Late detection of breast cancer, contributing to higher death and disability, is a significant global public health priority (Bhattacharya et al., 2022). The need-based targeting of the missed hopes under the widespread screening program in India deserves accurate, precise evidence. India reported approximately 1.46 million new cancer cases, with a projection of 1.57 million by 2025 (Sathishkumar et al., 2022). Globally, breast cancer incidence is projected to increase by more than 40% from 2020 to 2040, and mortality is expected to increase by over 50%, disproportionately impacting low- and middle-income countries (Kim et al., 2025). India had 192,020 new breast cancer cases, which account for almost 28.2% of all female cancers. Similarly, India reported 98,337 deaths due to breast cancer in 2022 (Mortality to Incidence Ratio ≈ 0.512), meaning roughly one death for every two women diagnosed (Sathishkumar et al., 2022; Kim et al., 2025). Despite a higher burden of breast cancer among Indian women (21.8% of the total cancer-related Disability-Adjusted Life Years (DALYs) in 2020) with significant regional disparity (685.5 per 100,000 in northern India and 677.6 per 100,000 in the south), nationwide screening uptake remains extremely low (Jena et al., 2024; International Institute for Population Sciences - IIPS/India and ICF, 2022). Given the benefit of early detection at a localized stage, in reducing cancer mortality by up to 30% (WHO, 2022), only 0.9% of women aged 30–49 in India had ever undergone a clinical breast examination (International Institute for Population Sciences - IIPS/India and ICF, 2022). Tamil Nadu had the highest rate at just 5.6%, and many states reported rates well below 1% (Gopika et al., 2022).

In India, the high cancer burden is attributed to lifestyle changes with varying prevalence by socioeconomic and demographic factors, including education, caste, rural residence, and access dimensions (Smith and Mallath, 2019; Aashima and Sharma, 2024; de Siqueira Filha et al., 2022; George et al., 2020). For example, indigenous women still avoid healthcare due to historical discrimination and a lack of culturally appropriate services (George et al., 2020). Under India’s National Program for Non-Communicable Diseases (NP-NCD), breast cancer screening is delivered via Clinical Breast Examination (CBE) at Health and Wellness Centers. CBE is free of cost and recommended every 5 years for women aged 30–65. Unlike high-income countries, India does not offer population-wide mammography; thus, uptake depends heavily on the availability of trained providers, outreach, and system readiness. An observational cross-sectional study by Negi and Nambiar (2021) using NFHS-4 data on 699,686 women reported that socioeconomic-related inequalities in breast cancer screening coverage were most evident among Christians (Slope Index of Inequality, or SII: 20.6, 95% CI: 18.5–22.7), currently married (SII: 14.1, 95% CI: 13.8–14.4), employed (SII: 14.6, 95% CI: 13.9, 15.3), and rural women (SII: 10.8, 95% CI: 10.5–11.1) compared to their urban, richer, and socially advantaged counterparts measuring the Slope Index of Inequality (SII) (Negi and Nambiar, 2021). Another study assessing the prevalence of cervical cancer screening among women aged 30 to 49 (n = 357,353) using NFHS-5 found wealth-based inequality patterns in screening prevalence (Muthuramalingam and Muraleedharan, 2023). Rahman et al. (2025) reported that the uptake of breast cancer and cervical cancer screening tests was low, at 9 and 20 per 1,000 women, respectively and higher among the richest (Ref. Poorest; SII: 1.1 [breast cancer], 1.8 [cervical cancer]), more pronounced in rural areas (Ref. Urban; Relative Concentration Index or RCI: 22.5 [breast cancer], 21.3 [cervical cancer]), educated women (Ref. No education; relative index of inequality or RII: 4.84 [breast cancer], 2.12 [cervical cancer]). The Northeastern region showed greater socioeconomic inequality (RCI: 32.8), while the Western region showed more education-based inequality (RCI: 30.9) using relative CI (RCI). According to NFHS-5 data, uptake of breast cancer examination was higher in the Southern region, with 23.7 breast cancer examinations per 1,000 women (95% CI: 22.7–24.8), and urban women were more likely to get screened than rural women, with 12.8 breast cancer examinations per 1,000 women (95% CI: 12.1–13.4) (Rahman et al., 2025). A systematic review found that lack of awareness, long waiting times, unavailability of timely appointments, and distance to facilities limit people’s ability to approach screening. Furthermore, social embarrassment, unavailability of female healthcare staff, fear of screening or diagnosis, and cultural beliefs hindered acceptability. Other barriers included insufficient availability of services, affordability issues, and poor appropriateness, failing to match patients’ needs. Additionally, lack of spousal or family support, anxiety, and psychosocial fear further reduced uptake of screening (Srinath et al., 2023).

