Enhancing Dose Predictions in Radiopharmaceutical Therapies by Applying Machine Learning to Time-Activity Curve Features: A Simulation Study using PBPK Modeling

Introduction: Pre-therapeutic dynamic PET images hold significant potential for predicting doses delivered in radiopharmaceutical therapies (RPTs), by providing key insights into the pharmacokinetics of radiopharmaceuticals. This study specifically focuses on methods to forecast Lu177-PSMA RPT absorbed doses, utilizing features extracted time-activity curves (TAC) from dynamic PET images, by performing realistic physiologically based pharmacokinetic (PBPK) modeling for pre-therapy imaging and therapeutic PSMA RPTs.

Methods: A comprehensive whole-body PBPK model was developed using MATLAB SimBiology and employed to simulate TACs for tumors, salivary glands, and kidneys. Model parameters were obtained from the literature or fitted using post-therapeutic SPECT data. In the creation of our dataset, realistic ranges of pharmacokinetic parameters were utilized, yielding relevant TAC data. The dataset was generated for three distinct radionuclides, namely, Cu-64, . Key TAC features, including Ti, Td, Time-To-Peak, A-Max, A-Mean, A-Median (A=Activity), Entropy, Skewness, Percentile1, Percentile10, and Percentile90, were extracted for further analysis (Ti is the time it takes for dose rate to reach half max value during an increase, while Td is the time it takes for the dose rate to drop to half max value during a decrease; a percentile is an activity-point in dataset indicating a certain percentage of values below it). The dose was calculated from post-therapeutic simulated Lu177-PSMA data and was used as predictive label. Univariate correlation of individual features with dose was obtained using Spearman correlation. Five machine learning (ML) algorithms, namely Regression tree, Ensemble of Trees, Gaussian Process Regression (GPR), Neural Network and SVM, were used to combine features for predictive modeling, and performances were measured using mean absolute error (MAE) and R² via 5-fold cross validation.

Results: All results are depicted in Figure 1-3, including simulated Pre/Post therapeutic TACs (1), correlation of features with dose and models performance (2) and best R² for three organs (3). In the case of radionuclides F-18 and Ga-68, no univariate correlated features were identified in relation to dose. However, for Cu-64, two specific features, namely, Ti and Td, which measure TAC increase/decrease half-times, were shown as highly correlated features. Results for MAEs of machine learning algorithms for different radionuclides show Cu-64 exhibits the lowest prediction errors across all machine learning algorithms, particularly in relation to the salivary gland and tumor. The Cu-64 model yielded the most precise predictions, showcasing minimal errors across specific organs: tumor (MAE; 3.49, ML=GPR), salivary gland (MAE; 0.40, ML=GPR), and kidney (MAE; 3.67, ML=SVM). The most favorable R² values were identified for Cu-64, as depicted in Figure 3. Within this scope, the R² scores for tumor (0.88, ML=GPR), salivary gland (0.76, ML=GPR), and kidney (0.60, ML= GPR) showcased noteworthy performances.

Conclusions: Enhanced predictive dosimetry can open the way to personalized RPTs, and can be achieved through sophisticated mathematical modeling, including PBPK approaches, representing an advanced and refined methodology. Integration of pre-therapeutic PET imaging, utilizing long-life radionuclides such as Cu-64, holds promising potential to significantly augment predictive performance within the domain of dosimetry prediction. Furthermore, integrating machine learning with key time-activity curve metrics has the potential to establish a novel framework for predictive modeling in precision RPTs.

• Download figure
• Open in new tab
• Download powerpoint