Heart failure (HF) is a complex clinical syndrome resulting from structural or functional cardiac abnormalities that impair ventricular filling or ejection. Key clinical manifestations include dyspnea, fatigue, and fluid retention (e.g., pulmonary and systemic congestion, peripheral edema). As a major global public health challenge, HF affects over 64 million individuals worldwide, with an estimated prevalence of 1–2% ^{[1, 2]}. The condition imposes a substantial economic burden on healthcare systems, driven by high rates of morbidity, hospital readmissions, and mortality, as well as significant impairments in functional capacity and health-related quality of life ^[3]. In developed countries, HF accounts for an estimated 1–2% of total annual healthcare expenditures ^[4]. Indirect costs—such as productivity loss, early retirement, and informal caregiving—further amplify its socioeconomic impact ^[3].

Ischemic heart disease(IHD) remains the leading cause of HF. Following acute myocardial infarction (AMI), the irreversible loss of cardiomyocytes triggers a reparative fibrotic response. This process activates neurohumoral pathways, leading to ventricular remodeling, cardiac dilation, and impaired contractility. Subsequent hemodynamic overload often results in pulmonary congestion, manifesting as dyspnea and chest tightness, and poses a serious threat to patient survival. Against the backdrop of population aging and rising prevalence of comorbidities such as obesity, both the incidence and prevalence of HF are expected to increase. Missed or delayed diagnosis can worsen clinical outcomes and elevate treatment costs, underscoring the critical importance of early screening and risk stratification. Timely identification of high-risk patients not only saves lives but also enables personalized treatment strategies that can improve quality of life ^[5].

Recent advances in artificial intelligence (AI) have created new opportunities for enhancing clinical decision-making. Machine learning (ML), a core subset of AI, leverages computational models to identify patterns in large datasets and supports tasks such as diagnosis, risk prediction, and treatment optimization ^{[6, 7]}. ML approaches can be broadly categorized into supervised, unsupervised, and reinforcement learning. In clinical settings, supervised learning algorithms—such as logistic regression, k-nearest neighbors, decision trees, support vector machines, and naive Bayes—are frequently used for classification tasks, including binary outcome prediction ^[8]. For instance, the extreme gradient boosting (XGBoost) algorithm has been employed to predict in-hospital mortality in acute HF ^[9], and random forest models have been applied to HF classification ^[10]. These methods can uncover complex, non-linear relationships in multidimensional clinical data, offering valuable insights for disease prediction and diagnosis.

In this study, we aimed to develop and compare six ML algorithms for early prediction of HF following AMI. Using routinely collected clinical data, we constructed prediction models and evaluated their performance to identify the most effective approach. We anticipate that a robust ML-based tool will assist clinicians in stratifying post-AMI risk, facilitating early intervention, and ultimately improving patient outcomes.

Inclusion criteria were as follows: (1) age ≥ 18 years; (2) initial diagnosis of AMI at admission; and (3) no evidence of HF at the time of admission.

Exclusion criteria comprised: (1) HF attributable to valvular heart disease or cardiomyopathy; (2) non-obstructive myocardial infarction; (3) concomitant rheumatic or autoimmune diseases; (4) history of malignant tumor; and (5) in-hospital mortality.

The study protocol was approved by the Ethics Committee of the Second Affiliated Hospital of Xuzhou Medical University (Approval No. 2024071501), which waived the requirement for informed consent due to the retrospective nature of the analysis.

Demographic, clinical, laboratory, and echocardiographic data were extracted from the hospital's electronic medical record system for all eligible participants and linked to a unique admission identifier. Data preprocessing involved cleaning and standardization: records with extensive missing data (> 30%) were excluded, while remaining missing values in continuous variables were imputed using the mean or median after assessing normality with the Shapiro-Wilk test. The robustness of this imputation was evaluated by comparing model performance pre- and post-imputation. Finally, all laboratory values were converted to international standard units to ensure consistency.

To develop the prediction model, a two-step feature selection process was employed. First, Least Absolute Shrinkage and Selection Operator (LASSO) regression with 5-fold cross-validation was applied to the initial 61 variables, identifying 24 non-zero coefficients as potentially significant predictors. These 24 variables were subsequently subjected to a multivariate logistic regression analysis, which further refined the feature set to 8 statistically significant indicators. These final 8 features were used as inputs for the following six machine learning algorithms: backpropagation neural network (BPNN), decision tree (DT), logistic regression (LR), random forest (RF), support vector machine (SVM), and XGBoost. The dataset was randomly partitioned into a training set (70%) for model construction and a validation set (30%) for performance evaluation. The models were trained on the training set, and their predictive performance was assessed on the validation set. Furthermore, 5-fold cross-validation was implemented during the training phase of each algorithm to optimize hyperparameters and ensure model robustness.

A total of 61 variables were initially extracted from patient baseline characteristics, initial laboratory tests, and echocardiographic examinations as potential predictors for model development. Missing quantitative data were imputed using the median value of each variable. Feature selection was performed in two stages: first, LASSO regression with 5-fold cross-validation was applied, which reduced the feature set to 24 non-redundant variables (Figs. 1 and 2). These variables included type of myocardial infarction (MI), diabetes, sex, age, diastolic blood pressure (DBP), neutrophil count (NEUT), albumin (ALB), globulin (GLB), aspartate aminotransferase (AST), alkaline phosphatase (ALP), blood urea nitrogen (BUN), serum creatinine (Scr), glucose (GLU), high-density lipoprotein cholesterol (HDLC), lipoprotein(a) (Lpa), residual cholesterol (RC), high-sensitivity C-reactive protein (hsCRP), prothrombin time (PT), high-sensitivity cardiac troponin I (hscTnI), creatine kinase-MB (CKMB), left atrial anterior–posterior diameter (LAAPD), left ventricular anterior–posterior diameter (LVAPD), left ventricular posterior wall thickness (LVPW), and right ventricular anterior–posterior diameter (RVAPD). Subsequently, logistic regression was used to further refine the predictor set, identifying the eight most clinically meaningful features: MI type, age, ALB, AST, BUN, GLU, PT, hsCRP, and LVAPD, which were used to construct the final prediction model for post-infarction heart failure (Fig. 3).

