Study protocol for a consortium linking health medical records, biospecimens, and biosignals in Korean patients with acute kidney injury (LINKA cohort)

Ethics statements

The study protocol for LINKA (named from the consortium LINking health medical records, biospecimens, and biosignals in Korean patients with Acute kidney injury) was reviewed and approved by the Institutional Review Boards of each participating center; Hallym University Sacred Heart Hospital (2022-11-015), Seoul National University Hospital (J-2312-022-1489, J-2312-014-1488), Seoul National University Boramae Medical Center (10-2022-112, 10-2023-84), Seoul National University Bundang Hospital (B-2312-871-401, B-2312-871-402), The Catholic University of Korea, Yeouido St. Mary’s Hospital (SC23TNDE0128), Korea University Guro Hospital (2023GR0392), Keimyung University Dongsan Hospital (2023-08-064), and Kyungpook National University Hospital (2023-07-027). In addition, an overarching ethics oversight committee specifically established for LINKA rigorously reviewed the study design and protocol. This committee will continuously oversee and address any ethical issues that arise during the course of the study, ensuring adherence to the highest standards of research ethics and participant safety.

Study design and population

This was a multicenter, observational cohort study that incorporated both a biobank and a databank, specifically targeting patients with AKI requiring CKRT. Nephrologists from eight major university-affiliated hospitals spanning metropolitan and regional areas across Korea participated in this study, along with the inclusion of a diverse patient population.

This study included retrospective and prospective cohorts. The retrospective cohort comprised adult patients aged ≥ 9 years who underwent CKRT between January 2017 and June 2022 in the intensive care unit. The inclusion criteria were based on AKI according to the Kidney Disease: Improving Global Outcomes (KDIGO) guideline [10,11], which defines AKI as an increase in serum creatinine by ≥0.3 mg/dL within 48 hours or an increase to ≥1.5 times baseline within the last 7 days, or a urine volume of less than 0.5 mL/kg/hr for 6 hours. We included AKI patients who had CKRT initiated due to critical changes in fluid, electrolyte, and acid-base balance, including severe hyperkalemia, metabolic acidosis, or complications arising from fluid overload that are unresponsive to medical management. Patients without baseline kidney function data, with unknown kidney function at 3 months, or those who died within 3 months of CKRT initiation were excluded. The study involving the prospective cohort commenced in January 2024 and involves the collection of biospecimens.

Common exclusion criteria for both retrospective and prospective cohorts comprise patients on maintenance dialysis for end-stage kidney disease (ESKD), those with suspected post-renal AKI based on clinical assessments including imaging studies or those undergoing acute intermittent hemodialysis before CKRT, and pregnant women. Through these exclusion criteria, we aimed to maintain a relatively homogenous study population with clinically severe AKI. Patients with advance directives against resuscitation, those enrolled in other randomized trials and those who refused consent were also excluded from this prospective cohort study. Inclusion and exclusion of patients will be conducted with appropriate consultation and discussion with the nephrology department. A schematic of the patient recruitment flowchart is shown in Fig. 1.

Figure 1.

Flowchart of participant selection and grouping in the LINKA (consortium LINking health medical records, biospecimens, and biosignals in Korean patients with Acute kidney injury) cohort.

AKI, acute kidney injury; CKRT, continuous kidney replacement therapy; ESKD, end-stage kidney disease; RCT, randomized controlled trial.

Sample size calculation

The sample size for this study was determined using diagnostic power calculations based on our previous research findings [12]. Among the 1,764 patients who underwent CKRT for severe AKI, 331 survivors with confirmed kidney function at 3 months were identified, resulting in a prevalence of 0.1876. Among the 331 patients, 125 developed AKD, equating to a proportion of 0.3776. Setting a width of 0.04 for the 95% confidence interval, the required sample size (n) was calculated to be 288 patients [13]. Further, factoring in the prevalence, the total sample size (N = n/prevalence) was calculated to be approximately 1,535 patients. Considering these aspects, we aimed to recruit approximately 1,500 participants in both the retrospective and prospective cohorts.

