Extract Foundation Model Clinical Sequence Tokens — extract_clinical_sequence

Formats clinical event sequences into tokenized sequences compatible with clinical BERT and Foundation Models (EHR-SHOT, CEHR-BERT, FEMR).

Usage

extract_clinical_sequence_tokens(
  con,
  features_parquet,
  lookback_days = 365,
  min_patient_freq = 1,
  domains = c("condition", "drug", "procedure"),
  max_seq_len = 512,
  include_index_date = FALSE,
  person_col = NULL,
  index_date_col = NULL,
  outcome_col = NULL,
  concept_table = NULL,
  anchor_date = "auto"
)

Arguments

con: Active DuckDB connection.
features_parquet: Path to cohort parquet file.
lookback_days: Historical lookback window in days (default 365).
min_patient_freq: Minimum patient frequency threshold (default 1).
domains: Clinical domains to include (default: condition, drug, procedure).
max_seq_len: Maximum sequence length for input_ids (default 512).
include_index_date: Whether events on index date are included (default FALSE).
person_col, index_date_col, outcome_col: Optional column aliases in cohort parquet.
concept_table: Concept table name.
anchor_date: Strategy for selecting index date.

Value

A list with input_ids (matrix), attention_mask (matrix), token_type_ids (matrix), labels, vocab, id_to_token, concepts, max_seq_len.