A high-risk multimorbidity post-MI phenotype (Cluster 2) was associated with significantly longer hospital stays (median 11 vs 7 days) compared to other machine learning-derived phenotypes.
Observational (n=1,460)
Does unsupervised machine-learning clustering identify distinct phenotypes of acute myocardial infarction that predict length of hospital stay?
Unsupervised machine learning can identify distinct clinical phenotypes in acute myocardial infarction patients, highlighting a multimorbid subgroup at high risk for prolonged hospitalization.
Absolute Event Rate: 11% vs 7%
p-value: p=2.24×10−¹4 to 8.74×10−²6
Abstract Background Acute myocardial infarction (MI) includes a wide range of patients. This diversity makes early risk assessment more difficult. Machine learning–based clustering may support early identification of patients at risk of prolonged hospitalisation. Purpose To derive data-driven baseline phenotypes in MI patients using clinical, laboratory, and angiographic variables, and to describe their key characteristics. Methods A total of 1,460 patients were included. Thirty-seven variables encompassed demographic data (age, sex), body mass index (BMI), comorbidities (diabetes, hypertension, chronic kidney disease, prior MI/PCI, chronic obstructive pulmonary disease, heart failure, atrial fibrillation), laboratory markers (glucose, low-density lipoprotein, triglycerides, C-reactive protein, N-terminal pro-B-type natriuretic peptide NT-proBNP, haemoglobin, leukocytes, platelets, creatinine, aspartate aminotransferase AST, alanine aminotransferase ALT) and angiographic features (vessels treated, stents implanted). Factorial Analysis of Mixed Data (FAMD) was applied. The optimal clustering solution (K=5) was selected based on silhouette width and clinical interpretability. Between-cluster differences were assessed using ANOVA/Kruskal–Wallis tests or Chi-square/Fisher tests. Continuous variables are reported as median IQR. Results FAMD followed by k-means clustering identified five distinct clinical phenotypes (silhouette = 0.29). Cluster 1 (C1, n=161) included patients with preserved EF (50% 39–55) and higher NT-proBNP (1319 411–4089 pg/mL). Cluster 2 (C2, n=205) represented the most adverse profile, with the oldest age (73 68–80 years), highest multimorbidity (4 comorbidities), reduced EF (38% 30–45) and very high NT-proBNP (4377 2002–11611 pg/mL), alongside frequent diabetes (70%), CKD (41%) and prior heart failure (77%). Cluster 3 (C3, n=126) showed an intermediate-risk profile with preserved EF (50% 36–60) and the lowest metabolic (glucose 119 103–151 mg/dL) and inflammatory activity (CRP 3.0 1–13 mg/L). Cluster 4 (C4, n=284) was characterised by extensive coronary disease and procedural history (CAD 93%, prior MI 86%, prior PCI 88%) with moderately impaired EF (48% 40–55). Cluster 5 (C5, n=684), the largest and youngest group (63 55–69 years), showed the lowest comorbidity burden (1), highest EF (52% 46–56), lowest NT-proBNP (668 211–1616 pg/mL) and minimal history of prior MI (5%) or PCI (1%), despite higher LDL (115 90–142 mg/dL). When comparing length of hospitalisation, patients in cluster 2 had significantly longer hospitalisations (11 8–14 vs. 7 days in all other clusters), with pairwise p-values of 2.24×10−¹4 (vs C1), 5.60×10−¹4 (vs C3), 8.74×10−²6 (vs C4) and 7.49×10−²³ (vs C5). Conclusions Unsupervised FAMD-based clustering identified five meaningful post-MI phenotypes, including a high-risk multimorbidity subgroup at increased risk of prolonged hospital stay.For image description, please refer to the figure legend and surrounding text.
Szymanska-Lyczkowska et al. (Mon,) conducted a observational in Acute myocardial infarction (n=1,460). High-risk multimorbidity phenotype (Cluster 2) vs. Other clinical phenotypes (Clusters 1, 3, 4, 5) was evaluated on Length of hospitalisation (p=2.24×10−¹4 to 8.74×10−²6). A high-risk multimorbidity post-MI phenotype (Cluster 2) was associated with significantly longer hospital stays (median 11 vs 7 days) compared to other machine learning-derived phenotypes.