Skip to contents

Map Diagnosis Text to ICD Candidates

Source: R/prep_diagnosis.R
prep_map_diagnosis_to_icd.Rd

Maps free-text diagnosis strings to ICD-10 candidate descriptions using one of three methods:

Usage

prep_map_diagnosis_to_icd(
  data,
  text_col = "diagnosis_text",
  reference = NULL,
  method = c("exact", "fuzzy", "python_embedding"),
  icd_desc_col = "description3",
  icd_code_col = "icd_code_who_eq",
  top_k = 5L,
  threshold = 0,
  model = "FremyCompany/BioLORD-2023",
  id_col = NULL
)

Arguments

data

Data frame.

text_col

Character. Column containing prepared diagnosis text (output of prep_diagnosis_text()). Default "diagnosis_text".

reference

Data frame or NULL. ICD-10 reference table. If NULL, loads from inst/extdata/icd10_who.csv.

method

Character. One of "exact", "fuzzy", "python_embedding". Default "exact".

icd_desc_col

Character. Column in reference to match against. Default "description3" (most concise ICD labels).

icd_code_col

Character. Column in reference holding ICD codes. Default "icd_code_who_eq".

top_k

Integer. Maximum ICD candidates to return per input string. Default 5. Ignored for "exact" (returns all exact matches).

threshold

Numeric. Minimum similarity score (0-1) to retain a candidate. Default 0.0 (keep all). For "fuzzy": similarity is 1 - normalised_distance.

model

Character. Sentence-transformers model name. Used only for "python_embedding". Default "FremyCompany/BioLORD-2023".

id_col

Character or NULL. Identifier column to carry through into the output. Default NULL (output contains only match columns).

Value

Long data frame with columns: diagnosis_text, icd_prediction, icd_code, icd_score, icd_rank, icd_method. If id_col is supplied, it is included as the first column.

Details

"exact"

Case-insensitive exact string match against ICD descriptions. Fast, high precision, low recall.

"fuzzy"

String distance matching via stringdist. Handles typos and minor variations. Requires stringdist.

"python_embedding"

Semantic embedding similarity using the Python alethia package via reticulate. Highest recall. Requires Python, alethia, and a sentence-transformers model.