Despite policy efforts to improve cancer screening coverage in India, significant disparities persist in access to these services, and a detailed predictive analysis of inter-group disparities in screening access remains scarce. The Government of India has responded with programs like the National Program for Non-Communicable Diseases (NP-NCD) and the Ayushman Bharat Health and Wellness Centers (HWCs), rendering population-based screening services with a plan of 200 district-level Day Care Cancer Centers, and exempting 36 lifesaving cancer drugs from customs duties to improve affordability (Press Information Bureau, 2024). Socioeconomic inequities based on caste, economic status, and education continue to shape healthcare access, particularly for preventive services like screening (Negi and Nambiar, 2021). These disparities are neither uniform nor well understood across the country. Limited availability of evidence-based support curbs the effect size of these interventions (Chanakira et al., 2024). Therefore, robust prediction of breast cancer screening uptake, identification of crucial factors that determine the likelihood of uptake, and the underlying inequalities, when applying the most effective tool, are major public health priorities in a resource-scarce setting.

For the last couple of years, studies show that machine learning performs predictions of health service uptake and identifies important predictors or features with higher accuracy and precision in comparison to conventional modelling techniques (Baykemagn et al., 2025; Fahey et al., 2022; Ijaiya et al., 2025; Zegeye et al., 2025). Analysis of a weighted DHS, 2022 dataset of 4 sub-Saharan African nations comprising 33,952 participants, inferred that random forest (RF) performed the best with 78% accuracy and 86% explainability (area under the receiver operating characteristic curve or AUROC) in identifying the predictors of cervical cancer screening uptake (Baykemagn et al., 2025). Likewise, the study using weighted data of 299,759 respondents from DHS sub-Saharan Africa again found RF as the best algorithm (83% accuracy, 89% AUROC) in predicting homebirth and identifying its determinants (Zegeye et al., 2025). Another study applied programmed data of four Nigerian states on 41,394 patients to predict interruptions in treatment among people living with HIV on antiretroviral therapy, who found AdaBoost as the best performer (AUROC = 84%) (Ijaiya et al., 2025). Similarly, a study applied an ensemble decision tree to electronic medical record data and randomized control trial survey data from 184 patients over time. It achieved a mean accuracy of 75.2% in predicting the risk of disengagement (Fahey et al., 2022). Studies focusing on the Indian subcontinent depict the success of machine learning in the prediction of cancer screening uptake (Chandar et al., 2025; Kolasseri and Bhimavarapu, 2025). A longitudinal cancer prevention program (population-based: 14,801 participated in the screening out of 57,270 invited individuals) in India found machine learning models predicting screening uptake with >70% accuracy. The study reflects the prospective future of ML to tailor public health programs and improve screening uptake. In another study, weighted RF performed the best in predicting the determinants of cervical cancer screening uptake in India using NFHS 4 data, with 94% accuracy and 70% AUROC (Kolasseri and Bhimavarapu, 2025). Furthermore, different studies successfully applied feature importance analysis, like Shapley Additive Explanations (SHAP), permutation importance analysis to identify the determinants of service uptake or risk factors after ML prediction. These studies indicate the utility of these tools in public health decision making (Hashtarkhani et al., 2025; Ijaiya et al., 2025; Kolasseri and Bhimavarapu, 2025; Saha et al., 2023; Sultana et al., 2025). Access to mammography, educational level, affordability, health visits, and proximity to a health center are likely to explain the most variance in cancer screening uptake, as per the findings (Arage et al., 2025; Hashtarkhani et al., 2025; Kolasseri and Bhimavarapu, 2025).

Given this backdrop, current research seeks to uncover the predictors, extent, and nature of inequity in access to breast cancer screening in India, using the nationally representative National Family Health Survey (Round 5) (An excerpt of literature review matrix is presented in Table 1). Our study brings novelty in predicting and exploring the determinants and performing decomposition of CI from population-based survey data, applying ML for more accurate prediction and feature identification for both the breast cancer screening uptake and inequality in uptake. Since the data were not collected through a breast cancer–specific intervention or a randomized controlled trial, deriving insights with higher accuracy using conventional modeling techniques is challenging. Our study is the first attempt to explore this direction with respect to breast cancer screening uptake among 30–49-year-old Indian women using DHS data. This study has a triad of aims: (1) prediction of screening uptake using ML models incorporating basic sociodemographic and accessibility factors, (2) estimation of CI as a measure of inequities by different rank variables using conventional methods and ML, and (3) prediction of structural determinants contributing to inequality in uptake through ML-based CI decomposition.

