Observational multicentre registry to update key interventions and new approaches with artificial intelligence for heart failure (OKINAWA-HF): a retrospective medical record-derived registry study protocol Authors: Yoshino Kinjo, Haruno Nagata, Yuichiro Toma, Chikashi Nago, Mamoru Manita, Masanori Kakazu, Minoru Wake, Moriichi Sugama, Osahiko Sunagawa, Shinichiro Ueda, Kenya Kusunose Published: 2026 ABSTRACT Abstract AbstractIntroductionThe efficacy of guideline-directed medical therapy (GDMT) for heart failure has been established in clinical trials, but trial populations differ substantially from real-world clinical practice. Consequently, the real-world status of GDMT implementation, the patient-level factors that determine whether guideline-recommended targets are achieved and the clinical circumstances that limit optimisation remain poorly characterised. We designed the Observational Multicenter Registry to Update Key Interventions and New Approaches with Artificial Intelligence for Heart Failure registry, a retrospective multicentre study based on medical records from seven teaching hospitals in Okinawa, Japan, to address a central scientific question: how GDMT is implemented at discharge in routine care for patients hospitalised with heart failure and which patient-level and treatment-limiting factors explain incomplete implementation. Analyses of postdischarge outcomes and artificial intelligence (AI)-derived imaging variables are explicitly positioned as secondary and exploratory components.Methods and analysisConsecutive patients hospitalised for new-onset or worsening heart failure will be enrolled. Clinical characteristics, laboratory data, medication use, chest X-ray and echocardiographic images will be collected. The primary analyses will describe patterns of GDMT implementation, target-dose achievement and documented reasons for non-implementation and examine patient-level clinical characteristics and treatment-limiting factors associated with incomplete GDMT implementation and target-dose non-achievement. As a secondary objective, associations between guideline-based GDMT optimisation and postdischarge clinical outcomes will be examined descriptively and as hypothesis-generating analyses. Exploratory analyses will evaluate AI-derived imaging variables obtained from routinely acquired chest X-ray and echocardiography as standardised imaging phenotypes and assess their potential incremental prognostic value beyond conventional clinical and imaging variables.Ethics and disseminationThe study is being conducted in accordance with the Declaration of Helsinki. Ethical approval was obtained through a centralised ethics review system at the University of the Ryukyus Hospital Ethics Committee, which served as the central reviewing body for this multicentre study. Under this centralised review framework, all participating institutions conducted the study based on this single ethical approval. On completion of data collection and analyses, the findings of this study will be disseminated through academic and clinical conferences, reports to funding agencies and publications in peer-reviewed journals. We will also conduct dissemination events to report the study findings and support their application in regional clinical practices.Keywords: Observational Study, Heart failure, Treatment Outcome, Echocardiography STRENGTHS AND LIMITATIONS OF THIS STUDY This retrospective study will use data from consecutively enrolled patients hospitalised for new-onset or worsening heart failure, which may help reduce selection bias and improve the representativeness of real-world clinical practice and allow the inclusion of patient groups that are often under-represented in clinical trials. This retrospective study will use data from consecutively enrolled patients hospitalised for new-onset or worsening heart failure, which may help reduce selection bias and improve the representativeness of real-world clinical practice and allow the inclusion of patient groups that are often under-represented in clinical trials. In this study, a registry constructed directly from medical records may provide more detailed characterisation of patient characteristics and allow more direct assessment of clinical outcomes. In this study, a registry constructed directly from medical records may provide more detailed characterisation of patient characteristics and allow more direct assessment of clinical outcomes. Artificial intelligence-based imaging analyses may provide standardised and reproducible imaging-derived variables, potentially reducing interhospital variability in image interpretation. Artificial intelligence-based imaging analyses may provide standardised and reproducible imaging-derived variables, potentially reducing interhospital variability in image interpretation. As a retrospective observational study, the influence of unmeasured or unmeasurable confounders cannot be fully excluded from the analysis. As a retrospective observational study, the influence of unmeasured or unmeasurable confounders cannot be fully excluded from the analysis. Because the cohort is restricted to the Okinawa region of Japan, the generalisability of the findings will require cautious interpretation. Because the cohort is restricted to the Okinawa