A machine-learning decision tree using LDL-C, CHD status, family history and age identifies clusters with up to 91.1% prevalence of elevated Lp(a) (Int J Cardiol 2025)

Background: Decision tree algorithms, obtained by machine learning, provide clusters of patients with similar clinical patterns by the identification of variables that best merge with a given dependent variable.

Methods: We performed a multicenter registry, with 7 hospitals form Spain, of patients with, or high-risk of having, coronary heart disease (CHD). Elevated Lp(a) was defined as >50 mg/dl. Machine learning based decision trees were obtained by Chi-square automatic interaction detection.

Results: We analyzed 2301 patients. Median Lp(a) was 26.7 (9.3-79.9) mg/dl and 887 (38.6 %) patients had Lp(a) >50 mg/dl. The machine learning algorithm identified 6 clusters based on LDLc, CHD, FH of premature CHD and age (Fig. 1). Clusters 1 (LDLc <100 mg/dl, no CHD and, no FH of CHD) and 3 (LDLc <100 mg/dl, CHD and, no FH and, age < 50 yo) had the lowest Lp(a) values (Fig. 2); patients classified in cluster 5 (LDLc >100 mg/dl, CHD and, FH of CHD) and 6 (LDLc >100 mg/dl) had the highest values. We collapsed clusters in 3 groups: group 1 with clusters 1 and 3; group 2 with clusters 2 and 4; group 3 with clusters 5 and 6. The 3 groups have significantly different (p < 0.001) and progressively higher Lp(a) values. The prevalence of Lp(a) >50 mg/dl was 15.4 % in group 1, 29.2 % in group 2 and 91.1 % in group 3; similarly, the prevalence of Lp(a) >180 mg/dl was 1.0 %, 3.0 % and 7.6 % respectively.

Conclusions: A decision tree algorithm, performed by machine learning, identified patients with, or at high risk of having, CHD have higher probabilities of having elevated Lp(a).