Phenotyping for novel classification of adults with Fontan circulation by machine learning
M Nakashima, N Toh, T Seki, A Saito, K Soma, K Saito, Y Hyodo, K Ejiri, Y Takaya, S Akagi, T Akagi, M Hatano, N Takeda, M Lachmann, S YuasaAbstract
Background
Adults with Fontan circulation show various clinical features and their heterogeneous clinical presentations make management strategies difficult to establish.
Purpose
We aimed to clarify potential phenotypes among adults with Fontan circulation and to develop a simple model applicable to clinical settings using machine learning.
Methods
This was a multicentre study, with derivation set: of a university hospital and validation set: of another university hospital. Adults with Fontan circulation were enrolled. Patients without cardiac catheterization or laboratory tests were excluded. We conducted an agglomerative hierarchical clustering analysis using characteristics at the first adult congenital heart disease centre visiting. Variables obtained in routine clinical practice with missing value ratio <0.3 were put up. Missing values were imputed using a random forest algorithm. Criteria of selecting variables were: 1) significant representation within the top five dimension in principal component analysis; and 2) no collinearity with other variables by Pearson correlation coefficient. Elbow method, Euclidean distance metric, and Ward's minimum variance method were applied. We constructed a decision tree predicting the clusters in derivation set using recursive partitioning analysis by 5-fold cross-validation. We validated the decision tree using validation set.
Results
Derivation set contained 104 patients (25±9 years, 53.8% of men). We selected six variables and determined the ideal number of clusters as three (Figure 1). Cluster 1 (n=13) had the worst aspartate aminotransferase and serum albumin, labelled as “Failing Fontan”. Cluster 2 (n=50) had similar characteristics to Cluster 3 but had relatively lower haemoglobin (almost within normal lesion of 14.3±1.6 g/dL but significantly lower than that in Cluster 3 of 17.0±1.2 g/dL, p<0.001), labelled as “Relative anaemia”. Cluster 3 (n=41) had favourable clinical parameters, labelled as “Favourable condition”. During median 4.7 years, the composite outcomes of all-cause deaths and hospitalization for heart failure occurred in 10 patients in derivation set: 7 in Cluster 1 (53.8%), 3 in Cluster 2 (6.0%) and no patient in Cluster 3 (0.0%), (log-rank test, p=0.012). A decision tree was constructed in derivation set (Figure 2). The validation set contained 106 patients (23±6 years, 60.4% of men). The characteristics of clusters by decision tree in validation set were similar to those of clusters by hierarchical clustering in derivation set. During median 3.5 years, the composite outcomes were observed in 14 patients: 5 in Cluster 1 (35.7%), 7 in Cluster 2 (14.3%) and 2 in Cluster 3 (4.7%), (log-rank test, p=0.025), which resembled those in derivation set.
Conclusions
Machine learning analyses classified adults with Fontan circulation into three potential phenotypes and predicted them in a simple model.Figure 1Figure 2