Six blood and age markers predict death after heart attack in stroke survivors

Patients who arrive at the hospital with a heart attack after already surviving an ischemic stroke are among the most difficult people in cardiology to assess. Their risk of dying is elevated by two overlapping vascular diseases, yet the standard bedside tools used to stratify heart attack patients were built for a single organ system and cannot capture the tangled interplay of cerebral injury, coronary occlusion, inflammation, kidney dysfunction, and poor nutrition that characterizes this population. A new prospective cohort study from Xinjiang Medical University in Urumqi, China, set out to close that gap by building and testing an interpretable machine learning model designed specifically for these patients, and its results offer both a practical risk score and a cautionary lesson about how far such models can be trusted.

The study, published in Frontiers in Cardiovascular Medicine, enrolled 543 consecutive patients who were admitted to the First Affiliated Hospital of Xinjiang Medical University with chest pain between February 2015 and June 2023, diagnosed with an acute myocardial infarction and a prior history of ischemic stroke, and underwent percutaneous coronary intervention within twelve hours of symptom onset. Of 2,368 patients screened during that period, 574 met the diagnostic criteria; 18 were excluded for reasons such as severe valvular disease, malignancy, or critical missing data, and 13 were lost to follow-up. All participants gave written informed consent, the protocol was approved by the hospital’s ethics committee, and the trial was registered in the Chinese Clinical Trial Registry.

The cohort was predominantly male, with 447 of 543 patients (82.3%) being men, and the mean age was 61.0 years. Blood samples were drawn within twelve hours of symptom onset and before any revascularization or medical treatment, which allowed the researchers to capture the patients’ baseline physiological state rather than a treatment-altered one. Over a median follow-up of 3.03 years, 66 patients died, corresponding to a mortality rate of 12.2%. Non-survivors were older on average (69.8 versus 59.8 years) and showed markedly lower hemoglobin, lower red blood cell counts, lower lymphocyte percentages, and lower prognostic nutritional index scores than survivors, alongside elevated inflammatory markers such as interleukin-6 and procalcitonin.

To decide which variables deserved a place in the final model, the researchers used a deliberately conservative hybrid strategy. Univariable Cox regression, the Boruta algorithm (a random forest-based feature selection method that compares real variables against permuted shadow features), and LASSO Cox regression were each applied independently, and only variables identified by all three methods were retained. That convergence produced six predictors: age, hemoglobin, the prognostic nutritional index (a composite of serum albumin and lymphocyte count), estimated glomerular filtration rate, the atherogenic index of plasma (a logarithmic ratio of triglycerides to high-density lipoprotein cholesterol), and the monocyte-to-lymphocyte ratio. Missing data were handled through multiple imputation by chained equations, with imputation models re-estimated separately within each bootstrap training sample to prevent information leakage into the validation step.

Six survival models were then built and compared: a standard Cox proportional hazards model, an elastic-net penalized Cox model, a random survival forest, a gradient boosting machine, an XGBoost survival model, and a survival support vector machine. Internal validation used 100 bootstrap resamples with out-of-bag evaluation, meaning each model was tested on the patients excluded from its own training resample. The random survival forest achieved the numerically highest concordance index at 0.720, but its calibration was poor, with a calibration slope of just 0.445, indicating that its predicted probabilities did not track observed mortality well. The elastic-net Cox model, by contrast, achieved a nearly identical C-index of 0.714 (95% confidence interval 0.616 to 0.790) while showing the best calibration (slope 0.798) and the lowest three-year Brier score of 0.096. Given that the two models were statistically indistinguishable in discrimination, the researchers selected the elastic-net Cox model as the final one on the grounds of better calibration, lower prediction error, and greater interpretability.

The final model’s time-dependent discrimination, weighted for censoring, was 0.745 at one year, 0.704 at two years, and 0.701 at three years. A three-year calibration curve showed reasonable agreement between predicted and observed mortality across the range of risks represented in the cohort, although uncertainty was relatively large in some risk groups. A decision curve analysis suggested that the model would deliver greater net clinical benefit than treating everyone or treating no one across part of the clinically relevant threshold range, most clearly between roughly 5% and 20% risk thresholds. The apparent C-index of 0.732 was close to the bootstrap-validated figure of 0.714, suggesting limited overfitting despite the modest number of deaths.

