Estimasi dosimetri ginjal yang akurat sangat penting dalam terapi [177Lu]Lu-PSMA-617 pada pasien kanker prostat. Oleh karena itu, Penelitian ini bertujuan untuk mengembangkan model farmakokinetika populasi berbasis Non-Linear Mixed-Effect Modeling (NLMEM) serta mengevaluasi akurasi estimasi dosimetri berbasis pengukuran biokinetik tunggal, atau dapat disebut juga Single-Time-Point (STP), menggunakan metode Generalized Additive Model (GAM), NLMEM, Machine Learning (ML), dan gabungan NLMEM + ML. Data aktivitas ginjal dikumpulkan dari 101 pasien pada lima time point (TP) pasca injeksi. Enam fungsi sum-of-exponentials (SOE) dengan parameterisasi berbeda diujikan menggunakan pendekatan Population-Based Model Selection (PBMS) NLMEM, dan fungsi terbaik dipilih berdasarkan evaluasi goodness-of-fit dan bobot Akaike. Fungsi SOE terpilih digunakan untuk menghitung Time-Integrated Activity Coefficient (TIAC) referensi. Empat metode dosimetri STP dibandingkan: (1) STP berbasis metode GAM (STPGAM), (2) STP berbasis NLMEM dengan fungsi SOE optimal (STPNLMEM), (3) STP berbasis metode ML dengan Support Vector Regression (SVR) (STPML), dan (4) metode hibrida yang menggabungkan metode NLMEM dan ML (STPNLMEM+ML). Untuk metode berbasis ML, variabel covariate pasien, serta hyperparameter model ML diseleksi secara sistematis. Evaluasi akurasi dilakukan dengan menghitung persentase Relative Deviation (%RD), persentase Root-Mean-Square Error (%RMSE) dan Mean Absolute Percentage Error (MAPE) terhadap TIAC referensi. Hasil menunjukkan bahwa fungsi SOE dengan tujuh parameter memberikan kecocokan terbaik dalam pendekatan PBMS NLMEM. Metode STPGM memiliki performa terendah di seluruh time point. Selanjutnya, STPNLMEM dan STPML memiliki peforma yang setara dan lebih baik dari STPGM. Sementara itu, metode gabungan STPNLMEM+ML menunjukkan kinerja terbaik secara konsisten, dengan nilai %RMSE (MAPE) yang paling rendah, yakni 21,72% (15,54%), 12,56% (9,08%), 12,88% (9,00%), 13,72% (10,23%), dan 17,80% (13,56%) dari TP1 hingga TP5. Kesimpulannya, kombinasi model farmakokinetika populasi berbasis NLMEM dan pendekatan ML terbukti mampu meningkatkan akurasi estimasi dosis ginjal hanya dengan pengukuran biokinetik tunggal. Metode gabungan STPNLMEM+ML menunjukkan akurasi tertinggi secara konsisten di semua time point, menjadikannya solusi yang efisien dan dapat diandalkan untuk penerapan dosimetri STP dalam praktik klinis terapi radionuklida.

---

Accurate renal dosimetry estimation is essential in [177Lu]Lu-PSMA-617 therapy for prostate cancer patients. Therefore, this study aims to develop a population pharmacokinetic model based on Non-Linear Mixed-Effect Modeling (NLMEM) and to evaluate the accuracy of Single-Time-Point (STP) dosimetry using various approaches: Generalized Additive Model (GAM), NLMEM, Machine Learning (ML), and a hybrid method combining NLMEM and ML. Renal activity data were collected from 101 patients at five post-injection time points: TP1 (1.94 ± 0.75 h), TP2 (18.90 ± 1.02 h), TP3 (43.12 ± 1.53 h), TP4 (66.41 ± 1.43 h), and TP5 (165.79 ± 24.68 h). Six different parameterizations of sum-of-exponentials (SOE) functions were evaluated using a Population-Based Model Selection (PBMS) approach within NLMEM, and the best-performing function was selected based on goodness-of-fit and Akaike weights. The selected SOE function was then used to calculate reference Time-Integrated Activity Coefficient (TIAC). Four STP dosimetry methods were compared: (1) GAM-based STP (STPGAM), (2) NLMEM-based STP using the optimal SOE function (STPNLMEM), (3) ML-based STP using Support Vector Regression (STPML), and (4) a hybrid method combining NLMEM and ML (STPNLMEM+ML). For ML-based methods, patient-specific covariates and ML hyperparameters were systematically selected. Accuracy evaluation was performed by calculating the percentage of Relative Deviation (%RD), Root-Mean-Square Error (%RMSE), and Mean Absolute Percentage Error (MAPE) against the reference TIAC. The results showed that the seven-parameter SOE function provided the best fit under the PBMS-NLMEM approach. STPGM demonstrated the lowest performance across all time points, while NLMEM and ML performed similarly and better than STPGM. The hybrid STPNLMEM+ML method consistently outperformed all other methods, achieving the lowest %RMSE (MAPE) values of 21.72% (15.54%), 12.56% (9.08%), 12.88% (9.00%), 13.72% (10.23%), and 17.80% (13.56%) from TP1 to TP5, respectively. In conclusion, combining NLMEM-based population pharmacokinetic modeling with machine learning approaches significantly improves the accuracy of renal dose estimation using a single biokinetic measurement. The hybrid STPNLMEM+ML method demonstrated consistently superior accuracy across all time points, making it an efficient and reliable solution for clinical implementation of STP dosimetry in radionuclide therapy.