Comparison of Clinician and Machine Learning Predictions of Allo-HCT Prognosis: A Prospective Observational Study

Le résumé fourni par la source

Introduction Recently, the performance of data-driven personalized prognostic prediction using machine learning models has been reported in the field of allogeneic hematopoietic stem cell transplantation (allo-HCT). However, it remains unclear which prediction is superior between the machine learning model and clinicians. Objective We prospectively compared the performance clinicians and a machine learning model in prognostic prediction. Methods This prospective observational study was conducted at Osaka Metropolitan University (OMU). Patients who underwent allo-HCT for hematological malignancies at OMU Hospital between January 2009 and March 2021 were consecutively included. For those who underwent allo-HCT between October 2019 and March 2021, OMU hematologists estimated the following outcomes before the procedure based on pre-transplant clinical information: 1-year overall survival (OS), 1-year progression-free survival (PFS), 1-year cumulative incidence of relapse, and 1-year cumulative incidence of non-relapse mortality (NRM). The same outcomes were also predicted using the machine learning models developed with data from patients who underwent allo-HCT between January 2009 and October 2019, using a Random Survival Forest (RSF) and Gradient Boosted Machine (GBM). Age, ECOG-PS, refined DRI, HCT-CI, conditioning intensity, HLA compatibility, donor source, and the number of transplantations were included as covariates in the machine learning models. Results A total of 57 patients were prospectively evaluated. The median age was 52 years (range: 21-72), and the median follow-up duration was 929 days (range: 8-1370). The median number of hematologists involved in predicting each patient's prognosis was 16 (range: 3-19). The post-transplant outcomes were as follows: 1-year OS, 75.2%; 1-year PFS, 62.8%; 1-year cumulative incidence of relapse, 26.7%; and 1-year cumulative incidence of NRM, 10.1%. The discrimination performances of the RSF and GBM models did not reach that of the clinicians (Table 1). However, the calibration plot demonstrated that the RSF model provided more balanced predictions compared to clinicians and GBM model. Clinicians tended to predict worse outcomes. When we constructed a composite model combining the RSF model with clinicians’ predictions, it most accurately predicted patient prognosis (Fig.1, Table 1). Conclusion For clinicians, collaborating with machine learning to predict the prognosis of allo-HCT may be valuable.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.