of energy infrastructure: access to electricity (overall, rural, and urban) emerges as a key predictive feature, suggesting that basic service connectivity is increasingly associated with HDI prediction at this income level. WASH-related mortality (water, sanitation and hygiene) also contributes to predictive performance, indicating persistent associations between living conditions and model-predicted HDI. Education-related variables, including primary completion, female lower-secondary completion, and tertiary enrolment, further highlight the predictive relevance of human capital-related conditions. Finally, income proxies (GNI per capita in US$ and PPP terms) also appear among the important features associated with HDI variation within the lower-middle-income group.

Figure 11 presents the top SHAP feature importance for the upper-middle-income group. The results indicate that early-life health outcomes remain highly influential, as under-five (male, female, and total) and neonatal mortality rates are among the top predictors of HDI. A salient shift at this income level is the prominence of digital connectivity: the share of individuals using the Internet ranks among the leading features, suggesting that connectivity and digital inclusion increasingly differentiate model-predicted HDI outcomes. Higher education indicators (tertiary enrolment) also contribute, highlighting the predictive relevance of skills and human capital in more diversified economies. In addition, multiple income proxies, GDP per capita and GNI per capita, in both nominal and PPP-adjusted terms, appear in the top rankings, indicating that living standards and purchasing power remain strongly associated with variation in model-predicted HDI within the upper-middle-income group.

Figure 12 presents the top SHAP feature importance for the high-income group. The ranking indicates that health outcomes remain highly informative even at advanced development levels, with under-five (total, female, and male) and neonatal mortality among the leading predictors of HDI. Digital inclusion also plays a central role, as the share of individuals using the Internet ranks prominently, indicating that connectivity and digital diffusion remain strongly associated with model-predicted HDI in high-income countries. In addition, multiple living-standard proxies are strongly represented in the top rankings: GDP per capita and GNI per capita, measured in nominal, PPP-adjusted, and constant-price terms, appear among the important predictive features associated with variation in model-predicted HDI within high-income countries. Notably, institutional and policy-related indicators emerge in this group, including the Women, Business and the Law index and the applied tariff rate on manufactured products, suggesting that these indicators increasingly differentiate model-predicted HDI outcomes in high-income countries. Higher education participation (tertiary enrolment) remains a relevant contributor, indicating that advanced human capital remains strongly associated with model-predicted HDI in high-income economies.

After identifying the most influential indicators through feature-importance and SHAP-based rankings, it is essential to clarify how these predictors are associated with model-predicted HDI. Importance scores quantify a variable’s contribution to predictive performance, but they do not reveal the direction of the association (positive vs. negative) nor whether the relationship is linear or characterized by thresholds. To examine directionality and potential non-linear effects, the ranking analysis is therefore complemented with Partial Dependence Plots (PDPs), which depict the marginal effect of an indicator on predicted HDI as implied by the trained model.

In the low-income group, Figure 13 shows that the Partial Dependence Plots (PDPs) reveal a strong and monotonic negative association between child/neonatal mortality and model-predicted HDI, with pronounced non-linearities suggesting threshold regions where changes in mortality correspond to sharper shifts in predicted HDI. In contrast, youth male literacy exhibits a positive marginal effect on predicted HDI, becoming more pronounced at higher attainment levels. Overall, these Partial Dependence Plot patterns indicate that early-life health conditions and basic human capital remain strongly associated with variation in model-predicted HDI in low-income settings.

Figure 14 shows that, within the lower-middle-income group, mortality indicators continue to track predicted HDI negatively, but the PDPs also reveal a clearer role for basic infrastructure. Access to electricity displays a marked improvement in predicted HDI once coverage moves into the upper range, followed by diminishing gains near universal access. This pattern points to a development-stage shift in which health remains important, while electrification becomes a key differentiator of HDI performance among lower-middle-income countries.

