library(survPhenoStab)
#> Loading required package: mclust
#> Package 'mclust' version 6.1.3
#> Type 'citation("mclust")' for citing this R package in publications.survPhenoStab implements the stability-driven
hierarchical clustering framework for multi-endpoint patient phenotyping
described in Roy Chowdhury et al., “A stability-driven hierarchical
clustering framework for multi-endpoint patient phenotyping in
oncology”. It provides:
embed_survival(): build a standardised, mixed
continuous-binary feature-space embedding from multiple time-to-event
endpoints.sim_competing_risks(): the paper’s competing-risks
data-generating mechanism, for simulation studies.cluster_configs() / cluster_assign():
enumerate and fit the distance-linkage-K grid.boot_jaccard(): bootstrap Jaccard cluster-stability
assessment.chaining_diagnostic(): the Delta_CA chaining-artefact
diagnostic.select_k(): cluster-number selection combining
silhouette, gap statistic and bootstrap stability.benchmark_methods(): comparison against k-means and
Gaussian mixtures.mc_replicate(): Monte Carlo replication of the
stability comparison across simulated datasets.Fit and assess one configuration:
ca <- boot_jaccard(Z, distance = "cosine", linkage = "average", k = 3, B = 100, seed = 2024)
ca
#> Configuration: cosine distance, average linkage, K = 3 (B = 100)
#> J_min = 0.520 (SD 0.241) [borderline]
#> 95% bootstrap percentile interval: [0.206, 0.843]
#> Cluster sizes: 132, 81, 87Compare against the conventional Ward-Euclidean default:
data(colon_survival)
Zc <- embed_survival(
time = colon_survival[, c("T_recur", "T_death")],
event = colon_survival[, c("E_recur", "E_death")]
)
ca_colon <- boot_jaccard(Zc, "cosine", "average", k = 3, B = 100, seed = 2024)
we_colon <- boot_jaccard(Zc, "euclidean", "ward.D2", k = 3, B = 100, seed = 2024)
data.frame(
method = c("cosine-average", "Ward-Euclidean"),
J_min = c(ca_colon$J_min, we_colon$J_min)
)
#> method J_min
#> 1 cosine-average 0.6035493
#> 2 Ward-Euclidean 0.7320246As in the manuscript’s real-data application, the cosine-average advantage seen in simulation does not universally transfer: Ward-Euclidean is competitive (often stronger) on this dataset, motivating per-dataset empirical selection via bootstrap Jaccard stability rather than a fixed default.
diag_single <- chaining_diagnostic(Z, "cosine", "single", k = 3, B = 100, seed = 2024)
diag_average <- chaining_diagnostic(Z, "cosine", "average", k = 3, B = 100, seed = 2024)
data.frame(
linkage = c("single", "average"),
ARI = c(diag_single$ARI, diag_average$ARI),
J_min = c(diag_single$J_min, diag_average$J_min),
delta_CA = c(diag_single$delta_CA, diag_average$delta_CA),
flagged = c(diag_single$flagged, diag_average$flagged)
)
#> linkage ARI J_min delta_CA flagged
#> 1 single 0.2050489 0.1142393 0.090809627 FALSE
#> 2 average 0.5151426 0.5198932 -0.004750603 FALSE