“Decoding candidemia in critically ill patients: unsupervised clustering identifies three unique phenotypes”: algorithmic tautology and feature selection bias Authors: Zejun Yang, Yaotian Zhang Published: 2026 To the Editor, We read with great interest the recent article by Reizine et al. published in Critical Care, which utilized unsupervised machine learning—specifically Factor Analysis of Mixed Data (FAMD) to navigate the statistical complexities of mixed datasets—to identify three unique clinical phenotypes of candidemia in critically ill patients [1]. The authors demonstrated that these three data-driven clusters exhibited markedly divergent 90-day mortality rates, providing a much-needed framework for stratifying highly heterogeneous populations in future clinical trials [1]. While we highly commend the authors’ ambition to leverage advanced computational methods, and acknowledge that incorporating established clinical risk scores into machine learning models has historical precedent, a rigorous critical appraisal of the study design reveals profound methodological vulnerabilities. Specifically, we harbor significant concerns regarding macroscopic algorithmic tautology and microscopic feature selection bias, which together compromise the true clinical utility of the identified phenotypes. From a macroscopic perspective, the translation of these computational clusters into actionable clinical intelligence is highly questionable. The unsupervised clustering essentially categorized the cohort into three intuitive groups: Phenotype 1 comprised severely immunosuppressed hematological patients with extreme illness severity; Phenotype 2 consisted predominantly of deteriorating cirrhotic patients with gastrointestinal sources of infection; and Phenotype 3 encapsulated a comparatively less severe cohort driven largely by catheter-related candidemia [1]. In the realm of critical care, it is a universally established axiom that profoundly neutropenic patients or patients with end-stage cirrhosis suffering from septic shock will experience drastically higher mortality (approximately 70% in Phenotypes 1 and 2) compared to relatively stable patients suffering from an easily removable infected central venous catheter (50% in Phenotype 3). Deploying sophisticated machine learning algorithms to “discover” that fundamentally distinct etiologies and baseline morbidities carry vastly divergent prognoses borders on clinical tautology. Without unmasking novel, non-obvious pathobiological endotypes or directly translating these clusters into distinct, phenotype-specific therapeutic algorithms, the model merely rebrands basic bedside intuition with complex mathematical nomenclature. As the intensive care community increasingly adopts artificial intelligence, the standard for clinical utility must move beyond recapitulating obvious clinical realities to generating genuinely actionable, counter-intuitive insights [2]. Microscopically, the striking divergence in survival curves among the three phenotypes is largely a statistical artifact driven by severe feature selection bias. The authors explicitly stated that they excluded outcome variables and post-baseline therapeutic interventions from their clustering algorithm to prevent outcome-driven grouping [1]. However, this safeguard was entirely nullified by the simultaneous inclusion of massive surrogate markers of mortality as baseline clustering features—most notably, the Simplified Acute Physiology Score II (SAPS II), the presence of septic shock, and the necessity for early intubation within 24 h [1]. We acknowledge that the authors focused specifically on candidemia outcomes, where SAPS II arguably serves as a validated baseline factor. However, its inclusion in an unsupervised model designed to predict mortality trajectories constitutes a severe methodological flaw. SAPS II is not a neutral biological descriptor; it is a highly validated, composite mathematical construct specifically engineered to predict ICU mortality by aggressively weighing extreme physiological derangements (e.g., Glasgow Coma Scale, profound hypotension, severe hypoxia) and severe comorbidities [3]. By feeding SAPS II into the FAMD algorithm as a primary baseline feature, the authors inadvertently forced the machine learning model to separate the cohort based on brute physiological severity rather than unmasking any latent, candidemia-specific biological signatures. Consequently, it is an absolute mathematical certainty that the resulting clusters would exhibit significantly divergent survival curves. The outcome differences (70% vs. 50% mortality) are not a revelation of hidden disease subtypes; rather, they are a self-fulfilling statistical prophecy orchestrated by the heavy weighting of pre-existing severity scores embedded in the input data. To achieve true precision medicine in sepsis and candidemia, future unsupervised machine learning models must boldly strip away overarching, mortality-tethered clinical scores (like SAPS II, SOFA, or APACHE) from their baseline feature sets [4]. We propose a pragmatic alternative mechanism: these composite severity scores should be excluded from baseline FAMD clustering and instead applied post-hoc to externally validate the clinical relevance of the newly discovered clusters, or utilized as independent covariates in downstream supervised predictive models. Only by relying purely on unbiased clinical, transcriptomic, or biomarker profiles can these algorithms transcend tautology and discover genuinely novel biological endotypes that pave the way for targeted immunomodulatory or antifungal therapies. Acknowledgements Not applicable. Abbreviations Artificial Intelligence Acute Physiology and Chronic Health Evaluation Factor Analysis of Mixed Data Glasgow Coma Scale Intensive Care Unit Simplified Acute Physiology Score II Sequential Organ Failure Assessment Author contributions YZ and ZY conceptualized the core arguments and drafted the manuscript. Both authors read and approved the final manuscript. Funding This research received no external funding. Data availability No datasets were generated or analysed during the current study. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement No datasets were generated or analysed during the current study.