In the upper-middle-income group, Figure 15 indicates a clear shift in the predictive patterns captured by the Partial Dependence Plots (PDPs). Mortality indicators (under-five and neonatal) remain negatively associated with model-predicted HDI, with most of the change occurring within relatively low mortality ranges and becoming more gradual beyond that range. At the same time, Internet use shows a strong positive marginal effect: predicted HDI rises rapidly as digital adoption increases from low to intermediate levels, and the gains become smaller as Internet use approaches higher levels. Overall, Figure 15 points to a development-stage transition in which early-life health conditions continue to matter, while digital connectivity increasingly distinguishes HDI outcomes among upper-middle-income countries.

In the high-income group, Figure 16 shows that the Partial Dependence Plots (PDPs) continue to associate child and neonatal mortality with lower model-predicted HDI; however, the strongest changes are concentrated within very low mortality ranges, after which the curves become largely stable. In parallel, Internet use displays a consistent positive marginal effect, with predicted HDI increasing steadily as digital connectivity expands, and the rate of improvement becoming more gradual at higher levels. Overall, Figure 16 suggests that, in high-income countries, residual variation in early-life health remains relevant, while digital connectivity plays a prominent role in differentiating HDI outcomes.

5. Discussion

In comparison with prior development research, this study contributes to the ongoing debate on how human development should be conceptualized and assessed across heterogeneous income contexts. Consistent with earlier critiques of GDP-centered development metrics [41,63,64], the findings reinforce that income-based indicators alone are insufficient to fully capture development performance. As illustrated in Figure 3 and Figure 4, persistent disparities between GDP per capita trajectories and HDI outcomes across income groups suggest that improvements in average income do not automatically translate into proportional gains in human development. While high-income economies maintain both higher GDP per capita and HDI levels, lower-income groups exhibit persistent development gaps, indicating that the conversion of economic expansion into health, education, and living-standard outcomes remains uneven. Taken together, these patterns support the argument that development assessment requires multidimensional indicators that incorporate social, infrastructure, and institutional dimensions, rather than relying exclusively on output-based measures.

These findings can also be situated in relation to alternative strands of development research. While a large body of work continues to rely on aggregate macroeconomic indicators or narrower sustainability proxies, other studies have increasingly called for multidimensional and context-sensitive approaches to development assessment. The present results support this latter perspective by showing that the SDI–HDI relationship cannot be reduced to a single uniform pattern across countries. At the same time, the use of an interpretable machine-learning framework extends beyond conventional linear modelling by revealing heterogeneous predictive structures and nonlinear associations that may remain underdetected in more restrictive empirical settings.

Building on calls for integrative sustainability measurement that combines biophysical and socio-economic dimensions [44], the results highlight the limits of relying on narrow sustainability proxies or single-domain indices. While parts of the sustainability-measurement literature emphasize resource- and energy-related signals, such approaches may not fully capture the institutional, social, and infrastructure dimensions through which sustainability-related conditions are associated with development outcomes. This study extends that perspective by leveraging a high-dimensional sustainability space with 408 SDIs spanning health, education, infrastructure, environment, and governance, and by showing that the SDI–HDI relationship varies systematically across income groups. In particular, SHAP-based rankings reveal a development-stage gradient: early-life health and basic human capital dominate in low-income settings; electrification and basic services gain prominence in lower-middle-income economies; and digital inclusion and institutional dimensions increasingly differentiate HDI outcomes in upper-middle- and high-income groups. Beyond expanding indicator coverage, the framework moves beyond static correlation reasoning by using a Random Forest structure that captures non-linearities and threshold-like patterns, which are further clarified through Partial Dependence Plots. This combination provides a more nuanced, development-stage-specific mapping of sustainability priorities.

Ref. [33] cautioned against one-size-fits-all development strategies and emphasized the need for policies tailored to context-specific constraints and capabilities. The evidence reported in this study is consistent with that view, showing that the SDI–HDI relationship varies systematically across income groups rather than following a uniform pattern. Although a common core of predictive features remains visible across different stages of development, most notably early-life health conditions (under-five and neonatal mortality) and living-standard proxies (GDP/GNI per capita), their relative importance shifts markedly with income level. In low-income settings, the dominant predictive signals are associated with survival and basic human-capital conditions, whereas lower-middle-income economies display a stronger role for basic infrastructure connectivity, particularly electrification. In upper-middle- and high-income groups, differentiation increasingly arises from digital inclusion (Internet use), higher education participation and, in advanced settings, institutional and regulatory dimensions. This heterogeneity implies that SDG-oriented policy design and performance management frameworks should be income-contingent: interventions that yield high marginal gains in HDI at early stages may differ from those that matter most once basic needs are largely met.

