AI Reveals That What Predicts Death From Fatty Liver Disease Differs Sharply Between Women and Men

Metabolic dysfunction-associated steatotic liver disease, better known as MASLD and formerly called non-alcoholic fatty liver disease, has quietly become the most common chronic liver condition on the planet. It begins with fat accumulating in the liver, often alongside obesity, diabetes, or high blood pressure, and for many people it progresses silently for years before announcing itself through cirrhosis, cardiovascular events, or premature death. Now a large Canadian study has added a striking twist to the story: the factors that best predict who will die from MASLD are not the same for women and men, and the divergence is not just biological. For women, social circumstances such as income and education appear to carry as much predictive weight as clinical measurements, while for men the risk profile is dominated by metabolic markers like waist circumference and blood pressure.

The findings come from a team led by Cindy Wen of Queen’s University in Kingston, Ontario, together with collaborators at McGill University, the University of British Columbia, the University of Calgary, Toronto Metropolitan University, and the University Health Network in Toronto. Their work, published in the journal Biology of Sex Differences, harnessed the Canadian Longitudinal Study on Aging, one of the largest and most detailed population cohorts in the world. Out of 30,097 participants in the dataset, the researchers identified 8,429 adults who met the criteria for MASLD: evidence of hepatic steatosis, at least one cardiometabolic risk factor, and alcohol consumption below sex-specific thresholds of 140 grams per week for women and 210 grams per week for men. Only about a third of the MASLD group, 35.3 percent, were female, reflecting the known epidemiology of the disease.

Over a median follow-up of 7.67 years, 631 of those participants died, giving the team enough events to model mortality risk with real statistical power. But rather than building a single one-size-fits-all prognostic model, as most clinical risk calculators do, the researchers trained separate machine learning models for women and men, and then subdivided each sex into middle-aged adults between 45 and 64 years and older adults aged 65 and above. The goal was to capture how the architecture of risk changes not only between the sexes but across the life course, a dimension that conventional risk scores routinely ignore.

The technical machinery behind the study is worth unpacking, because it illustrates how modern prognostic modeling differs from the regression-based risk scores familiar from cardiology clinics. The team considered 25 candidate predictors spanning clinical, sociodemographic, and lifestyle domains, chosen on the basis of prior literature, expert input, and data availability. These ranged from upstream social determinants of health such as income and education, through intermediate factors like physical activity and diet, to surrogate clinical markers including albumin, blood pressure, body mass index, waist circumference, and markers of cardiometabolic multimorbidity. The cohorts were split 75 percent for training and 25 percent for testing, with five-fold cross-validation used to tune hyperparameters, a standard safeguard against overfitting that ensures the model’s performance is not an artifact of memorizing the training data.

Three different survival-learning algorithms were compared, and a random survival forest emerged as the strongest performer. Random survival forests extend the logic of decision-tree ensembles to time-to-event data: hundreds of trees are grown on bootstrapped samples of the cohort, each split chosen to maximize differences in survival between branches, and the ensemble aggregates their risk estimates. Performance was evaluated in the held-out test set using several complementary metrics. The concordance index, which measures how well the model ranks pairs of individuals by risk, reached 0.73 for women and 0.79 for men. The integrated Brier score, which penalizes both discrimination and calibration errors across the follow-up period, came in at 0.032 for women and 0.035 for men, while the mean time-dependent area under the curve was 0.78 and 0.82 respectively. In plain terms, the models meaningfully stratified people with MASLD into higher- and lower-risk groups in both sexes, with somewhat sharper discrimination among men.

The most provocative results came from the interpretability analysis. To open the black box, the team applied SHAP values, or Shapley additive explanations, a technique borrowed from cooperative game theory that assigns each predictor a contribution to each individual’s predicted risk, averaged across the cohort to reveal global importance and directionality. For women, particularly those in middle age, the risk landscape was strikingly broad. Alongside clinical variables such as serum albumin, a marker of liver synthetic function and overall physiological reserve, and the burden of coexisting cardiometabolic disease, the models flagged income, education, and lifestyle behaviors as important drivers of mortality risk. For men, the picture was narrower and more conventionally metabolic: waist circumference, blood pressure, albumin, and body mass index dominated the rankings.

That asymmetry carries real biological and social plausibility. Albumin levels reflect hepatic synthetic capacity and are a well-established prognostic marker in chronic liver disease, so its prominence in both sexes is reassuring from a clinical standpoint. But the strong showing of socioeconomic variables among women echoes a broader literature linking economic precarity and lower educational attainment to worse health outcomes, potentially through pathways such as reduced access to care, chronic stress, food insecurity, and fewer opportunities for physical activity. The fact that these social determinants mattered most for middle-aged women suggests a window of vulnerability during which social circumstances may compound the metabolic burden of MASLD, a hypothesis the authors argue deserves targeted intervention research.

The study is not without caveats, and the authors are careful about what their models can and cannot claim. The Canadian Longitudinal Study on Aging captures adults aged 45 and older, so the findings may not generalize to younger populations. MASLD was defined using non-invasive criteria rather than liver biopsy, the traditional gold standard, which is standard practice in large cohorts but introduces some misclassification. Missing data were handled under assumptions of missingness at random, and the observational design means the models identify associations and risk stratification, not causal effects that can be modified by policy. The authors also note that the incremental predictive utility of adding social determinants to existing clinical models, and the practical barriers to implementing such models in clinics, remain open questions for future research.

Even so, the implications are considerable. Prognostic models for MASLD in current use are largely agnostic to sex and blind to the social context of patients’ lives. This study demonstrates that both omissions come at a cost in accuracy and equity. If the variables that best identify a middle-aged woman at high risk of death include her income and education, then risk prediction becomes inseparable from social policy, and the path to prevention may run through resources that have nothing to do with a prescription pad. For men, the dominance of modifiable metabolic markers suggests that weight, blood pressure, and central adiposity remain the highest-yield targets. The authors frame their work as a bridge between social medicine and precision medicine, an approach that could make MASLD risk prediction more individualized and more equitable at the same time. As machine learning tools migrate from research datasets into clinical dashboards, the lesson from this Canadian cohort is clear: who you are, in both body and circumstance, shapes how your liver disease will unfold, and our predictive models are finally learning to see that.

Subject of Research: Sex-specific machine learning prediction of all-cause mortality in metabolic dysfunction-associated steatotic liver disease using the Canadian Longitudinal Study on Aging

Article Title: Sex differences in mortality risk profiles across the life course in metabolic dysfunction-associated steatotic liver disease: a machine learning analysis of the Canadian Longitudinal Study on Aging

Article References: Wen, C., Tu, W., Sebastiani, G., Burnside, J., Rapino, C., Ramji, A., Swain, M. G., Patel, K., Moodie, E. E. M., Flemming, J. A., & Saeed, S. (2026). Sex differences in mortality risk profiles across the life course in metabolic dysfunction-associated steatotic liver disease: a machine learning analysis of the Canadian Longitudinal Study on Aging. Biology of Sex Differences. https://doi.org/10.1186/s13293-026-00992-9

Image Credits: AI Generated

DOI: 10.1186/s13293-026-00992-9

Keywords: MASLD, fatty liver disease, sex differences, machine learning, mortality prediction, social determinants of health, Canadian Longitudinal Study on Aging, random survival forest, SHAP values, precision medicine, cardiometabolic risk, health equity