The growing global burden of AMI is closely linked to contemporary shifts in lifestyle and increasing exposure to psychosocial stressors, contributing to its rising incidence. As a leading cause of HF, AMI imposes a substantial economic and healthcare burden worldwide. This challenge is further compounded by population aging, which has been widely recognized as an independent risk factor for HF. Epidemiological studies indicate that adults over 60 years of age face a markedly elevated risk of developing HF compared to younger individuals ^[13], and aging is strongly associated with the pathogenesis and progression of the disease ^[14]. In our study, the mean age of both the HF and non-HF groups exceeded 60 years, with patients in the HF group being significantly older. Those affected by HF frequently experience recurrent hospitalizations, long-term pharmacotherapy, and severely impaired functional capacity, leading to considerable loss of productivity and diminished quality of life. These factors underscore the critical importance of early prediction of HF following AMI ^{[15, 16]}.

The widespread adoption of electronic medical record (EMR) systems in healthcare institutions has generated vast repositories of clinical data, which—while requiring rigorous information security measures—provide a valuable foundation for applying ML in clinical research. Conventional statistical methods often face limitations in handling high-dimensional and complex multimodal health data, which can restrict the predictive performance of resulting models. In contrast, ML techniques offer a powerful alternative by leveraging computational algorithms to automatically learn patterns from data without relying solely on pre-specified hypotheses.

ML models are trained by establishing mappings between input features and output labels, enabling the prediction of outcomes through classification or regression mechanisms ^[17]. These methods can capture both linear and non-linear relationships, but their performance heavily depends on appropriate feature selection and model design. Consequently, model development is an iterative, empirically driven process that involves continuous adjustment of algorithms and hyperparameters based on performance feedback ^[18].

Current diagnosis of heart failure relies on comprehensive assessment of patient symptoms, physical signs, laboratory tests, imaging findings, and medical history. In this study, we developed and validated an XGBoost-based machine learning model using routinely collected clinical data—including baseline characteristics, laboratory results, and echocardiographic parameters—to predict early-onset heart failure following AMI.

Feature importance analysis revealed that hsCRP was the most influential predictor in our model, with significantly higher levels observed in the HF group compared to the non-HF group. As a sensitive inflammatory marker, hsCRP reflects underlying chronic inflammation, which contributes to atherosclerotic plaque vulnerability and progression of coronary artery disease. Elevated hsCRP levels have been documented in patients with non-ST-segment elevation acute coronary syndrome (NSTEMI-ACS) and have shown predictive value for new-onset HF ^[19]. Similarly, elevated hsCRP in stable ACS patients has been associated with a two-fold increased risk of new or worsening HF within two years ^[20].

AST and ALB were also identified as significant predictors in our model. Elevated AST following AMI may result from oxidative stress, cardiomyocyte necrosis, or ischemia-reperfusion injury, although direct evidence linking AST to HF progression remains limited ^{[21, 22]}. Thus, AST may serve as an auxiliary predictive marker in this context. Hypoalbuminemia, frequently observed in HF patients (affecting up to 40% in some cohorts ^[23]), is associated with higher NYHA functional class, reduced renal function, and increased comorbidity burden ^[24]. Higher serum albumin levels have been correlated with improved outcomes in HF with preserved ejection fraction (HFpEF) ^[25], supporting ALB’s relevance as a prognostic indicator.

LVAPD also contributed to model performance. Following AMI, compensatory ventricular dilation may initially maintain stroke volume but can progress to adverse remodeling and functional deterioration. Previous studies have linked LVAPD to adverse outcomes in both HFpEF and HF with reduced ejection fraction (HFrEF) ^{[26, 27]}, consistent with our findings.

BUN emerged as another relevant predictor. In HF, reduced cardiac output activates neurohormonal systems such as the sympathetic nervous system and renin–angiotensin–aldosterone system, promoting renal sodium retention and elevated BUN. These pathways are associated with worse cardiac function and prognosis ^{[28, 29]}, and BUN has been consistently linked to HF outcomes ^{[30, 31]}.

Diabetes and elevated admission blood glucose are well-established risk factors for HF, with diabetic patients facing a two-fold increase in HF risk in men and up to five-fold in women after age adjustment ^[32–35]. Hyperglycemia at admission may reflect impaired myocardial energy metabolism and oxidative stress, further aggravating cardiac injury ^[36–38] and contributing to HF onset ^[39].

Finally, MI type (STEMI vs. NSTEMI) was incorporated into the model. STEMI, typically resulting from transmural ischemia, often leads to more extensive myocardial damage and higher HF incidence compared to NSTEMI ^[40]. Although MI type had the lowest feature importance in our model, it still provided complementary predictive value, consistent with prior studies ^[41].

In this study, we developed and compared six machine learning models for predicting in-hospital heart failure following acute myocardial infarction. The XGBoost algorithm demonstrated superior performance, forming the basis of a final prediction model that incorporates eight key clinical features: hs-CRP, age, AST, LVAPD, BUN, albumin, glucose, and MI type. This model exhibits strong predictive capability and clinical utility, offering a reliable, data-driven tool for early risk stratification that may support clinical decision-making and improve patient management.