To understand what the model was actually doing, the team applied Kernel SHAP (SHapley Additive exPlanations) analysis, a technique that quantifies each variable’s marginal contribution to individual predictions. Age emerged as the largest overall contributor, followed by hemoglobin, the prognostic nutritional index, estimated glomerular filtration rate, the atherogenic index of plasma, and the monocyte-to-lymphocyte ratio. At the three-year horizon, age and hemoglobin alone accounted for 27.6% and 24.1% of the total mean absolute SurvSHAP(t) values, respectively. Older age and higher monocyte-to-lymphocyte ratios pushed predictions toward higher mortality risk, while higher hemoglobin, nutritional index, and kidney function values pushed predictions toward lower risk. Time-dependent permutation analysis confirmed that age and hemoglobin contributed most strongly to discrimination, each reducing the time-dependent AUC by roughly 0.04 when permuted.

One finding demanded particular caution. The atherogenic index of plasma is conventionally treated as a marker of an atherogenic lipid profile, yet in this cohort higher values were predominantly associated with lower model-predicted mortality risk, a direction consistent with the lower AIP levels observed among non-survivors at baseline. The authors note that a previous study of stroke patients reported a similar inverse association, sometimes described as a lipid paradox, and they list plausible explanations including acute-phase changes in lipid concentrations, prior lipid-lowering treatment, nutritional status, residual confounding, and correlations among the predictors themselves. Because AIP showed relatively low permutation importance, the researchers explicitly caution against interpreting this inverse pattern as evidence that higher AIP independently protects against death.

When the final model was benchmarked against the GRACE risk score, the standard tool for acute coronary syndrome prognosis, the new model achieved a C-index of 0.714 versus 0.648 for GRACE, and a three-year time-dependent AUC of 0.701 versus 0.608. The paired differences were 0.065 and 0.093 respectively, but their confidence intervals crossed zero, meaning statistical superiority in discrimination was not demonstrated. Calibration, however, favored the new model, with a three-year calibration slope of 0.931 versus 0.676 for GRACE. An exploratory incremental-value analysis, adding the new model’s risk score to GRACE, yielded a three-year continuous net reclassification improvement of 0.558 and an integrated discrimination improvement of 0.049, though these figures were computed in the full development cohort and require external confirmation.

The authors are candid about the study’s limits. It was a single-center study with only 66 deaths, which produced wide confidence intervals and raised the risk of overfitting, particularly for the more flexible machine learning algorithms. Bootstrap resampling, while reducing optimism, does not create new independent information or substitute for validation in a separate population. Stroke severity was not measured with instruments such as the NIHSS or modified Rankin Scale, and information on medication adherence, rehabilitation, and post-discharge management was incomplete. The model has also not been tested in real-time clinical practice. Independent external validation and prospective implementation studies are required before the six-variable score can be used at the bedside, but the work demonstrates that a small, interpretable panel of routinely collected blood markers can meaningfully stratify mortality risk in a population that conventional scores were never designed to serve.

Subject of Research: Machine learning prediction of all-cause mortality in acute myocardial infarction patients with prior ischemic stroke

Article Title: Mortality prediction in patients with acute myocardial infarction and prior ischemic stroke: a prospective cohort study with explainable machine learning and SHAP analysis

Article References: Han, C., Muladili, A., Song, N., Shan, C., Liu, F., Li, X., Yang, Y., & Zhao, Q. (2026). Mortality prediction in patients with acute myocardial infarction and prior ischemic stroke: a prospective cohort study with explainable machine learning and SHAP analysis. Frontiers in Cardiovascular Medicine, 13, 1949226. https://doi.org/10.3389/fcvm.2026.1949226

Image Credits: AI Generated

DOI: 10.3389/fcvm.2026.1949226

Keywords: acute myocardial infarction, ischemic stroke, machine learning, elastic-net Cox model, SHAP analysis, mortality prediction, percutaneous coronary intervention, prognostic nutritional index, GRACE score, prospective cohort, risk stratification, cardiovascular medicine