Prospective Validation of Machine Learning Model to Predict Platinum Induced Nephrotoxicity in Cancer Patients

Overview

This study aims to investigate the utility of predictive models for chemotherapy-induced nephrotoxicity in the Taiwanese cancer population. The investigators will prospectively collect clinical data from enrolled participants, including demographic information, comorbidities, laboratory data, and chemotherapy treatment details. After chemotherapy administration, participants' renal function will be monitored over time to assess the development of nephrotoxicity, based on changes in serum creatinine (SCr) and other relevant clinical criteria. The primary objective is to evaluate and compare the predictive performance of a machine learning model and clinical physicians, using the area under the receiver operating characteristic curve (AUROC) as the main metric for discrimination performance.

This is a prospective cohort study conducted at Wan Fang Hospital, designed to evaluate the predictive performance and clinical utility of established machine learning models for detecting nephrotoxicity in cancer patients receiving platinum-based chemotherapy. The models were developed using a Long Short-Term Memory (LSTM) neural network architecture, trained on a retrospective dataset from January 1, 2009 to January 31, 2022. All model parameters were locked after training to ensure reproducibility, prevent data leakage, and maintain the integrity of prospective validation. Cancer patients receiving platinum-containing agents (such as Cisplatin and Carboplatin) between October 2023 and August 2025 will be recruited, with follow-up until November 2025. After confirming eligibility based on the inclusion and exclusion criteria, written informed consent will be obtained from each participant prior to data collection. Each administration of platinum chemotherapy is treated as a separate prediction case. For every administration, the following renal outcomes will be predicted: 1. Acute Kidney Injury (AKI) : occurring within 14 days after chemotherapy, defined according to the Common Terminology Criteria for Adverse Events (CTCAE) version 5.0. 2. Acute Kidney Disease (AKD) : occurring within 89 days after chemotherapy, defined according to the Acute Disease Quality Initiative (ADQI) 2016 criteria. After obtaining informed consent, the investigators will collect general patient information (gender, age, height, weight), cancer-related information (stage), chemotherapy details (administration date, course, dosage, concomitant chemotherapy drugs, and the number of administrations), and laboratory data (SCr, glomerular filtration rate \[GFR\]). Relevant follow-up data will be used to evaluate treatment effects and disease prognosis. All data will be de-identified and recorded using a research identification number. In addition to model predictions, clinicians' risk assessments will be obtained for the same chemotherapy administrations. Four physicians will participate in this process. Each chemotherapy case will be randomly assigned to one of the four physicians for independent prediction. Physicians will be blinded to model predictions and required to return their assessment regarding the risk of AKI and AKD within one week. If the interval between two chemotherapy administrations exceeds 42 days (6 weeks), or if there is a change in chemotherapy regimen (e.g., switching from Cisplatin to Carboplatin or vice versa), the participant will be treated as a new subject with a separate research identification number for subsequent analysis. The occurrence of AKI and AKD will be predicted using the LSTM-based machine learning model and compared with physicians' predictions. The performance of both the model and physicians will be evaluated using metrics such as AUROC, area under the precision-recall curve (AUPRC), and sensitivity. Participants will also be classified into nephrotoxicity risk groups based on both the model's and physicians' predictions. The study aims to verify the utility of the model and analyze the causes of disparities between predicted outcomes and observed outcomes. Categorical variables will be analyzed using the Chi-square test or Fisher's exact test based on the expected cell counts. Continuous variables will be compared using the independent t-test or the Mann-Whitney U test, depending on the distribution of the data. A two-sided p-value \< 0.05 will be considered statistically significant. The 95% confidence intervals (CIs) for AUROC and AUPRC will be calculated using DeLong's test and the bootstrap method with 1,000 iterations. All statistical analyses will be performed using Python (version 3.11.12) and R (version 4.4.3).