To address the question of whether the most influential SDIs are associated with higher or lower levels of human development, it is important to distinguish predictive importance from directionality. Feature-importance or SHAP-based rankings identify which indicators contribute most to predictive performance, but they do not, by themselves, establish whether the relationship is positive or negative, nor whether effects are linear. For this reason, SHAP-based rankings are complemented with Partial Dependence Plots (PDPs) to assess marginal direction and potential threshold-like patterns in the SDI–HDI relationship.

The PDP evidence is consistent with the predictive patterns identified by the model: mortality indicators are associated with lower model-predicted HDI, whereas access and capability indicators (e.g., electrification and Internet use) tend to be associated with higher predicted HDI, often with diminishing returns at higher levels. Taken together, these results provide an interpretable, development-stage-specific account of how sustainability-relevant indicators relate to HDI, while acknowledging that the findings reflect conditional associations within a predictive framework. Accordingly, these relationships should be interpreted as predictive associations, and clarifying causal pathways remains an important avenue for future research.

Based on these findings and discussions, the study makes the following academic and practical contributions:

• Broadening Development Metrics: Moving beyond conventional economic aggregates, this study integrates non-economic SDG indicators, including health-related measures (e.g., neonatal and under-5 mortality), social indicators (e.g., demographic and living-condition measures), and environmental indicators (e.g., emissions-related variables), to support a broader, multidimensional assessment of development performance.
• Integrating Machine Learning into Development Analysis: Moving beyond purely descriptive and linear approaches, this study applies an interpretable Random Forest framework, supported by SHAP rankings and Partial Dependence Plots (PDPs), to model HDI using a high-dimensional SDG indicator space, thereby enabling scalable development assessment that captures complex, non-linear relationships across heterogeneous income contexts.
• Income-Group-Specific Insights: The study provides stage-specific insights into the key factors associated with human development by showing that the relative importance of SDG indicators changes systematically across income groups, challenging one-size-fits-all development diagnostics and underscoring the need for strategies tailored to each development stage and its most pressing needs.
• Strengthening Policy and Performance-Oriented Development Diagnostics: By combining HDI-based development measurement with a comprehensive SDG indicator space and an interpretable machine-learning framework, this study proposes an integrated, data-driven design for development diagnostics that can support policy prioritization and monitoring across heterogeneous income contexts, thereby providing a stronger evidence base for performance-oriented development planning and implementation.

These contributions create new avenues for research on sustainable development–human development linkages using performance-oriented indicators and provide practical value by informing policy design, strategic planning, and balanced performance-management frameworks that account for complex, multidimensional influences beyond traditional economic measures.

6. Conclusions

Understanding how sustainable development relates to human development across heterogeneous development contexts has become increasingly important for the design of effective and well-targeted policy strategies. From this perspective, this study addressed an important gap in the literature by moving beyond narrow or single-domain sustainability proxies and by examining the relationship between sustainable development indicators (SDIs) and human development across four World Bank income groups—low-, lower-middle-, upper-middle-, and high-income countries. Unlike conventional empirical approaches that often rely on restrictive linear assumptions, the analysis adopted a Random Forest framework capable of capturing non-linear relationships and complex interactions among variables. Using 408 SDIs drawn from the World Bank SDG database and the Human Development Index (HDI) from UNDP as a measure of human development, the study provided a broader and more flexible assessment of the factors associated with development outcomes across different stages of development.