NFHS is cross-sectional, nationally representative, and conducted every 5 years under the Demographic and Health Survey (DHS) program, which collects information on demographic, reproductive, maternal, and child health, nutrition, household characteristics, and gender-based violence using structured questionnaires. The NFHS-5 surveyed 724,115 women aged 15–49 years, yet in line with ICMR’s cancer screening policy, this evaluation is limited to 68,526 participants, including women aged 30–49 years, mirroring the approach in similar population studies (Rahman et al., 2025). According to the Indian Government’s operational framework, CBE is recommended for all women between 30 and 65 years of age at five-year intervals (Ministry of Health & Family Welfare Government of India, 2023). The reason for excluding women under 30 lies in the fact that the incidence of breast cancer before 30 years of age is low, and screening of younger women with clinical breast exam or imaging is not validated or cost-effective in the context of the Indian population (Khanna et al., 2024). This study follows the CLAIM reporting checklist for AI/machine learning in health research. A completed checklist is provided in Supplementary Table 1.

Our primary objective was to achieve higher accuracy and explainability in predicting breast cancer screening uptake and capturing nonlinearity in inequality decomposition, not causal inference. In our analysis, we explored predictors of breast cancer screening uptake and inequality in uptake across India among middle-aged women aged 30–49 years using NFHS-5 data, leveraging advanced machine learning approaches such as SHAP decomposition to inform health policy aimed at increasing breast cancer screening uptake in India.

Breast cancer screening uptake markedly varies by economic status, educational attainment, and social position. Wealth and higher education remained positively associated with breast cancer screening. Screening rates of 0.75% were observed among wealthier women, compared with 0.31% in poorer ones. Screening rates of 1.2% in highly educated women, in contrast to 0.24% in those with no education. This affirms their reliability as predictors rather than chance associations consistent with national findings by Negi and Nambiar (2021) and Rahman et al. (2025). Like in Nepal, the uptake of breast cancer screening was found to be associated with occupational status, ethnicity, education, and prior information on breast cancer screening. In their study, lower odds were observed among agricultural laborers (Ref. non-working; aOR = 0.59, [95% CI: 0.42, 0.82]), higher odds among Dalit women (Ref. Brahmin/Chhetri; aOR = 2.08, [95%CI: 1.37, 3.16]), and educated women (Ref. No education; Basic: aOR = 1.49, [95%CI: 1.04, 2.13], Secondary: aOR = 1.96, [95%CI: 1.33, 2.88], Higher education: aOR = 2.80, [95% CI: 1.51, 5.19]) (Lamichhane et al., 2024).

The findings of our study demonstrate that economic and educational inequalities are similarly oriented and are explained by the approachability/acceptability (information via CHWs, female provider preference, hesitancy to seek care unaccompanied) and availability/affordability (proximity, transport, ability to pay) factors. Major accessibility barriers were identified from the χ² test of independence on the observed data, as well as permutation importance and SHAP after predictive analysis. These factors were a lack of access to service information, the disempowered status of women, hesitancy in unaccompanied care seeking, lack of proximity or transport facilities, financial constraints, and perceived quality of care. In other words, findings, like women reporting no accessibility barriers considered in the framework, reveal higher screening rates. This aligns with previous evidence synthesis and observational studies that identify similar supply- and demand-side barriers across LMICs, particularly in India (Almohammed, 2024; Grosse Frie et al., 2013; Mahalakshmi and Suresh, 2020).

CI estimates reveal substantial pro-rich, pro-educated screening uptake, indicating that economic and educational deprivation shape the uptake. Conversely, relatively higher screening uptake among the socially disadvantaged subgroup might reflect targeted efforts aimed at caste-based disparities. Availability and accommodation, acceptability, approachability, quality of care, and spatial access-related barriers steepen the economic and educational gradients. Our finding is also supported by the literature review by Almohammed (2024), which revealed a recurring pattern of obstacles related to awareness, cultural barriers, unavailability of female staff, transport, and cost concerns (Almohammed, 2024). Targeted interventions to enhance affordability, acceptability, availability, and accommodations might be favoring screening uptake among scheduled castes and tribes; however, the decomposition reveals a negligible contribution of the factors considered in the theoretical framework. It demands further detailed investigation related to caste-based inequality predictors in the future.

The results of SHAP analysis reveal that spatial and structural determinants—predominantly state of residence (0.110), rural–urban location (0.100), and autonomy in healthcare seeking (0.064)—are the most influential predictors of breast cancer screening uptake. This emphasizes the role of subnational disparities and women’s autonomy in seeking preventive health services. This is similar to the findings in Patil et al. (2023) and Negi and Nambiar (2021), who indicated that subnational disparities, infrastructural limitations, constrained decision making power, and financial access critically shape preventive health behaviors in India (Negi and Nambiar, 2021; Patil et al., 2023). Our model also reveals that interaction with community health workers (CHWs) (SHAP = 0.053) and less autonomy in seeking unaccompanied care (SHAP = 0.039) exert comparable effects, emphasizing the role of informational and psychosocial empowerment of women in bridging gaps in screening uptake. Our finding is consistent with Rahman et al. (2025), who substantiated education (0.056) and socioeconomic status (0.034) as fundamental drivers (Rahman et al., 2025).

