PT - JOURNAL ARTICLE AU - Roberto Fedrigo AU - John Harpel AU - Vijay Reddy AU - Mark Berger AU - Arman Rahmim AU - Carlos Uribe AU - Qing Liang TI - Feasibility of Machine Learning-Assisted Personalized Dosimetry for Targeted Radioimmunotherapy with Anti-CD45 Iodine (<sup>131</sup>I) Apamistamab [Iomab-B] in Patients with Active Relapsed or Refractory Acute Myeloid Leukemia; a Phase III Clinical trial DP - 2021 May 01 TA - Journal of Nuclear Medicine PG - 73--73 VI - 62 IP - supplement 1 4099 - http://jnm.snmjournals.org/content/62/supplement_1/73.short 4100 - http://jnm.snmjournals.org/content/62/supplement_1/73.full SO - J Nucl Med2021 May 01; 62 AB - 73Objectives: The SIERRA phase III clinical trial uses an iodine-131 labelled antibody (Apamistamab, Iomab-B) to deliver high doses of radiation to leukemia cells of older patients with relapsed or refractory acute myeloid leukemia (AML). This is done as a conditioning regimen for bone marrow transplant. The therapeutic infused activity is personalized based on an initial dose assessment conducted using a low activity (7-20mCi) Iomab-B infusion followed by 3 planar acquisitions (right after infusion, 24 h, 72-96 h). The goal is to deliver the highest dose possible to cancer cells and a maximum of 24 Gy to the liver. Simplifying the dose calculation process would minimize the workload and improve patient convenience. One way to do this is to reduce the number of imaging sessions required. Machine learning has become a powerful tool for predictive modelling within healthcare and nuclear medicine imaging. The availability of demographic, anatomic structural (CT), and lab test information within the context of SIERRA opens the possibility to improve dosimetric prediction using machine learning. Thus, this study aimed to investigate the feasibility of using machine learning and a single timepoint image (first image) in combination with additional patient clinical information to predict the infused activity that is required to optimize the radiation dose delivered to cancer cells while sparing healthy organs. Methods: 74 patients and 212 features (organ uptake and volume from planar and CT images respectively, demographic information, blood, and marrow tests) were evaluated. Features with greater than 20% of entries missing were discarded, and remaining missing values were replaced with the median. Interquartile range of features was scaled using the RobustScaler preprocessing method of Python’s scikit-learn library. A novel feature selection pipeline was designed and implemented using Python’s mlxtend library (SequentialFeatureSelector method). The LASSO linear regression method (α=0.3) was used to train and test the model (70/30 train/test split). Results: 8 robust features were selected to account for variations in prescribed activity: liver mass, initial liver uptake, serum aspartate aminotransferase value (pre and post dosimetric infusion), body surface area (BSA), serum lactase dehydrogenase, lymphocyte count, and neutrophil count. The machine learning model had r2=0.71 and RMSE=179.8mCi (mean prescribed activity=716.4mCi). Error of less than 30% was reported in 63.6% of patients, which represents the fraction of patients within the uncertainty threshold of optimized gamma-camera imaging. SIERRA-specific constraints (maximum manufactured activity is 1030mCi) led to further reduced model error within the clinical setting (RMSE=167.4mCi). The machine learning model outperformed conventional single-variable methods, such as those based on body or liver mass (RMSE=337.2mCi or 278.3mCi, respectively). Furthermore, the current method appears to significantly improve upon the predictive capability of liver dosimetry with single initial time point (RMSE=331.8mCi). Conclusion: This work evaluates the feasibility of an alternative dosimetry method that can be used to simplify the clinical protocol and reduce the time for dose calculation for this phase III radioimmunotherapy trial. View this table:Test set performance metrics for Iomab-B dose prediction, using models of varying complexity.