Surrogate-guided ensemble Latent Dirichlet Allocation

Surrogate-guided ensemble Latent Dirichlet Allocation

sureLDA(
  X,
  ICD,
  NLP,
  HU,
  filter,
  prior = "PheNorm",
  weight = "beta",
  nEmpty = 20,
  alpha = 100,
  beta = 100,
  burnin = 50,
  ITER = 150,
  phi = NULL,
  nCores = 1,
  labeled = NULL,
  verbose = FALSE
)

Arguments

X

nPatients x nFeatures matrix of EHR feature counts

ICD

nPatients x nPhenotypes matrix of main ICD surrogate counts

NLP

nPatients x nPhenotypes matrix of main NLP surrogate counts

HU

nPatients-dimensional vector containing the healthcare utilization feature

filter

nPatients x nPhenotypes binary matrix indicating filter-positives

prior

'PheNorm', 'MAP', or nPatients x nPhenotypes matrix of prior probabilities (defaults to PheNorm)

weight

'beta', 'uniform', or nPhenotypes x nFeatures matrix of feature weights (defaults to beta)

nEmpty

Number of 'empty' topics to include in LDA step (defaults to 10)

alpha

LDA Dirichlet hyperparameter for patient-topic distribution (defaults to 100)

beta

LDA Dirichlet hyperparameter for topic-feature distribution (defaults to 100)

burnin

number of burnin Gibbs iterations (defaults to 50)

ITER

number of subsequent iterations for inference (defaults to 150)

phi

(optional) nPhenotypes x nFeatures pre-trained topic-feature distribution matrix

nCores

(optional) Number of parallel cores to use only if phi is provided (defaults to 1)

labeled

(optional) nPatients x nPhenotypes matrix of a priori labels (set missing entries to NA)

verbose

(optional) indicating whether to output verbose progress updates

Value

scores nPatients x nPhenotypes matrix of weighted patient-phenotype assignment counts from LDA step

probs nPatients x nPhenotypes matrix of patient-phenotype posterior probabilities

ensemble Mean of sureLDA posterior and PheNorm/MAP prior

prior nPatients x nPhenotypes matrix of PheNorm/MAP phenotype probability estimates

phi nPhenotypes x nFeatures topic distribution matrix from LDA step

weights nPhenotypes x nFeatures matrix of topic-feature weights