Data collection parameters

Table 1 outlines the parameters for data collection in this study, including demographic and clinical information, biospecimens, treatment details, and key outcomes at designated time points from CKRT initiation to final follow-up. The outcomes include weaning status from CKRT (continuing CKRT, transitioned to intermittent dialysis, or not undergoing dialysis) at 7 days, 3 months, 12 months, and the final follow-up, progression to AKD, CKD, ESKD, and all-cause mortality. AKD and CKD are defined by the KDIGO guidelines as an incomplete recovery of renal function after AKI: AKD persists for 7 days to 3 months, and CKD for 3 months or more, irrespective of the cause [10,11]. The retrospective cohort was used for data extraction from existing medical records. Conversely, participants in the prospective cohort or their legal guardians will provide informed consent for the use and collection of medical information and biospecimens. Participants in this cohort had the right to withdraw upon request, ensuring secure handling and potential destruction of their data and biospecimens. A comprehensive framework of the LINKA cohort is illustrated in Fig. 2.

Table 1.

Information on parameters for the LINKA (consortium LINking health medical records, biospecimens, and biosignals in Korean patients with Acute kidney injury) cohort

Figure 2.

Comprehensive framework of the LINKA (consortium LINking health medical records, biospecimens, and biosignals in Korean patients with Acute kidney injury) cohort.

AI, artificial intelligence; CKRT, continuous kidney replacement therapy.

Biospecimen collection procedures, quality control, and standard operating procedures

Biospecimen collection will be conducted as part of the prospective cohort study utilizing a standardized multicenter system for the collection of human-derived biospecimens. The plasma, serum, and urine samples will be collected and stored in cryotubes in a deep freezer (–80 °C). A minimum of five vials of plasma and serum will be stored, each with a volume of 300 μL. Whereas, a minimum of ten vials of urine samples will be stored, each with a volume of 1 mL.

Specimen collection points include at the time of CKRT initiation, at 1 week, 3 months (±1 month), and 1 year (±1 month) after initiation. This study aimed to standardize the collection and management of biospecimens across institutions by adhering to the Central Biobank operating procedures for biospecimen collection and management. Each blood and urine sample will be assigned a 7-digit SPREC (Standard PREanalytical Code) to encode pre-processing information for quality control of the biospecimens [14]. Visual checks for hemolysis will be conducted for serum and plasma samples. The collection and storage protocols have been designed to comply with the containers and standards specified for each type of resource. The reliability of the tests will be ensured via an external quality control conducted by the Korean Association of External Quality Assessment Services.

Biosignal collection procedures

Biosignal collection involves the measurement of vital signs of patients undergoing CKRT in the intensive care unit, including blood pressure, pulse rate, respiratory rate, body temperature, and oxygen saturation. These measurements will be recorded at minimum hourly intervals, with additional measurements performed as required. Electrocardiogram (ECG) and photoplethysmography (PPG) signals will be collected from monitoring devices such as Solar 8000M (GE Healthcare) and Tram-Rac4A (GE Healthcare) and stored at up to 500 Hz. The CKRT machines provide data on the pressure at four ports: ultrafiltration rate, clearance rate, blood flow rate, dialysate and replacement fluid rates, and transmembrane pressure, which is recorded every minute. They could offer real-time insights into both the machine’s operational status and the physiological response of the patient. Changes in machine settings will also be logged in minutes, providing a much more precise time unit compared to those recorded in the actual medical chart. Additionally, information on medication, treatment interventions, and outcomes, such as recovery from AKI and death, will be collected, ensuring ease of access, review, and annotation by researchers through organized database management.

Database development with de-identification

To support our research efforts, we developed a protocol for constructing a high-quality extensive database that leverages cloud-based technologies. We assembled a committee consisting of research experts and advisors to supervise this initiative with a specific focus on standardizing the extraction and storage of biosignal data, including vital signs, ECG, and PPG, in formats such as XML, HL7, and PDF. ECG data will be transformed for integration into HL7 standards in accordance with the ECG waveform data standard regulations certified by the American National Standards Institute. XML transformation facilitates the collection and refinement of electronic health information, thereby promoting compatibility and interoperability within multicenter healthcare systems. Thus, we can efficiently manage data and detect anomalies in real-time using advanced XML decryption and HL7 format processing coupled with real-time server-based monitoring.

