Oops, you're using an old version of your browser so some of the features on this page may not be displaying properly.

MINIMAL Requirements: Google Chrome 24+Mozilla Firefox 20+Internet Explorer 11Opera 15–18Apple Safari 7SeaMonkey 2.15-2.23

Poster Display session

780P - Comparative assessment of machine learning model performance in clinical classification of acute lymphoblastic leukemia subtypes

Date

07 Dec 2024

Session

Poster Display session

Presenters

Seoyoung Lim

Citation

Annals of Oncology (2024) 35 (suppl_4): S1679-S1697. 10.1016/annonc/annonc1699

Authors

S. Lim1, S. Shin2, Y.J. Choi2, S. Lee2, J.R. Choi2

Author affiliations

  • 1 Graduate School Of Medical Science, Brain Korea 21 Plus Project, Yonsei University College of Medicine, 120-752 - Seoul/KR
  • 2 Laboratory Medicine, Severance Hospital - Yonsei University College of Medicine, 03722 - Seoul/KR

Resources

This content is available to ESMO members and event participants.

Abstract 780P

Background

B-cell precursor acute lymphoblastic leukemia (BLL) is a heterogeneous disease, harboring 26 molecular subtypes characterized by unique gene expression profiles (GEPs) and diverse genetic alterations. Diagnosing subtypes defined as GEPs through traditional genetic tests can be challenging. Therefore, there is a need to consider the bioinformatics analysis of the RNA sequencing (RNA-seq) method with subtype diagnosis in detail. Recently, machine learning models for subtype classification based on GEPs have been developed, but comparisons of their performance can be inaccurate due to overlapping test and training datasets.

Methods

Our study included 69 patients diagnosed with BLL at our institution. RNA-seq analysis identified subtypes with genomic events like gene fusions, hotspot SNVs, and virtual karyotypes. We evaluated the performance of several open-source BLL subtype classifiers (ALLCatchR, ALLSorts, ALLIUM, and ALLSpice) on our RNA-seq data. The difficulties in assessing each model came from the different numbers and types of classifiable subtypes and their very own methods to represent prediction results. So, we introduced a method of multi-label encoding to turn it into a unified format of prediction results. A label based evaluation was then conducted to assess the reprocessed multi-label prediction results by calculating the evaluation index for each subtype and averaging these values.

Results

Molecular subtypes were assigned based on analysis of leukemic cells at the time of diagnosis concerning morphology, immunophenotype, molecular genetics, and cytogenetics. A total of 62 patients (89.9 %) had an established molecular subtype, and 7 patients (10.1%) were denoted as Not Other Specified. To validate subtype classifiers, only samples categorized into certain subtypes (n = 62) were used. As a result, ALLSorts showed the highest accuracy at 97.7%. All models performed well but they had lower sensitivity than accuracy and specificity.

Conclusions

In conclusion, the positive predictive results and performance of these GEP-based BLL subtype classifiers have shown potential relevance in clinical decisions. However, several considerations remain before applying these to actual clinical decisions.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Graduate School of Medical Science, Brain Korea 21 PLUS Project, Yonsei University College of Medicine, Seoul, Republic of Korea.

Funding

Has not received any funding.

Disclosure

S. Lee, J.R. Choi: Financial Interests, Personal, Full or part-time Employment: Dxome. All other authors have declared no conflicts of interest.

This site uses cookies. Some of these cookies are essential, while others help us improve your experience by providing insights into how the site is being used.

For more detailed information on the cookies we use, please check our Privacy Policy.

Customise settings
  • Necessary cookies enable core functionality. The website cannot function properly without these cookies, and you can only disable them by changing your browser preferences.