Multi-Label Classification of ICD-10 Coding & Clinical Notes Using MIMIC & CodiEsp

ICD-10, is the 10th revision of the International Classification of Diseases (ICD) coding standard used to get proper treatment and charged accordingly for any medical service. Clinical notes for patients’ assessment and treatment are captured by the physicians in free-form texts. Abbreviations and/or misspellings of words in this free-form text creates ambiguity and leads to the complexity and errors within the medical billing and coding process. In our study, multi-label classification experiments are conducted using preprocessing and deep learning techniques against two known corpora: the MIMIC-III (Medical Information Mart for Intensive Care) and CodiEsp. Clinical notes use abbreviation normalization and misspelled words are corrected. Logical hierarchy of lower-level ICD codes (labels) are converted to their ICD Chapters (first-level hierarchy). Our study performs several experiments with our preprocessed training datasets against several deep learning models. Our results showed that the deep learning attention mechanism is effective in enhancing ICD-10 predictions and the HAN GRU yields the best in F1 measures and 10-fold cross validation.