Machine Learning Reveals Stark Vulnerability Strata in Cameroon’s Under-Five Mortality

Child survival in Cameroon is not a matter of chance distributed evenly across the population, according to a new study that applies machine learning to two decades of demographic survey data. Researchers Aoudou Njingouo Mounchingam and Vissého Adjiwanou, publishing in SSM – Population Health, have shown that the country’s under-five mortality is organized into sharply differentiated empirical strata of vulnerability, with children in the highest-risk group facing a cumulative probability of dying before age five of 27.8 percent, compared with just 2.1 percent among children in the lowest-risk stratum. The gap, more than thirteenfold, is among the starkest illustrations yet of how social disadvantage accumulates to shape whether a child lives or dies in early childhood.

The study draws on pooled data from the 2011 Demographic and Health Survey-Multiple Indicator Cluster Survey and the 2018 Demographic and Health Survey, two nationally representative household surveys conducted by Cameroon’s National Institute of Statistics in partnership with international agencies. Together, the surveys cover 21,465 children born in the five years preceding each wave, of whom 1,646, or 7.46 percent, died before reaching their fifth birthday. Because each survey records complete birth histories, the researchers were able to construct time-to-event measures of child survival, treating children who died as events and those still alive at the time of interview as right-censored observations.

What sets the analysis apart is its methodological approach. Most research on child mortality in sub-Saharan Africa relies on regression models that estimate the average effect of each individual predictor, such as maternal education or birth interval, while holding other factors constant. The authors argue that such variable-centered strategies obscure the reality that vulnerability rarely stems from a single adverse characteristic. Instead, mortality risk emerges from the accumulation of reproductive constraints, limited parental resources, restricted access to health services, and adverse community environments that cluster together within particular households and communities.

To capture these configurations, the researchers turned to a random survival forest, a machine learning method that extends tree-based ensembles to right-censored survival data. Unlike the conventional Cox proportional hazards model, which assumes that relative risks between individuals remain constant over time and that covariate effects are linear and additive, the random survival forest makes no such assumptions. The choice was not merely stylistic. Tests based on Schoenfeld residuals revealed widespread and statistically significant violations of the proportional hazards assumption in the Cox benchmark model, with a global test statistic of 172.35 and a p-value below 0.001, indicating that the relative contribution of risk factors changes substantially over the course of early childhood.

The random survival forest produced a continuous risk score for each child, summarizing the estimated probability of death before 59 months. When the researchers ranked children along this gradient and partitioned them into strata, three distinct groups emerged. Children in the low-risk stratum, slightly more than half of the sample, experienced a cumulative mortality of just 2.1 percent. The intermediate-risk stratum, representing roughly a quarter of children, faced a mortality of 6.3 percent. The high-risk stratum, also about a quarter of the sample, accounted for a disproportionate share of deaths, with 1,210 of the 1,646 observed deaths and a cumulative mortality of 27.8 percent. Kaplan-Meier survival curves confirmed a clear and monotonic separation of trajectories, with differences emerging within the first months of life and persisting throughout the first five years.

The social composition of each stratum proved internally coherent and revealing. Children in the low-risk stratum were predominantly embedded in advantaged contexts, characterized by overrepresentation of higher maternal and paternal education, greater maternal media exposure, wealthier households, urban residence, higher birth weight, more antenatal care visits, and greater maternal decision-making autonomy. The intermediate-risk stratum presented a more paradoxical picture: despite favorable educational and informational resources, these children were disproportionately exposed to low birth weight, short birth intervals, young motherhood, and missed antenatal care, illustrating that social resources alone cannot fully offset early-life biological vulnerabilities.

The high-risk stratum concentrated disadvantage across every dimension the researchers measured. Low birth weight appeared at markedly higher levels than in the intermediate group, alongside strong overrepresentation of short preceding birth intervals reflecting closely spaced fertility. Both mothers and partners in this group were more likely to have no formal education, mothers were disproportionately likely to receive no antenatal care and to deliver at home rather than in health facilities, and women without any household decision-making autonomy were overrepresented. The authors describe this as a configuration of compounded vulnerability, in which biological risk factors, constrained reproductive conditions, and entrenched social disadvantage reinforce one another.

Perhaps the most sobering finding concerns temporal persistence. Comparing the 2011 and 2018 survey waves, the researchers found that although overall under-five mortality in Cameroon declined from 144 to 80 deaths per 1,000 live births between 2004 and 2018, the social structuring of vulnerability remained remarkably stable. High-risk strata in both waves were characterized by the same accumulation of reproductive, educational, and healthcare disadvantages, while low-risk strata concentrated protective resources. Progress in child survival, in other words, occurred without fundamentally altering the distribution of mortality risk across social groups, raising questions about whether aggregate improvements mask persistent inequity.

The findings carry significant implications for public health policy, both in Cameroon and across sub-Saharan Africa, a region that accounts for nearly half of the global burden of under-five mortality despite representing only about 17 percent of the world’s population. In 2023, approximately 4.8 million children worldwide died before their fifth birthday, mostly from preventable causes, and progress toward the Sustainable Development Goal target of 25 deaths per 1,000 live births by 2030 remains far off track in Cameroon, where the 2018 level stood at 80 deaths per 1,000 live births. The study suggests that interventions targeting individual risk factors in isolation may be insufficient; instead, programs should address the concentration of disadvantages within identifiable high-risk population strata through integrated maternal and child health approaches.

The authors are careful to note the limitations of their approach. The strata are empirical groupings along a continuous risk gradient, not latent classes or causal typologies, and should be interpreted as descriptive representations of cumulative vulnerability rather than distinct causal entities. The analysis also relies on retrospective birth histories, which may be subject to recall bias, particularly for early neonatal deaths that are heavily concentrated in the high-risk stratum. Nevertheless, the study demonstrates that machine learning methods, used as exploratory tools rather than black-box predictors, can make mortality heterogeneity empirically observable in ways that complement conventional regression. Future research, the authors suggest, could map the spatial distribution of vulnerability strata using survey cluster coordinates, examine transitions between strata over time with longitudinal data, and apply similar person-centered survival approaches to other health outcomes and populations across the region.

Subject of Research: Empirical vulnerability strata in under-five mortality in Cameroon identified using survival-based machine learning on Demographic and Health Survey data

Article Title: Empirical vulnerability strata in under-five mortality in Cameroon: Evidence from DHS data

Article References: Mounchingam, A. N., & Adjiwanou, V. (2026). Empirical vulnerability strata in under-five mortality in Cameroon: Evidence from DHS data. SSM – Population Health, 35, Article 101965. https://doi.org/10.1016/j.ssmph.2026.101965

Image Credits: AI Generated

DOI: 10.1016/j.ssmph.2026.101965

Keywords: under-five mortality, Cameroon, DHS data, random survival forest, machine learning, child survival, social inequality, maternal education, antenatal care, sub-Saharan Africa, survival analysis, health disparities