Application of machine learning-based phenotyping in individualized fluid management in critically ill patients with heart failure

Heart failure (HF) is a major public health challenge, with fluid management being one of the most critical aspects of treatment. Fluid management is particularly a complex and challenging issue in critically ill patients, especially when cardiac pump function fails to meet the body's needs.^1-3 Clinicians often face multiple challenges when formulating fluid management strategies, including significant individual variations, complex dynamic changes, and diverse monitoring indicators. Most current intervention studies targeting fixed fluid management in HF patients have reported negative outcomes,^{4, 5} reflecting the heterogeneity of severe HF patients and highlighting the urgent need for precision medicine. Therefore, our study aims to identify distinct characteristics of critically ill HF patients through retrospective analyses and develop targeted treatment strategies based on the optimal fluid balance ranges identified by longitudinal infusion data for each patient phenotype (Figure 1).⁶

The advancement of artificial intelligence and machine learning (ML) technology offers innovative solutions to these challenges. Unsupervised ML has emerged as a powerful tool in medical research, capable of identifying patterns in complex, high-dimensional data without explicit labelling. The patient data were extracted from two intensive care unit databases, integrating both numerical and categorical variables to maintain comprehensive clinical characteristics. The K-Prototypes algorithm was selected for its ability to effectively combine the principles of K-Means and K-Modes principles, thereby enhancing clustering quality by considering the differential contributions of various variable types to the total distance between samples.⁷ Furthermore, fluid management is a dynamic process, where daily interventions and test results can affect subsequent outcomes. To address this, we analyzed 7-day fluid balance records using the G-formula parameter.⁸ a sophisticated statistical approach to eliminate confounding effects between time-varying exposures and outcomes, thus providing more reliable clinical guidance.

Our analysis identified four distinct phenotypes of HF patients, each exhibiting significant differences in clinical characteristics and prognosis. The optimal fluid balance ranges for each phenotype aligned closely with their distinct clinical features. Phenotype A, characterized by severe inflammation and aggressive interventions including high rates of vasoactive drug use and mechanical ventilation, showed optimal outcomes with a moderate fluid balance of between –1000 and 500 mL per day. This finding indicates that a positive fluid balance is associated with adverse effects on mechanical ventilation duration and mortality. Phenotype C, despite having milder clinical parameters but combined with advanced age and multiple comorbidities, demonstrated high mortality and required careful fluid restriction (–1500 to 500 mL daily), highlighting the significant impact of age and frailty on HF prognosis. Phenotype D, presenting with severe metabolic disorders including acidosis and renal insufficiency, required stricter fluid management ranging from –2000 to –500 mL per day.

To facilitate clinical application, we developed a streamlined classification approach using nine clinical indicators identified through feature screening: age, blood urea nitrogen, hematocrit, vasoactive drug use, renal disease, creatinine, diastolic blood pressure, mechanical ventilation status and anion gap. The XGBoost model demonstrated good predictive efficacy in both internal and external validation, with the area under the curve values ranging from 0.918 to 0.943 and from 0.802 to 0.907, respectively. A web-based typing tool was developed to facilitate rapid phenotype identification at the bedside, supporting prompt decision-making in fluid management.

The advancement in ML-based phenotyping and personalized fluid management strategies for HF has opened new avenues for research and development. While our current findings demonstrate promising results, further investigation and technological advancement are necessary in several key areas. First, prospective validation studies are essential and should include diverse patient populations and healthcare settings to ensure the broad applicability of findings. Second, the integration of novel biomarkers offers an opportunity to improve phenotype classification accuracy. Beyond conventional clinical parameters, novel molecular markers, genetic signatures and sophisticated imaging metrics can provide deeper insights into disease mechanisms and treatment responses. From a technological perspective, continuous refinement of predictive models remains paramount. The application of advanced ML architectures, including deep learning and ensemble methods, has the potential to improve prediction accuracy and model robustness.

In conclusion, our work provides a potential framework for the implementation of precision medicine in intensive care cardiology. By establishing population-level evidence, we provide a practical approach to personalized fluid management in critically ill HF patients. The combination of robust phenotypic identification and user-friendly web-based tools provides clinicians with the basis for implementing more targeted treatment strategies to improve patient outcomes.

Chengjian Guan wrote the original draft. Bing Xiao reviewed and edited the manuscript.

The authors declare no conflict of interest.

The project was supported by the S&T Program of Hebei No. 22377728D.

Not applicable.