Multi-dimensional COVID-19 short- and long-term outcome prediction algorithm

COVID-19 is an evolutionarily unprecedented natural experiment testing human immunological fitness and the potential to recover from a virus-activated “cytokine storm syndrome” [1], a manifestation of viral sepsis [2]. Elderly persons with co-morbidities such as diabetes mellitus, renal failure, chronic obstructive pulmonary disease, heart conditions, and obesity are at higher risk of severe COVID-19 disease than the general population [3–14], likely because of weaker immune responses, called immune-senescence. This immunological pattern is characterized by decreased phagocyte and lymphocyte function, antigen presentation, cellular replication response to cytokine stimuli, persistent T cell exhaustion, constant low-level inflammation with elevated baseline levels of cytokines such as interleukin-1 (IL-1), IL-6 and tumor-necrosis factor alpha (TNF-a) and prolonged inflammatory states after infections, allowing viral infections to easily evade eradication by the immune system and develop into serious systemic infections [15]. Cytokine-storm with pro-inflammatory coagulopathies presumably also play a role in younger persons without pre-existing conditions who develop severe COVID-19 phenotypes including strokes [16].

In addition to causing pneumonia [3,17], sepsis and septic shock [3,18], COVID-19 has been reported to cause injury and dysfunction in virtually every organ system, including the cardiovascular [19–22], neurological [7,16,23,24], renal [25,26], hepatic [11,27,28], gastro-intestinal [18,29], clotting [16] and immune [18] systems.

Short- and long-term organ dysfunction and survival outcome prediction is important to stratify COVID-19 positive patients for disease severity, treatment escalation and resource allocation. Clinical tools alone, such as respiratory, cardiovascular and other organ function scores, may not be informative enough to precisely assess the prognosis of this novel disease. As of March 31, 2020, a systematic review identified 10 clinical models of predictive outcomes and concluded that these models were “poorly reported and at high risk of bias, raising concerns that their prediction could be unreliable when applied in daily practice” [30]. There is a need for a comprehensive precision risk stratification and prediction that is independently validated, guided by complete short- and long-term outcome event observation including organ dysfunction and death [31,32]. Several biomarkers have been suggested, including higher leukocyte counts, levels of C-reactive protein, procalcitonin, creatinine kinase, myohemoglobin, high-sensitivity troponin I, N-terminal pro-B-type natriuretic peptide, aspartate aminotransferase, creatinine [20], higher white blood cell and neutrophil counts, lymphopenia, higher neutrophil-to-lymphocyte ratio as well as lower percentages of monocytes, eosinophils, basophils [11], and viral load of SARS-CoV-2. In 76 patients studied Jan 21 to Feb 4, 2020 in China, mean viral load of severe cases was 60 times higher than that of mild cases. While 90% of mild cases tested negative by day 10, all severe cases tested positive beyond day 10 [33].

We hypothesize that short- and long-term outcomes such as organ dysfunction and death in COVID-19 patients are related to the potential to cope with immunological stressors such as novel infections and that this coping potential is a resultant of the combined impact of the primary culprit (i.e. SARS-CoV-2 infection), secondary organ dysfunction, co-morbidities, frailty, disabilities as well as chronological age, jointly termed functional recovery potential [34], governed by the balance between innate and adaptive immunity networks [35]. and reflected in peripheral blood mononuclear cell (PBMC) biology.

