Geel, Jennifer A.Hramyka, ArtsiomDu Plessis, JanGoga, YasminVan Zyl, AnelHendricks, Marc G.Naidoo, ThanushreeMathew, RemaLouw, LizetteNeethling, BeverleySchickerling, Tanya M.Omar, Fareed E.Du Plessis, LiezlMadzhia, ElelwaniNetshituni, VhutshiloEyal, KatherineNgcana, Thandeka V.Z.Kelsey, TomBallott, Daynia E.Metzger, Monika L.2025-03-202025-03-202024-10-24Geel, J.A., Hramyka, A., Du Plessis, J. et al. 2024, 'Machine learning to predict interim response in pediatric classical Hodgkin lymphoma using affordable blood tests', JCO Global Oncology, vol. 10, no. e2300435, pp. 1-10. https://DOI.org/10.1200/GO.23.00435.2687-894110.1200/GO.23.00435http://hdl.handle.net/2263/101618PRIOR PRESENTATION : Presented at 55th Annual Conference of the International Society of Pediatric Oncology, Ottawa, Canada, October 11-14, 2023.DATA SHARING STATEMENT : The dataset for this study is available on request.PURPOSE : Response assessment of classical Hodgkin lymphoma (cHL) with positron emission tomography-computerized tomography (PET-CT) is standard of care in well-resourced settings but unavailable in most African countries. We aimed to investigate correlations between changes in PET-CT findings at interim analysis with changes in blood test results in pediatric patients with cHL in 17 South African centers. METHODS : Changes in ferritin, lactate dehydrogenase (LDH), erythrocyte sedimentation rate (ESR), albumin, total white cell count (TWC), absolute lymphocyte count (ALC), and absolute eosinophil count were compared with PET-CT Deauville scores (DS) after two cycles of doxorubicin, bleomycin, vinblastine, and dacarbazine in 84 pediatric patients with cHL. DS 1-3 denoted rapid early response (RER) while DS 4-5 denoted slow early response (SER). Missing values were imputed using the k-nearest neighbor algorithm. Baseline and follow-up blood test values were combined into a single difference variable. Data were split into training and testing sets for analysis using Python scikit-learn 1.2.2 with logistic regression, random forests, na¨ıve Bayes, and support vector machine classifiers. RESULTS : Random forest analysis achieved the best validated test accuracy of 73% when predicting RER or SER from blood samples. When applied to the full data set, the optimal model had a predictive accuracy of 80% and a receiver operating characteristic AUC of 89%. The most predictive variable was the differences in ALC, contributing 21% to the model. Differences in ferritin, LDH, and TWC contributed 15%-16%. Differences in ESR, hemoglobin, and albumin contributed 11%-12%. CONCLUSION : Changes in low-cost, widely available blood tests may predict chemosensitivity for pediatric cHL without access to PET-CT, identifying patients who may not require radiotherapy. Changes in these nonspecific blood tests should be assessed in combination with clinical findings and available imaging to avoid undertreatment.en© 2024 by American Society of Clinical Oncology. Licensed under the Creative Commons Attribution 4.0 License.Blood testClassical Hodgkin lymphoma (cHL)Positron emission tomography-computerized tomography (PET-CT)PediatricPatientsSDG-03: Good health and well-beingArticle