Application of text mining techniques to extract leukaemia information from reports of flow cytometry investigations conducted between 2015 and 2019 at a Johannesburg academic hospital

Abstract

Leukaemia data for epidemiological and public health planning is limited in South Africa. The ap- plication of text mining techniques could improve leukaemia data representation. This study aimed to apply text mining techniques to extract clinical and epidemiological data from narrative labora- tory reports of immunophenotyping flow cytometry completed at the Charlotte Maxeke Johannesburg Academic Hospital (CMJAH) laboratory from 2015 to 2019. An expert driven regular expressions (RegEx) based algorithm and topic modelling were used. The RegEx algorithm was used to place the reports into categories of Acute Lymphocytic Leukaemia, Acute Myeloid Leukaemia, Chronic Lym- phoid Leukaemia or Chronic Myeloid Leukaemia. The algorithm achieved F1 scores comparable to those of similar tasks in the literature. It achieved 0.74 for leukaemia and 0.93 for non-leukaemia classes respectively. Topic modelling was evaluated using coherence scores, inspection of visuali- sation of clusters, and interpretability of topics. Topic modelling uncovered more information from the corpus, relative to the regex based algorithm. Age standardised incidence rates for leukaemia in the City of Johannesburg, calculated using extracted data ranged from 0.54 to 0.87 per 100 000 population, which was slightly lower than the figures reported by the National Cancer Registry. This study has shown that text mining techniques can be applied to improve representativity of leukaemia data. It is recommended that further work be carried out to refine both the RegEx algorithm and topic models.