Use of artificial intelligence to improve access to initial leukemia diagnosis in low- and middle-income countries.

Abstract

e14117 Background: Disparities in access to cancer diagnostic methods affect especially low and middle-income countries (LMIC). Honduras, a country with limited resources of laboratory, and trained personnel to diagnose leukemia have an estimated incidence of leukemia of 5.8 per 100,000 person-years and the mortality is 4.8 per 100,000 person-years. This country has a single immunophenotyping center and similar to many other LMICs, the preliminary diagnosis of leukemia is performed by a morphological study of blood. Machine learning (ML) is a form of artificial intelligence (AI) that has the potential to improve medical attention in several medical fields. The purpose of this study is to evaluate the impact of ML use, as a method to reduce the time interval and the access to a diagnosis of leukemia in Honduras, where the current average time to perform an initial diagnosis of leukemia is from 2 weeks to 3 months. Methods: A quantitative correlational study was designed. With the help of local informatic developers, an automated image processing algorithm was designed, which was fed with a total of 1009, digital images of bone marrow aspirates and peripheral blood of patients with proven leukemia through immunophenotyping. The images were captured using a microscope Carl Zeiss lens with a 100x objective, the process included the segmentation of leukemia cells for its analysis. Bivariate outcomes were assessed using the Pearson chi-squared test. Results: From 2016 to 2018, a total of 341 samples of patients with any symptom of leukemia were included. After anonymization of patient identification, the samples were analyzed using our algorithm by comparing the cells with its database. Posteriorly, an expert hematologist performed an analysis of the sample. A total of 20 samples (5.8%), were diagnosed with a preliminary diagnosis of leukemia. Of the 20 samples, a total of 10(50%) were acute myeloid leukemia, 6 samples (30%) lymphoblastic leukemia and the remaining 4 samples (20%) were compatible with chronic myeloid leukemia. The average time to make an initial diagnosis of leukemia was 75% and 24% in 24 and 48 hours respectively. Local hematologists managed to make treatment decisions earlier with benefit to the patients. In 19 of the samples (95%), there was correspondence between the morphology diagnosis obtained by our algorithm and the immunophenotyping diagnosis. Conclusions: This preliminary study demonstrates that the use of artificial intelligence constitutes an important element to improve access and shortening of the time required to obtain a diagnosis of leukemia in LMCIs, and represents a method to reduce disparities in access to diagnosis of hematological cancers.