Prediction of Postoperative Creatinine Levels by Artificial Intelligence after Partial Nephrectomy

• Published in: Medicina (Kaunas, Lithuania) Vol. 59; no. 8; p. 1402
• Main Authors: Shin, Tae Young, Han, Hyunho, Min, Hyun-Seok, Cho, Hyungjoo, Kim, Seonggyun, Park, Sung Yul, Kim, Hyung Joon, Kim, Jung Hoon, Lee, Yong Seong
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 31-07-2023; MDPI
• Subjects: acute kidney injury; Algorithms; Artificial intelligence; Creatinine; Datasets; Ischemia; Kidney diseases; Medical records; partial nephrectomy; Patients; postoperative renal function; Robots; Surgery; Tumors; South Korea
• ISSN: 1648-9144; 1010-660X; 1648-9144
• DOI: 10.3390/medicina59081402

Background and Objectives: Multiple factors are associated with postoperative functional outcomes, such as acute kidney injury (AKI), following partial nephrectomy (PN). The pre-, peri-, and postoperative factors are heavily intertwined and change dynamically, making it difficult to predict postoperative renal function. Therefore, we aimed to build an artificial intelligence (AI) model that utilizes perioperative factors to predict residual renal function and incidence of AKI following PN. Methods and Materials: This retrospective study included 785 patients (training set 706, test set 79) from six tertiary referral centers who underwent open or robotic PN. Forty-four perioperative features were used as inputs to train the AI prediction model. XG-Boost and genetic algorithms were used for the final model selection and to determine feature importance. The primary outcome measure was immediate postoperative serum creatinine (Cr) level. The secondary outcome was the incidence of AKI (estimated glomerular filtration rate (eGFR) < 60 mL/h). The average difference between the true and predicted serum Cr levels was considered the mean absolute error (MAE) and was used as a model evaluation parameter. Results: An AI model for predicting immediate postoperative serum Cr levels was selected from 2000 candidates by providing the lowest MAE (0.03 mg/dL). The model-predicted immediate postoperative serum Cr levels correlated closely with the measured values (R2 = 0.9669). The sensitivity and specificity of the model for predicting AKI were 85.5% and 99.7% in the training set, and 100.0% and 100.0% in the test set, respectively. The limitations of this study included its retrospective design. Conclusions: Our AI model successfully predicted accurate serum Cr levels and the likelihood of AKI. The accuracy of our model suggests that personalized guidelines to optimize multidisciplinary plans involving pre- and postoperative care need to be developed.