Invited Commentary <\/p>\n\n\n\n
Health Informatics<\/p>\n\n\n\n
April 7, 2025<\/p>\n\n\n\n
One of the most difficult medical challenges we have faced for decades is regarding the human host response to infectious and noninfectious insults. We are still lacking understanding on what constitutes adaptive and maladaptive host responses as well as on how the host response itself alters (or not) the survival outcomes related to these insults. For example, an exuberant inflammatory response is common after major surgical procedures, such as a cardiac bypass surgery, or after a major infection, such as bacterial sepsis1<\/a><\/sup>; both can cause high fever and severe leukocytosis, both can cause shock and lactic acidosis, and both can cause large increases in inflammatory markers such as C-reactive protein, procalcitonin, ferritin, lactate dehydrogenase, interleukin 6 (IL-6), and tumor necrosis factor\u2013\u03b1.2<\/a><\/sup> However, the former is triggered by a surgical treatment associated with excellent survival outcomes, and the latter is produced by a severe infection associated with poor survival outcomes.3<\/a><\/sup> Yet, both syndromes are commonly labeled cytokine storms<\/em>. How is that possible?<\/a><\/p>\n\n\n\n Long and colleagues4<\/a><\/sup> address this important question by comparing different cytokine storm syndromes in 3 cohorts of patients with malignant hematologic neoplasms: COVID-19 cytokine storm (COVID-CS), cytokine release syndrome from chimeric antigen receptor T-cell therapy (CAR-T CRS), and malignant neoplasm\u2013associated hemophagocytic lymphohistiocytosis (MN-HLH). Of note, a cytokine storm is not always hyperinflammatory and can even be hypoinflammatory; thus, the most accurate term is cytokine dysregulation<\/em>, but we will use cytokine storm<\/em> herein to match the nomenclature used by Long et al.4<\/a><\/sup> The authors used their institution\u2019s electronic medical records (EMR) platform, which integrated clinical and research data, and collected data retrospectively between March 2020 and November 2022 from 671 consecutive patients who had malignant hematologic neoplasms plus 1 of the 3 syndromes. Hierarchical agglomerative clustering was the method used to group data points into clusters based on their similarities; clustering is a statistical method used to group a set of factors in such a way that factors within the same group (or cluster) are more similar to each other than to those in other groups. Random survival forest was used for the survival analysis with high-dimensional data, in addition to traditional Kaplan-Meier and Cox proportional hazards regression models. The study showed that patients with MN-HLH had the most dysregulated inflammatory and coagulation markers and the worst survival prognosis, while there were no significant differences in markers between COVID-CS and CAR-T CRS (ie, they had overlapping clustering). Importantly, these results remained similar after adjusting for age, gender, and race, which showed significantly higher death hazard ratios (HRs) for both MN-HLH (HR, 8.12) and COVID-CS (HR, 2.93) compared with CAR-T CRS. Also, patients receiving tocilizumab had numerically higher death rates in all 3 cohorts compared with those not receiving tocilizumab; additionally, in the COVID-CS cohort, tocilizumab, compared with no tocilizumab receipt, was associated with significantly worse survival, which could be due to illness severity or due to specific tocilizumab effects in the cohort with COVID-CS.<\/a><\/p>\n\n\n\n The limitations of the study by Long et al4<\/a><\/sup> include the inherent problems with the retrospective design, including selection bias, confounding by indication, competing risks, and survival bias. While the authors made major efforts to adjust for baseline characteristic imbalances, it is still possible that residual confounding and selection bias played a role in the survival outcome results. This study has multiple strengths. The seamless integration of hospital EMR with research projects is a great example of the utility of this interface for current and future research on biomarkers and host response. The use of clustering techniques in conjunction with traditional techniques such as Kaplan-Meier and Cox proportional hazards regression models to adjust and confirm clustering findings is an important asset of this study because clustering alone can miss the influence of known clinical factors that impact outcomes. The study period covered most SARS-CoV-2 strains, from Alpha through Omicron, which suggests that the COVID-CS occurred independent of variant types.<\/a><\/p>\n\n\n\n What do this study\u2019s findings add to our current knowledge and clinical practice? The new findings make a substantial contribution to our knowledge because now we have better means to understand patients\u2019 inflammatory and coagulation profiles related to these 3 specific clinical syndromes. Regarding clinical practice, this study provides a more precise approach for health care practitioners to prognosticate and communicate with patients experiencing these clinical syndromes. In fact, once the syndromic diagnosis is made, clinicians can provide a more informative prognostic course and plan the care in conjunction with patients and families.<\/a><\/p>\n\n\n\n The question now is how we move this host response field forward. Inflammatory and coagulation responses are essential profiles to be evaluated in both infectious and noninfectious etiologies of cytokine storm syndromes. First, the findings of the study by Long et al4<\/a><\/sup> need to be replicated in external hospital cohorts to better define their generalizability, which is critical for clinical applicability. Second, we need fast and large growth of research opportunities to further develop a secure and accurate seamless interface between EMR and research projects. Third, biomarker profile randomized clinical trials (RCTs) will be necessary because the absence of randomization (most biomarker studies to date) precludes the distinction between outcomes related to baseline patient characteristics (prognostic) and outcomes related to biomarkers and host responses (predictive); this distinction is sine qua non if we want to make causality progress.5<\/a><\/sup> Fourth, the more concretely defined the inflammatory and coagulation profiles, the more enriched the future RCT populations will be, which in turn will maximize the discovery of precision treatments for different cytokine storms.<\/a><\/p>\n\n\n\n Notably, not all cytokine storms are hyperinflammatory. This is supported by the minimally elevated IL-6 levels across all cohorts in the study by Long et al4<\/a><\/sup> and worse survival in the patients with COVID-CS receiving tocilizumab. While baseline illness severity may partially explain this finding, another explanation may be that IL-6 is not causally associated with these conditions. Additionally, some markers may suggest other pathophysiologic processes. Though ferritin is an acute-phase reactant often conceptualized as synonymous with inflammation, ferritin is primarily stored intracellularly6<\/a><\/sup> and may be released from damaged or dying cells.7<\/a><\/sup> Thus, substantial cytolysis may in part account for the significantly elevated ferritin in MN-HLH and worse survival compared with CAR-T CRS and COVID-CS in the study by Long et al.4<\/a><\/sup> A better understanding of causality is critical to treatment of these and similar conditions, as infectious processes may require a different therapeutic approach (antimicrobial plus immunomodulator) compared with cytokine storms from noninfectious etiologies (underlying disease therapy plus immunomodulator). Additionally, not all immunosuppressive drugs are equal immunomodulators, and not all immunomodulators are equal immunosuppressive drugs.<\/a><\/p>\n\n\n\n