Unsupervised Linkage between ICD- and Alpha-ID Codes and Real-World Diagnoses from Medical Reports by Means of the "word2vec" Method

Background Transformation into a standardised code system such as ICD-10 or Alpha-ID is required before medical reports can be scientifically analysed. This is due to the use of different terminologies and the frequent use of synonyms. The so-called "word vector embedding" seems to be suitable for the generation of the required thesaurus, because synonymous diagnoses can be identified independently of the spelling after suitable training of the underlying neural network. Methods All letters from a total of 50,000 patients were extracted anonymously. Diagnoses consisting of several words were merged into single words by means of phrase recognition and the "word2vec" model was trained on the text corpus of 352 megabytes. A total of 3742 diagnoses and ophthalmological interventions were extracted semi-automatically. The ophthalmological ICD and Alpha-ID codes were downloaded together with the official descriptions from the DIMDI website and the ophthalmological diagnoses/interventions were automatically linked with the nearest ICD- and Alpha-ID codes in the "word2vec" model. Results The "word2vec" model assigned 90% of the doctor's letter diagnoses correctly to appropriate ICD-10 codes. At the finer level of Alpha-ID, the rate of correct assignments was only 76%. The interventions were assigned to the correct indication in 92% of cases. Rare diseases, unusual designations and code degeneration in the official DIMDI file were identified as sources of error for incorrect or missing allocations. Discussion A diagnostic thesaurus can be generated with the "word2vec" method from a corpus of anonymised medical reports and the official Alpha-ID file from the DIMDI website. This thesaurus could be used for automatic extraction of diagnoses from doctor's letters in the future, given appropriate manual revision.