The findings show that the SDI–HDI relationship is both complex and strongly conditioned by development stage. Although the Random Forest model achieved robust predictive performance across all income groups, the relative importance and marginal effects of the leading indicators differed substantially from one group to another. Early-life health conditions and basic human capital emerged as particularly influential in low-income countries, whereas access to essential infrastructure, especially electrification, became more prominent in lower-middle-income settings. In upper-middle- and high-income groups, digital inclusion, higher education, and institutional dimensions became increasingly important in differentiating variation in model-predicted HDI. The directional analysis further showed that the effects of SDIs are not uniformly positive: mortality-related indicators were associated with lower predicted HDI, while capability- and access-related indicators, such as literacy, electricity access, and Internet use, were generally linked to higher HDI levels. Taken together, these results confirm that development priorities are not universal, but instead vary systematically across income groups.

These findings carry important policy implications. They suggest that development strategies should not be guided by a one-size-fits-all logic, but rather by approaches that are aligned with the structural conditions, constraints, and capabilities specific to each stage of development. From a policy and performance-management perspective, the results support more precise prioritization, better-targeted interventions, and monitoring frameworks that are sensitive to cross-group heterogeneity. In this sense, the study not only contributes to the empirical literature on sustainable development and human development, but also offers a more differentiated basis for the design of data-driven development strategies.

Beyond these broader development implications, the findings also have important implications for entrepreneurship and small business policy. In light of the strong cross-income differences identified in access to infrastructure, digital connectivity, education, and institutional conditions, the results suggest that the foundations supporting entrepreneurial activity also vary across development stages. In lower-income contexts, strengthening basic infrastructure, health, and human capital may help create the minimum conditions necessary for entrepreneurial emergence and small business survival. In more advanced income settings, by contrast, greater emphasis on digital inclusion, higher education, and institutional effectiveness may be more conducive to innovation-driven entrepreneurship and to improving the productivity and resilience of small and medium-sized enterprises. These findings therefore suggest that entrepreneurship policy should not be treated as a standalone agenda, but should be integrated into broader, stage-specific sustainable development strategies.

Beyond its empirical findings, this study contributes to the literature in three main ways. First, it advances development research by moving beyond narrow sustainability proxies and GDP-centred assessment toward a high-dimensional and human-development-oriented analysis of the SDI–HDI relationship. Second, it contributes methodologically by combining a comparative income-group design with an interpretable machine-learning framework capable of capturing nonlinearities, heterogeneous predictive structures, and marginal patterns across development stages. Third, it offers practical value by providing a more differentiated basis for policy prioritization, monitoring, and performance-oriented development planning. In this sense, the novelty of the paper lies not only in the use of machine learning, but in the joint integration of broad SDI coverage, cross-income heterogeneity, and interpretable analysis within a single development-assessment framework.

At the same time, these contributions should be interpreted in light of several limitations, which also point to promising directions for future research. First, although it relies on widely recognized and internationally comparable data sources, namely the World Bank SDG database and the UNDP human development dataset, it necessarily assumes that the published indicators are sufficiently reliable and consistent across countries and over time. Missing observations were addressed using an iterative imputation technique, in which initial estimates were progressively refined based on observed patterns until a stable solution was reached; however, some degree of measurement uncertainty may still remain. Second, while the income-group design improves comparability and helps identify broad development-stage patterns, it may also mask important heterogeneity within each income category. Countries belonging to the same income group do not necessarily share identical institutional configurations, structural constraints, or policy environments, which limits the extent to which group-level findings can fully capture national specificities. Third, although the Random Forest framework, complemented by SHAP-based rankings and Partial Dependence Plots, provides strong predictive and interpretive value, it remains a non-causal approach and therefore does not establish causal relationships between SDIs and HDI. Future research could build on these limitations by incorporating causal identification strategies, exploring within-group heterogeneity at the country or regional level, and comparing the present Random Forest framework with alternative interpretable or hybrid modelling approaches. Because the dataset is structured as repeated country-year observations over 1990–2020, within-country temporal dependence remains an important methodological consideration. Although grouped validation by country was adopted to reduce leakage across repeated observations, future research could extend this framework through more explicit time-aware validation strategies to further assess temporal generalization.