In a comparative analysis of traditional and ML-based inequality estimates, the ML-based concentration indices revealed a better comprehension of inequality patterns than linear techniques did. The Ordinary Least Squares-based concentration indices showed strong pro-rich (0.168) and pro-educated (0.261) inequality in breast cancer screening uptake, whereas the Decision Tree–based concentration indices were comparatively lower (0.108 and 0.182, respectively). This observed difference suggests that a significant portion of the socioeconomic disparities is mediated through the adjustable interdependencies among systemic factors such as autonomy, affordability, and proximity. These findings are in accordance with those of O'Donnell et al. (2008) and O'Donnell et al. (2016), who suggest caution in using linear CIs, which, by ignoring non-linear and interaction effects, may overestimate inequality (O'Donnell et al., 2008; O'Donnell et al., 2016). Similar considerations were documented in ML-based inequality research in maternal health in sub-Saharan Africa, underscoring the higher generalizability of ML-based decomposition in equity analysis (Zegeye et al., 2025).

This understanding is further enhanced by the DT-based application of SHAP decomposition to CIs as found in Candio (2024) and Zhu et al. (2024). Rural location (−0.009), limited autonomy (−0.010), and distance to facilities (−0.005) depict pro-poor CI features as well as negative contributions, indicating that higher targeting of poor, disempowered, rural women living in hard-to-reach areas is likely to reduce inequality. On the other hand, education and economic status were the primary positive contributors to the disparity in economic and educational rankings, and CI features are pro-rich and pro-educated. Additionally, disparity can be reduced by enablers such as CHW interaction (−0.001, −0.0003) and empowerment (−0.010 and −0.012) when economic status or education was used as the ranking criterion. However, access to water (0.003, 0.002) and sanitation (0.002, 0.001). and CHW engagement (0.003, 0.005) depicting pro-richness, indicating poor families are still deprived of these services. This indicates that community outreach and empowerment alleviate the structural burden of social hierarchy if properly targeted at poorer and less educated women, supporting the findings of similar studies (Mainaz et al., 2024; Palaniraja et al., 2025). In a nutshell, feature-specific CI contributions show that the burden of limited proximity, transport, hesitancy barriers, affordability barriers, and reduced autonomy falls disproportionately on disadvantaged women. These patterns were consistent across economic and education-based subgroups. However, when inequality was ranked by ethnicity, the same factors indicated the need for further in-depth research. This reflects a mixed empirical pattern, and this study recommends further research to identify the mediators and moderators under access barriers that align strongly with caste divisions. Overall, these observations confirm the robustness of directionality even when subgroup CI magnitudes varied.

Our findings contrast with some previous research that predominantly considered wealth or education as the primary contributors to inequality (Rahman et al., 2025; Sen et al., 2022). Although these are powerful predictors, the current analysis shows that inequalities are multidimensional and mediated by various structural and behavioral aspects. The lower CI magnitudes in ML-based models suggest that, rather than being solely due to economic disparities, some of the “wealth effect” observed in non-linear analysis really represents fundamental barriers to empowerment and accessibility. This question challenges the oversimplified narrative of socioeconomic theories. It emphasizes the need to incorporate psychosocial and structural factors into frameworks that account for basic social, economic, and educational equity gradients in a nonlinear manner. Therefore, as evident from the study of Chandar et al. (2025), the present study shows enhanced prediction of uptake/non-uptake of breast cancer screening programs in rural India, identifying those less likely to attend the programs, which can enhance CHW’s role in management to alter the screening preference of at-risk middle-aged women through positive deviance.

This study demonstrated that machine learning–driven prediction can efficiently identify predictors of breast cancer screening uptake and inequalities in uptake in India. Economic, educational, and social gradients shape screening participation, while accessibility, autonomy, CHW contact and service discussions, and hesitancy factors act as crucial mediators of inequity. The ML-based CI decomposition analysis highlights that financial and spatial barriers disproportionately affect poorer and less educated women, while targeting through informational and psychosocial enablers, e.g., CHW-led discussions and empowering them to make decisions, can alleviate disparities. The methodological integration of explainable AI and the concentration index, along with learner-based CI decomposition, brought methodological advancements. It generated more accurate, precise, and actionable insights into complex inequity structures that impede preventive care uptake. Findings underline the demanding need for multisectoral involvement to mitigate affordability, proximity, and empowerment gaps at the community level. Rebuilding awareness through data-driven targeting can eradicate inequities and advance equitable cancer prevention and control in resource-scarce settings.