Back
It is crucial to understand the trajectory of kidney function, as this facilitates the optimal and timely management of patients with CKD (for example, discussions of prognosis, medications to protect kidney function, timely discussion about kidney replacement therapy etc). To address these essential needs, we've developed the Kidney Disease Progression Prediction (KDPP), part of the Digital Kidney Care Suite.
KDPP integrates two deep-learning models: KDPP-RP, which predicts the risk of rapid kidney disease progression, and KDPP-IR, forecasting the initiation of renal replacement therapy (maintenance hemodialysis, peritoneal dialysis, or kidney transplantation). Both utilize routine laboratory parameters from CKD stage 3-5 patients. These models stem from a derivation dataset (includes training, testing, and internal validation) of 9,529 patients (2003-2020) at China Medical University Hospital (CMUH), Taiwan. Additionally, models were externally validated using data from 986 patients (2017-2022) at Asia University Hospital (AUH), Taiwan. We developed a deep learning algorithm employing a neural network architecture that incorporates a gated recurrent neural network for sequential eGFR data handling. Fully connected layers were employed to process baseline variables and yield probabilistic predictions across 1, 2 and 5 year timeframes. Inputs encompass demographic, comorbidities, lifestyle factors, laboratory parameters, and kidney function trends. Predictions are articulated as risk scores, categorizing patients into specific risk tiers (low/moderate/high). We assessed model performance using AUC (area under receiver operating characteristic curve), as the primary metric. The performance of low and high risk stratification for KDPP-RP and KDPP-IR was evaluated using sensitivity and specificity, along with positive and negative predictive value (PPV/NPV).
Baseline characteristics of the derivation cohort from CMUH had a median age of 69.3 years, 43% female, 53% diabetic, and 80% hypertension. In validation datasets, KDPP-RP achieved AUCs of 0.76, 0.82, 0.82 for 1-year, 2-year, and 5-year timeframes at CMUH, and the corresponding AUCs were 0.77, 0.77, and 0.78 at AUH. KDPP-IR had AUCs of 0.95, 0.95, and 0.96 at CMUH, and 0.95, 0.95, and 0.90 at AUH. Both models showed varying sensitivity and specificity at CMUH, with KDPP-RP's low-risk group achieving a 5-year NPV of 0.79 and a high-risk group with PPV between 0.68 and 0.86 (Table 1). In contrast, KDPP-IR's low-risk group reached a 1- and 2-year sensitivity of 0.99, and a consistent NPV of 1.00, while the high-risk group had sensitivity from 0.69 to 0.80, high specificity exceeding 0.94, and consistently high PPV, except for the 1-year timeframe. The adjusted Kaplan-Meier curves exhibited excellent risk differentiation with a statistical significance of p < 0.001 (Figure 1).
The KDPP models demonstrated strong performance and generalizability within Taiwan's hospitals. Future plans include conducting additional validation studies in different countries to assess and optimize their broader applicability.
The KDPP tool was granted ‘BREAKTHROUGH DEVICE DESIGNATION’ by the US FDA in May 2023.