region of Japan, the generalisability of the findings will require cautious interpretation. Introduction Heart failure is one of the leading causes of death worldwide.1 It is highly prevalent and often has a poor long-term prognosis associated with frequent hospital readmission.1 2 Heart failure also leads to a decline in patients’ quality of life. In this context, appropriate management of heart failure is important for improving patient outcomes. In recent years, clinical trials have demonstrated that several new drugs improve outcomes for heart failure patients, alongside conventional heart failure medications. Consequently, these therapies are now recommended in clinical guidelines for heart failure management.3,6 However, clinical trial results must be interpreted with caution because the patient profiles in clinical trials do not necessarily correspond with those encountered in real-world clinical practice. For example, concerns have arisen in real-world settings regarding the safety of spironolactone, as demonstrated in the severe heart failure clinical trial.7 8 Heart failure patients are predominantly elderly and have multiple comorbidities, such as renal dysfunction, and consequently, take many additional medications based on evidence-based guidelines. Adverse drug events are more likely to occur in such patients, leading to drug administration being interrupted or discontinued. As a result of these concerns, there is a possibility that guideline-directed medical therapy (GDMT) may be underused, recommended drugs may not be prescribed or target doses may not be achieved. These issues highlight an unmet need to better understand real-world GDMT use and its association with patient outcomes in real-world clinical practice. Given that there are several difficult-to-resolve challenges in performing clinical trials for heart failure management, we designed a retrospective registry study enrolling consecutive patients hospitalised for new-onset or worsening heart failure across seven teaching hospitals in Okinawa. The central scientific question of the Observational Multicenter Registry to Update Key Interventions and New Approaches with Artificial Intelligence for Heart Failure (OKINAWA-HF) registry is to determine how guideline-directed medical therapy is implemented at discharge in routine clinical practice among patients hospitalised for new-onset or worsening heart failure and to characterise patient-level clinical characteristics and treatment-limiting factors associated with incomplete implementation. Accordingly, the primary analyses are descriptive analyses of GDMT implementation, target-dose achievement and clinically documented reasons for non-implementation or non-intensification, and examine patient-level clinical characteristics and treatment-limiting factors associated with incomplete GDMT implementation and target-dose non-achievement. Secondary analyses will describe postdischarge clinical outcome patterns according to prespecified GDMT implementation categories. The purpose of these analyses is to characterise real-world implementation patterns of guideline-recommended therapies rather than to assess the quality or appropriateness of individual treatment decisions. Exploratory analyses will evaluate artificial intelligence (AI)-derived imaging variables obtained from routinely acquired chest X-ray and echocardiography as standardised imaging phenotypes. In addition, we will assess whether these AI-derived phenotypes provide incremental prognostic information beyond conventional clinical and imaging variables. Thus, the AI-based analyses are not a separate primary objective but rather an exploratory extension intended to support standardised phenotyping and generate hypotheses for future investigation. Methods and analysis Study design and setting The OKINAWA-HF study will be conducted as a retrospective registry study consecutively enrolling patients hospitalised for new-onset or worsening heart failure at seven teaching hospitals in Okinawa. At the study planning stage, we estimated that approximately 2500 patients would be enrolled based on the annual case volume of the participating centres and the predefined enrolment period. Previous Japanese heart failure registries have reported rates of all-cause mortality and readmission for worsening heart failure after discharge that correspond to an anticipated event rate of approximately 30% during follow-up, which was considered sufficient to support the planned multivariable analyses.2 9 Patients and registration The administrative study period, including data collection, follow-up ascertainment and analysis, runs from April 2025 to December 2030. We plan to enrol patients hospitalised for new-onset or worsening heart failure diagnosed according to the Framingham Heart Study criteria.10 Eligible index hospitalisations will include those occurring between April 2021 and March 2025 (figure 1). Hospitalisation for heart failure will be defined by the initiation of heart failure treatments, including diuretics, inotropic agents, vasodilators or emergency dialysis within 1 week after admission. If patients experienced multiple hospitalisations for heart failure during the enrolment period, the index hospitalisation will be taken as the first