Unsupervised clustering to classify patients with candidemia in intensive care: algorithmic tautology or the generation of highly anticipated new data?

We thank Drs Yang and Zhang for their thoughtful commentary [1] on our recent work [2]. Their critique raises important methodological considerations regarding unsupervised machine learning in critical care research, and we welcome the opportunity to clarify our approach. We address their two principal concerns (algorithmic tautology and the inclusion of SAPS II) with additional data demonstrating the robustness of our findings.

Algorithmic tautology or consolidation of a concept?

The real-life management of candidemia in the ICU remains imperfect, and the optimal strategy is unclear. The definition of complicated and uncomplicated candidemia is challenging and remains debated [3, 4]. We aimed to clarify decision-making beyond ‘clinical intuition’.

We therefore respectfully disagree with the assertion that our phenotypes merely reproduce ‘obvious clinical realities.’ Providing scientific proof of these intuitions remains both valuable and necessary. First, clinical intuition is inherently subjective and prone to bias, particularly for candidemia, a relatively uncommon infection in the ICU. Our data-driven approach provides quantitative, reproducible definitions of patient subgroups. Second, our analysis revealed unexpected microbiological associations: Candida glabrata was enriched in Phenotype 2 (24.2% vs. 15.4% overall) and Candida parapsilosis in Phenotype 3 (8.0% vs. 4.3% overall).

Third, despite fundamentally different underlying pathobiology (immunosuppression vs. cirrhosis), Phenotypes 1 and 2 have similarly high mortality, challenging the assumption that immunosuppression is the primary driver of poor outcomes. Fourth, and most importantly, our phenotypes provide a data-driven framework for phenotype-stratified clinical trials. Patients with Phenotype 3 may be ideal candidates for non-inferiority trials testing shorter antifungal courses or novel agents, whereas Phenotypes 1 and 2 may be more suitable for trials evaluating combination therapy or immunomodulatory strategies. Such clustering enabled us to design a multicenter target trial emulation showing that short-course antifungal therapy was not significantly associated with 90-day mortality in non-complicated candidemia [5].

Sensitivity analysis excluding SAPS II

The authors suggest that the inclusion of SAPS II may have artifactually driven phenotype separation and outcome differences, arguing that this constitutes a “self-fulfilling statistical prophecy.” We formally tested this hypothesis by performing a post-hoc sensitivity analysis repeating the entire clustering procedure with SAPS II explicitly excluded from the baseline feature set (Table 1).

Table 1.

Comparison of clustering results with and without SAPS II

Metric	With SAPS II (Original)	Without SAPS II
Total N	492	492
Cluster 1, n	70	77
Cluster 2, n	223	162
Cluster 3, n	199	253
Mortality across the three clusters
Mortality Cluster 1, %	72.9	70.1
Mortality Cluster 2, %	70.4	58.6
Mortality Cluster 3, %	50.3	62.8

In this analysis, phenotype 1 (characterized by hematological malignancy, neutropenia, high fever, and low leukocyte counts) remained robustly identifiable even without SAPS II and had similarly high mortality (70.1% vs. 72.9% in the original model). In addition, the distinction between Phenotypes 2 and 3 becomes slightly blurred in the absence of SAPS II, with a redistribution of patients between these groups. The exclusion of severity scores would therefore lead to this clinically relevant dimension of heterogeneity being missed. Finally, the microbiological associations originally observed (particularly C. glabrata enrichment in the cirrhotic/digestive cluster) persisted in the sensitivity analysis (31.5% vs. 24.2%), supporting the biological validity of our phenotypes beyond severity scoring.

On the inclusion of SAPS II in unsupervised clustering: a common practice in critical care

• The inclusion of composite severity scores, such as SAPS II or SOFA, in unsupervised clustering algorithms is a well-established practice in critical care research. Several recent studies have adopted this approach: In a multicenter cohort of critically ill patients with invasive group A Streptococcus infections, Urbina et al. included SOFA score to identifythree distinct clinical phenotypes [6].

• Similarly, in a study on necrotizing soft-tissue infections in a tropical setting, Pommier et al. included SAPS II in their unsupervised clustering analysis to identify three phenotypes associated with different outcomes [7].

These examples demonstrate that, rather than being a methodological flaw, the inclusion of severity scores is a legitimate approach to capturing the full spectrum of patient heterogeneity in critical illness.

As Seymour et al. demonstrated in their landmark work on sepsis phenotypes, baseline severity scores such as SOFA are legitimate descriptors when the goal is to characterize the entire clinical syndrome at presentation [8].

Perspectives for moving forward in this context

We have already acknowledged the limitations of our study, including its retrospective design and the absence of external validation. However, this analysis demonstrates that our phenotypes are not merely a statistical artifact of SAPS II weighting but reflect genuine patient heterogeneity. The observation that excluding SAPS II increases heterogeneity within the catheter-related cluster (through merging with cirrhotic patients) and obscures clinically meaningful differences in mortality supports the validity of our original phenotypes.We agree with Yang and Zhang that future studies should explore clustering without composite severity scores to identify more granular, biology-driven endotypes, and that transcriptomic, proteomic, or metabolomic biomarkers would enrich phenotype characterization.However, we maintain that illness severity is an intrinsic component of the clinical presentation in critical care and should not be dismissed as a “bias” when the goal is to characterize the entire clinical syndrome. We propose a pragmatic compromise: future studies could be performed with two-layer clustering: a first clustering on biological and microbiological variables (excluding severity scores) to identify disease-specific signatures, and a second clustering to stratify these clusters by severity. We believe that our framework constitutes a meaningful step toward precision medicine in the context of candidemia in the ICU.

Author contributions

F.R. RP and J.P.G. conceived, designed, and supervised the study and coordinated the data collection. F.R. performed the statistical analysis. F.R. and J.P.G. wrote the first draft of the article. All authors revised the manuscript and approved the final version of the manuscript.

Funding

No funding was received for this work.

Data availability

The datasets from this study are available from the corresponding author on request.

Declarations

Ethical approval

The study was carried out in accordance with the ethical standards of the Declaration of Helsinki, and the database was approved by the local ethics committee [Comité d’éthique CHU Rennes N° 23.99].

Competing interests

The authors report no conflict of interest related to this work JPG received honoraria from Gilead, MundiPharma, Pfizer, and Shionogi for presentations and conferences in congresses.