This package provides a solution for performing code embedding on EHR data, specifically for medical codes such as ICD (International Classification of Diseases) codes, CPT (Current Procedural Terminology) codes, and national drug codes. The goal is to translate these codes into embedding vectors to better capture their underlying correlations, facilitating downstream analysis. The method used in developing this package is based on the GloVe (Global Vectors for Word Representation) algorithm, which leverages global statistical information by analyzing code-to-code co-occurrence patterns across the entire dataset.
The package includes the following functions, outlined as steps to
follow for embedding the codes: 1. Create code co-occurrence matrix This
steps allow users to obtain a sparse matrix that record the
co-occurrence of the codes in the dataset. users may choose to decide
visualing code occurrance pattern by creating heatmap or decide dealing
with rare codes based on the resulting matrix. - cooccur_hpc() This
function computes a co-occurrence matrix for a dataset that containing
medical codes, based on the time window for each patient’s data. It is
optimized for high-performance computing (HPC) environments using the
slurmR package. - coccur_local() This function computes a
co-occurrence matrix for a dataset that containing medical codes, based
on the time window for each patient’s data. It is optimized for local
computing environments.
- Create sparse matrix for co-occurence pattern.
- cooccur_pairs() This function takes an input symmetric co-occurrence matrix and returns a data frame representing the co-occurrences between pairs of items. Each pair consists of a row index and a column index (i.e., co-occurring pairs) with an associated count, representing the number of co-occurrences.
- Derive Marginal Counts for Codes and Joint Counts for Code Pairs
- getsg() This function calculates the marginal counts for each code and the joint counts for each pair of codes in the co-occurrence matrix.
-
Calculate the Pointwise Mutual Information (PMI) value for code pair. -pmi_df() This function computes the Pointwise Mutual Information (PMI) value for each code in the co-occurrence matrix. It returns a data frame that includes the following details: the code, its co-occurring codes, the marginal count for each code, the joint count for the code pairs, and the PMI value for each pair.
-
Transform Code-Code PMI into Matrix In this step, the PMI values from the previous step are transformed into a symmetric matrix. This can be done in two ways: using the raw PMI values directly or applying the shifted positive PMI (SPPMI) method, commonly used in word embedding for paragraphs and notes. SPPMI is calculated as max(PMI - log(k), 0).
- pmi_matrix() This function transforms the information obtained in the previous step into a symmetric matrix, where each row/column represents a code, and each entry contains the corresponding PMI value for each code pair.
- sppmi_matrix() This function transforms the information obtained in the previous step into a symmetric matrix, where each row/column represents a code, and each entry contains the corresponding SPPMI value for each code pair.
- Embedding the codes
- truncated_svd() This function performs truncated Singular Value Decomposition (SVD) on the PMI/SPPMI matrix to generate code embedding vectors. Users can specify the number of dimensions and the maximum number of iterations for the embedding process. Additionally, they can choose to perform a full SVD and decide whether to remove zero-information vectors.
This package includes an example dataset specifically designed for testing and demonstration purposes, it comprises 2,000 unique patient IDs, 500 unique ICD-9 codes, and time sequences corresponding to each patient’s recorded events.