hospitalisation. The observational follow-up period will commence from the index hospitalisation date and continue until death, transfer to another facility at which subsequent outcomes could not be ascertained, the last documented clinical contact or December 2030, whichever occurs first (figure 2). For analyses based on GDMT at discharge, follow-up will begin at discharge from the index hospitalisation and continue until the first occurrence of the outcome of interest, transfer to another facility at which subsequent outcomes could not be ascertained, the last documented clinical contact or December 2030, whichever occurs first. Patient data will be extracted from each hospital’s medical information management system in accordance with these diagnoses and definitions. Relevant patient information will be obtained directly from the medical records. A flow chart of the enrolment of patients with heart failure, the inclusion and exclusion criteria and the outcome measures evaluated in this study. If patients had multiple hospitalisations for heart failure during the eligible index hospitalisation period, the index hospitalisation will be defined as the first hospitalisation. The timeline of data collection activities, guideline-directed medical therapy (GDMT) assessments and subsequent outcome evaluations, along with the timeline of routine clinical care. Data collection Variables including patient characteristics, comorbidities, relevant clinical data, treatments, chest X-ray and echocardiographic images will be collected (online supplemental table 1 and figure 2). In addition to treatment-related variables, data relevant to congestion phenotype, right-heart involvement and cardiorenal interactions will be collected when available. Congestion-related variables will include natriuretic peptide levels, loop diuretic use and dose, chest X-ray findings, AI-derived chest X-ray estimates of pulmonary artery pressure and pulmonary artery wedge pressure and echocardiographic indices such as E/e′, left atrial size and inferior vena cava findings. Right-heart involvement will be assessed using available echocardiographic variables, including right ventricular size and function, tricuspid regurgitation severity, estimated pulmonary artery pressure and inferior vena cava findings. Cardiorenal variables will include serum creatinine, serum potassium, dialysis status and diuretic use. These variables will be included as covariates or used in stratified or exploratory analyses, depending on data availability and missingness. These data will be extracted from the electronic medical records and collected and managed using Research Electronic Data Capture (REDCap) tools hosted at the University of the Ryukyus Hospital.11 12 REDCap is a secure, web-based software platform designed to support data capture for research studies, providing (1) an intuitive interface for validated data capture; (2) audit trails for tracking data manipulation and export procedures; (3) automated export procedures for seamless data downloads to common statistical packages and (4) procedures for data integration and interoperability with external sources. The data entered into the REDCap system are deidentified, with site-specific study IDs retained and the reidentification key stored only at each originating site. Access to the anonymised data for analysis will be restricted to the principal investigator and a prespecified core analysis team. Collaborators at each hospital will have access only to their own institutional data, limited to data entry, verification and quality control; cross-site record-level access will be prohibited. Image data, including echocardiographs, chest X-rays and electrocardiograms, will be sent to the University of the Ryukyus for centralised evaluation and AI-assisted analysis. Personal identifiers, including names, dates of birth and medical IDs, will be removed from all images to ensure patient anonymity. Methods of data transfer will be determined according to each facility’s circumstances, using secure cloud storage or hardware-encrypted physical media with passwords exchanged via a separate channel. No identifiable or site-level raw data will be shared outside the study team; only aggregated results will be disseminated. Assessment of GDMT implementation, target-dose achievement and guideline-based GDMT optimisation GDMT will be defined as the use of renin-angiotensin system inhibitors (angiotensin-converting enzyme inhibitors, angiotensin receptor blockers and angiotensin receptor-neprilysin inhibitors), beta blockers, mineralocorticoid receptor antagonists and sodium-glucose cotransporter 2 inhibitors.13 Using data from the OKINAWA-HF registry, we will assess GDMT implementation based on the number and specific combinations of GDMT classes prescribed at discharge. Target-dose achievement is operationalised as achievement of guideline-recommended target or maintenance dose and serves as a reference standard for comparison across patient subgroups and institutions.14 Treatment-limiting factors will be abstracted from medical records according to a predefined framework, including hypotension or low systolic blood pressure, bradycardia, renal dysfunction or dialysis, hyperkalaemia, documented drug intolerance or adverse