In compliance with the guidelines set forth by the Ministry of Health and Welfare, our database development process will incorporate rigorous de-identification measures to safeguard the participants’ privacy. By implementing these techniques, patient information such as names, registration numbers, and visit information collected from various hospitals will be pseudonymized, and data that could potentially identify individuals will be either aggregated, reduced, or masked. This will ensure compliance with the Personal Information Protection Act, and that the database encompassing clinical information, along with unstructured biosignal data, undergoes a rigorous process of de-identification and standardization, thereby establishing a comprehensive and secure repository of health information.

Development of prediction model

We initially utilized logistic and Cox proportional hazards models to identify the risk factors influencing the progression of AKI to AKD, CKD, or ESKD, as well as the risk of mortality. To enhance our analysis, we included interaction terms and time-varying covariates to enable a more nuanced assessment of how risk factors evolve over time or in response to specific interventions. Furthermore, by examining the subgroups within our cohort, we aim to identify disparities in disease progression that could inform more targeted treatment approaches. By leveraging these conventional statistical models, we aimed to identify significant predictors to potentially guide early intervention and inform personalized treatment strategies.

Following the statistical analysis, we will further explore the utility of advanced artificial intelligence (AI) models to predict outcomes. This analysis will incorporate various AI techniques, including gradient boosting machines and graph convolutional networks, to process and analyze a wide spectrum of clinical data, ranging from vital signs to laboratory test results. To address the issue of missing data, which can significantly affect prediction accuracy, we will employ interpolation methods such as forward-fill and gated recurrent unit networks [15]. The most appropriate AI model for prediction will be selected based on its performance and undergo rigorous validation using data from both retrospective and prospective cohorts to ensure robustness and reliability. To enhance the model explainability, we will utilize methods such as Shapley Additive Explanations, providing insights into the relevance and impact of various input features on model prediction [16].

Development of reinforcement learning model

Currently, we are developing a reinforcement learning model with the primary objective of reducing progression to AKD, CKD, and ESKD. The reward function is designed to minimize disease progression via patient-specific interventions such as adjusting CKRT settings and administering vasoactive drugs. It integrates observational features from blood tests and biosignals for a holistic assessment of the treatment impact.

To refine this reinforcement learning, we employed an offline learning method, transformer models such as decision transformers and trajectory transformers, that can effectively handle the intricacies of our dataset [17,18]. Additionally, we considered the use of dead-end discovery methods to enhance the explanatory power of variables and screen for high-risk patient groups [19]. This method identifies optimal treatments by analyzing preprocessed data for key variables, ensuring the adaptability and accuracy of the model, with the ultimate goal of reducing mortality and improving renal recovery through personalized CKRT interventions.

Development of guidelines

The development of guidelines for the prevention and management of AKD and its subsequent disease spectrum involves a collaborative decision-making process. The guidelines aim to define disease characteristics, assess epidemiology and risk factors, and categorize and evaluate underlying causes. They will also include comprehensive management strategies covering various aspects such as blood test types and frequency, inpatient and outpatient care, fluid and nutrition treatment, and pharmacological and dialytic interventions. Additionally, the guidelines address options for CKRT, criteria for initiation and termination, preventive measures, and the evaluation and management of the transition to AKD and CKD.

The methodology for developing these guidelines involves several steps. First, a research group is formed which determines the development strategies, including whether to develop guidelines from scratch or adapt existing ones. In this study, prior guidelines were evaluated and core questions were selected using the PICO (Patient, Intervention, Comparison, and Outcome) framework. Literature reviews were conducted using specific databases and search strategies, and the research findings were evaluated, including meta-analyses, where applicable. Guideline recommendations were drafted based on the evidence and a consensus on the evidence level was reached. The draft guidelines underwent an external review and approval to ensure robustness, evidence-based content, and practicality for clinical applications. This structured approach aimed to standardize care for patients with AKD, thereby promoting optimal outcomes and preventing progression to CKD and ESKD.