events, acute clinical instability and frailty-related vulnerability indicators when available. Guideline-based GDMT optimisation is operationally defined using two prespecified indicators: (1) implementation of guideline-recommended GDMT classes at discharge and (2) achievement of guideline-recommended target doses at discharge (figure 2). Therefore, all patients who failed to meet the predefined optimisation criteria will be classified as non-optimised regardless of the underlying reason. Guideline-based GDMT optimisation is intended to characterise real-world implementation patterns of guideline-recommended therapies rather than to assess the quality or appropriateness of individual treatment decisions. Consequently, study findings will be interpreted as reflecting the overall prognostic implications of GDMT optimisation status in routine clinical practice, encompassing both the effects of incomplete GDMT implementation and the influence of clinical conditions that limit optimisation rather than the isolated pharmacological effect of GDMT itself. Because eligibility for and recommendations regarding GDMT differ across heart failure phenotypes, all descriptive and outcome analyses will be performed separately for the heart failure phenotype classified according to left ventricular ejection fraction (LVEF). LVEF will be categorised according to current guideline definitions as heart failure with reduced ejection fraction (LVEF ≤40%), mildly reduced ejection fraction (LVEF 41%–49%) and preserved ejection fraction (LVEF ≥50%).13 AI-based imaging analyses Based on preliminary assessments of data availability across participating centres, admission chest X-ray and echocardiographic images were expected to be available for more than 90% of enrolled patients, whereas discharge imaging studies were available in a smaller proportion of patients. These availability rates were considered sufficient to support the planned AI-based imaging analyses. The AI-based analyses will use pre-existing, non-commercial, in-house deep-learning models previously developed and reported by Kusunose and colleagues.15,17 In addition, the commercially available automated analysis platform (Us2.ai) will be used to obtain a comprehensive set of echocardiographic parameters without manual input. Us2.ai is a Food and Drug Administration-cleared, European Conformity-marked software for fully automated echocardiographic analysis.18 The AI component is included as an exploratory extension of the registry and is not part of the primary objective. Its conceptual rationale is threefold. First, AI-based analysis may provide standardised imaging-derived variables from routinely acquired chest X-ray and echocardiography, thereby reducing interhospital variability in image interpretation. Second, AI-derived variables may support imaging-based phenotyping of congestion, pulmonary haemodynamics, ventricular function and right-heart involvement, which are clinically relevant to GDMT tolerance and treatment decision-making. Third, these variables may provide exploratory prognostic enrichment beyond conventional clinical and imaging variables. The AI analyses are not intended to guide treatment decisions, adjudicate clinical outcomes or replace routine clinical imaging interpretation in the primary analyses. All in-house AI models will be applied as fixed, previously developed algorithms without retraining using outcome data from OKINAWA-HF registry. To assess model robustness and external validity in this distinct population, we will report the proportion of successfully analysed images and perform quality-control procedures in a randomly selected subset of cases. In this subset, AI-derived measurements will be compared with expert manual readings. Agreement for continuous measurements will be assessed using intraclass correlation coefficients, mean absolute error and Bland–Altman analyses, whereas agreement for categorical outputs will be assessed using kappa statistics and discrimination metrics, as appropriate. In addition, the consistency of AI-derived measurements will be examined across relevant demographic and imaging subgroups. Because these AI models are used as standardised measurement tools rather than being developed or updated within this registry, formal model redevelopment and external predictive validation are beyond the scope of the present study. Findings derived from AI-based analyses will therefore be interpreted as exploratory and hypothesis-generating. To address the large number of AI-derived imaging variables, prespecified clinically relevant parameters, including estimates of pulmonary haemodynamics and ventricular function, will be prioritised for evaluation. Findings involving a broader set of AI-derived variables will be interpreted cautiously in light of the potential for multiplicity and false-positive findings. Chest X-ray assessments Chest X-ray data obtained both near hospital admission and before discharge will be collected when available. Standard digital chest radiographs (posterior–anterior) will be acquired from hospital admission data and archived in accordance with the Digital Imaging and Communications in Medicine standard for central analysis. Acquisition will follow routine clinical practice; view position, patient posture (erect/semierect/supine) and exposure parameters will be recorded when available. Chest X-ray images will be analysed using pre-existing, previously developed in-house deep-learning models15 19 20 to obtain AI-derived estimates related to cardiopulmonary haemodynamics. AI-derived parameters will include risk scores (percentages) for the following measures: (a) elevated pulmonary artery pressure,19 (b) elevated pulmonary artery wedge pressure20 and (c) heart failure-related adverse events,15 as defined in the previous publications. Echocardiography assessments Transthoracic echocardiographic data obtained both near hospital admission and before discharge will be collected when available. Echocardiography was performed using commercially available ultrasound machines. Basic echocardiographic parameters were measured as recommended by the guidelines.21 Echocardiographic images will be analysed using two independent AI-based systems in this study. First, pre-existing, previously developed in-house deep-learning models will be used to obtain AI-derived estimates of (i) regional wall motion abnormalities and (ii) LVEF.16 17 Second, Us2.ai will be used to obtain a comprehensive set of echocardiographic parameters without manual input. It employs advanced deep learning algorithms to generate a comprehensive set of cardiac measurements (over 45 echocardiographic parameters, including standard indices and strain analysis) without manual input. It is noted that the primary classification of heart failure phenotype will be based on the clinically reported LVEF documented in the routine echocardiography report closest to the index hospitalisation. Accordingly, AI-derived measurements will be treated as supplementary standardised imaging parameters rather than replacements for routine clinical assessment. These parameters will be used to evaluate the incremental explanatory and prognostic value of AI-based imaging metrics beyond conventional clinical and imaging assessments. Outcome measures The primary outcome will be defined as a composite of readmission for worsening heart failure and all-cause mortality occurring after discharge. Secondary outcomes will include readmission for worsening heart failure or all-cause mortality, evaluated separately after discharge. Other secondary outcomes will include myocardial infarction, stroke, end-stage renal disease and cardiovascular death. The diagnosis of myocardial infarction and stroke will be based on the WHO’s criteria. End-stage renal disease will include the initiation of dialysis or kidney transplantation. Cardiovascular death will include sudden cardiac death and death attributed to acute coronary syndrome, heart failure, arrhythmia, stroke, aortic dissection, ruptured aortic aneurysm or pulmonary vascular embolism. Clinical outcomes will be initially abstracted from medical records at each participating hospital using standardised REDCap case report forms and prespecified outcome definitions. Site investigators will classify outcome events according to these standardised definitions. Independent adjudication by a central clinical events committee is not planned because of the retrospective registry design. Although outcome adjudication will be performed locally, the use of prespecified definitions and standardised case report forms is expected to reduce intersite variability in event classification. Nevertheless, some degree of residual variability cannot be excluded. Analyses Analyses are organised according to the conceptual hierarchy of study objectives. Descriptive analyses and analyses of predictors of GDMT implementation and target-dose achievement constitute the primary analyses. Survival analyses evaluating associations between guideline-based GDMT optimisation and clinical outcomes, and exploratory AI-imaging analyses, are secondary and exploratory and are subject to the confounding limitations inherent to observational study designs. Descriptive statistics Descriptive statistics will be used to summarise patient characteristics, medical history, key laboratory tests at admission and discharge, cardiac function and medications, including GDMT implementation and target-dose achievement. Outcomes will include the number of events and incidence rates for readmission for worsening heart failure after discharge, death, cardiovascular death, myocardial infarction, stroke and end-stage renal disease. Categorical variables will be expressed as counts and proportions and will be compared by using χ2 tests. Continuous variables will be expressed as means and SDs or medians with IQRs and compared between groups by using student’s unpaired t-tests or Wilcoxon rank-sum tests, as appropriate. P values <0.05 will be considered statistically significant, with 95% CIs calculated in all analyses. Predictors of GDMT implementation and target-dose achievement Multinomial logistic regression analyses will be performed to examine associations between clinical variables and GDMT implementation patterns. Logistic regression analyses will be used to evaluate factors associated with target-dose achievement. In this study, missing data patterns and proportions will be summarised for all candidate variables. Complete-case analysis will be used when missingness for all model variables is below 5% and results in the exclusion of fewer than 10% of eligible participants. If missingness in variables prespecified for multivariable adjustment is ≥5% or complete-case analysis would exclude ≥10% of eligible participants, multiple imputation by chained equations will be used as the primary approach. Variables with more than 30% missing data will not be included in the primary multivariable models. Outcome variables, including dates of events, will not be imputed. Survival analysis Patients will be categorised according to predefined guideline-based GDMT optimisation patterns based on discharge GDMT implementation and target-dose achievement at discharge. Associations with postdischarge clinical outcomes across these GDMT optimisation categories will be evaluated using the Kaplan–Meier method, log-rank test and Cox proportional hazards models. Adjusted HRs will be interpreted as measures of association rather than estimates of causal treatment effects. No causal estimand will be specified for these analyses, and the results will not be used to conclude that GDMT implementation or dose achievement independently improves prognosis. This cautious interpretation is required because GDMT initiation and intensification are influenced by clinical status, frailty, renal dysfunction, hypotension, hyperkalaemia, drug intolerance and physician judgement, which may be incompletely captured in retrospective medical records. Missing data handling for survival analyses will follow the same approach described above. To account for potential between-centre heterogeneity, Cox proportional hazards models stratified by study centre will be used in the primary time-to-event analyses. Sensitivity analyses will include mixed-effects Cox models or shared frailty models with study centre as a random effect. For analyses of heart failure readmission alone, death without prior readmission will be treated as a competing event in sensitivity analyses using Fine–Gray subdistribution hazard models. Exploratory analyses of AI-derived imaging variables To evaluate the incremental explanatory and prognostic value of AI-derived imaging parameters beyond conventional clinical and imaging assessment, multivariable models with and without AI-derived variables will be compared. Incremental model performance will be assessed using discrimination, calibration and model-fit indices, including changes in the C-statistic or time-dependent area under the receiver operating characteristic curve for time-to-event outcomes, likelihood ratio tests and reclassification metrics, as appropriate. These analyses will be exploratory in nature. Patient and public involvement None. Ethics and dissemination The study is being conducted in accordance with the Declaration of Helsinki. Ethical approval was obtained through a centralised ethics review system at the University of the Ryukyus Hospital Ethics Committee, which served as the central reviewing body for this multicentre study (approval number: 24–2357, approved on 30 September 2024). Under this centralised review framework, all participating institutions conducted the study based on this single ethical approval. Owing to the retrospective nature of the study and the use of deidentified data, the requirement for written informed consent has been waived. An opt-out consent procedure will be implemented, whereby information about the study will be publicly disclosed on the websites of the participating institutions. The findings of this study will be disseminated through academic and clinical conferences, reports to funding agencies and publications in peer-reviewed journals. We will also conduct dissemination events to report the study findings and support their application in regional clinical practices. The reporting of the study’s results will adhere to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. Discussion We designed the OKINAWA-HF registry as a retrospective multicentre study based on medical records from hospitals in Okinawa. This study will use a consecutive enrolment design, in which all patients hospitalised for new-onset or worsening heart failure are included in a predefined period, which may help mitigate selection bias and enhance the generalisability to real-world clinical practice. Several clinical trials have previously demonstrated the efficacy of heart failure medications on clinical outcomes in chronic heart failure populations. In major clinical trials, such as Empagliflozin Outcome Trial in Patients With Chronic Heart Failure With Reduced Ejection Fraction (EMPEROR-Reduced), the Dapagliflozin and Prevention of Adverse Outcomes in Heart Failure (DAPA-HF) and the Prospective Comparison of ARNI with ACEI to Determine Impact on Global Mortality and Morbidity in Heart Failure (PARADIGM-HF), the mean age of participants ranged from 64 to 67 years.4,6 Moreover, these trials generally excluded patients with comorbidities, such as impaired renal function and hypotension, from enrolment. However, in real-world clinical practice, patients with heart failure tend to be older and have multiple comorbidities. Consequently, evidence regarding the efficacy and safety of GDMT in such patients remains limited. Given the limitations of clinical trials, we designed a retrospective registry study. However, insurance claims data, often used in real-world database studies, present several limitations. These limitations include the fact that the medications and tests administered are merely confirmed based on claim records; that diagnoses are identified by administrative codes, but validation of these codes is essential and that outcome measurements rely on indirect algorithms. These limitations are particularly problematic in heart failure research. Although a registry constructed from medical records may yield fewer cases compared with insurance claims databases, it may provide more detailed characterisation of patient characteristics and allow more direct outcome assessment. Therefore, we designed a registry constructed directly from medical records. Chest X-ray and echocardiography are essential for the non-invasive assessment of cardiac structure and haemodynamics in patients with heart failure. However, it is well known that their interpretation requires specialised expertise and often varies across facilities.22 Recent advances in AI-driven technologies can help identify subtle changes that may be overlooked by human observers, thereby providing valuable support for clinical decision-making.23 Additionally, deep learning approaches for analysing chest X-ray images have demonstrated their ability to add prognostic information to other conventional clinical prognostic factors in heart failure.15 In this study, AI-based analyses are not intended to guide treatment decisions, adjudicate clinical outcomes or replace routine clinical imaging interpretation in the primary analyses. These analyses may provide a standardised extraction of imaging-derived markers that may reflect congestion, cardiopulmonary haemodynamics, ventricular function and right-heart involvement across participating hospitals. In addition, exploratory analyses may assess whether AI-derived phenotypes provide incremental prognostic information beyond conventional clinical and imaging variables. Limitations of the study This study will have several limitations. First, as a retrospective observational study, the influence of unmeasured or unmeasurable confounders cannot be fully excluded from the analysis. An inherent limitation of retrospective cohort studies is that data collection is restricted to what has already been recorded; perfect ascertainment is therefore impossible, and missing values are inevitable. To address this, we specifically selected variables with high documentation rates in medical records and used extended intervals for longitudinal data analysis to minimise inconsistencies. Second, although multivariable adjustment, centre-stratified models and sensitivity analyses will be performed, residual confounding and confounding by indication are unavoidable. In particular, patients who receive more complete GDMT or higher doses may differ systematically from those who do not with respect to clinical stability, frailty, renal function, blood pressure, treatment tolerance and clinician judgement. Therefore, associations between GDMT implementation and outcomes should not be interpreted as causal effects. Third, because the cohort will be restricted to Okinawa, the generalisability of the findings to other regions of Japan or to non-Japanese populations will require cautious interpretation. Strengths of the study This study will have several strengths. First, the use of a consecutive enrolment design may help reduce selection bias, improve the representativeness of real-world clinical practice and allow the inclusion of patient groups that are often under-represented in clinical trials. Second, a registry constructed directly from medical records may provide more detailed characterisation of patient characteristics and allow more direct assessment of clinical outcomes. Third, AI-based imaging analyses may provide standardised imaging assessment, potentially reducing interhospital variability in image interpretation. Finally, Okinawa is a geographically defined island region with a relatively homogeneous healthcare system, where hospital care for heart failure is concentrated in a limited number of institutions. This regional focus may reduce interinstitutional variability in practice patterns and thereby improve the consistency of observed associations. Supplementary material Acknowledgements We thank Andrew Hunt, PhD, from Edanz (https://jp.edanz.com/ac) for editing a draft of this manuscript. Footnotes Funding: This study is supported by a grant from the Ministry of Education, Culture, Sports, Science and Technology (MEXT) project for the promotion of clinical research and education to SU and KK (Grant No. GPR6-11) and Japanese Circulation Society Grant for Cardiovascular AI Research to KK (Grant No. 2024). Prepublication history and additional supplemental material for this paper are available online. To view these files, please visit the journal online (https://doi.org/10.1136/bmjopen-2026-117452). Provenance and peer review: Not commissioned; externally peer reviewed. Patient consent for publication: Not applicable. Patient and public involvement: Patients and/or the public were not involved in the design, conduct, reporting or dissemination plans of this research. References Supplementary Materials