We have previously developed peripheral blood mononuclear cell (PBMC) -based molecular outcome prediction algorithms for other forms of organ failure [34,36–41]. During the development of the test for advanced heart failure survival prediction, age, respiratory rate, white blood cell count and diastolic blood pressure were identified as clinical predictor candidates for algorithm development. In the transcriptome discovery phase of the test development, 12 genes that represented the shared set between 28 genes predicting early functional recovery and 105 genes correlating with 1-year survival (AGRN, SAP25, ANKRD22, BATF2, FRMD6, HEXA-AS1, DNM1P46, NAPSA, KIR2DL4, RHBDD3, TIMP3, and BCORP1) were selected [36]. Based on these 4 clinical parameters and 12 differentially expressed genes, a predictive algorithm was constructed and independently validated in a larger, independent mixed advanced heart failure cohort. In this independent validation study, prediction of 1-year survival using 4 clinical parameters alone achieved an AUC=0.69. Adding the 12 genes to the clinical model improved the AUC to 0.90. For bedside test development, we calculated and calibrated a prototype clinical score for each of the 77 patients so that higher scores correlated with higher 1-year survival. Since our central postulate is that new onset or worsening frailty, organ dysfunction and death in advanced heart failure and symptomatic COVID-19 may be mediated by overlapping biological mechanisms resulting from innate and adaptive immune cell dysfunction, we hypothesize that a similar strategy may be useful to develop a COVID-19 outcome prediction algorithm.

The clinical variables selected during development of our advanced heart failure survival prediction algorithm, including age, respiratory rate, white blood cell count and diastolic blood pressure, are putatively involved in COVID-19 outcome prediction [3,4,6–8,10–12,14,42,43]. On the molecular immunology level, early comprehensive blood transcriptome profiling in COVID-19 patients suggests decreased levels of T cells and B cells, increased monocyte counts, high levels of inflammatory genes expressed in T cells and monocytes [13,22,44]. This pattern resembles the pattern we described in advanced heart failure patients undergoing mechanical circulatory support surgery [40]. Single-cell RNA sequencing of PBMC suggested a heterogeneous interferon-stimulated gene signature, HLA class II downregulation, a novel B cell-derived granulocyte population and low expression level of pro-inflammatory cytokines, suggesting that circulating leukocytes do not significantly contribute to the potential COVID-19 cytokine storm [45], contrary to other authors reporting inflammatory CD14 CD16 monocytes with high expression of pro-inflammatory cytokines such as IL-6 [10,43,46]. Cell-free DNA and neutrophil extracellular traps correlated with acute phase reactants including C-reactive protein, Ddimer, and lactate dehydrogenase, as well as absolute neutrophil count [47]. The role of mitochondrial cell-free DNA, originating from organ injury and initiating damage-associated molecular patterns of pro-inflammatory cytokine production and hyper-inflammation, has not yet been elucidated in COVID-19 [48]. We propose to build on these important early datasets and our previous algorithm development work in the urgent task of systematically developing and independently validating algorithms for COVID-19-related adverse short- and long-term outcomes such as new onset or worsening frailty, organ dysfunction and death.

In addition to facilitating the rapid development of precise outcome prediction algorithms for COVID-19, this strategy may help to better interpret the clinical role and relevance of inflammatory pathways that have been described in relation to COVID-19. For example, the IL-6 pathway, a primary correlate of unfavorable outcomes in COVID-19 patients [4,46] and the primary cytokine driving hepatic C-reactive protein (CRP) production, controls a crucial juncture in a proven pathway linking inflammation to clinical events and may be an important all-cause mortality prediction biomarker and therapeutic target in COVID-19 patients [9]. It appears that IL-6 is a key pathway associated with various disease conditions, biologically associating COVID-19 with chronic cardiovascular conditions [49]. While the correlation of the IL-6 pathway with outcomes in severe COVID-19 cases emerges as an important pattern, the answer to the question whether IL-6 activation is protective or detrimental in severe COVID-19 disease is open [50]. The biology of IL-6 is complex [51,52]. On one hand, the interaction of the soluble circulating cytokine IL-6 with the membrane anchored receptor system, the IL-6 receptor (IL-6R) and the glycoprotein 130 (gp130) co-receptor, leads to adaptive and protective tissue repair responses [53]. On the other hand, the anchoring of soluble IL-6/IL-6-receptor molecules to the membrane-bound gp130 co-receptor alone, termed “trans-signalling”, potently activates inflammatory responses. The lack of soluble IL-6R in type 2 pneumocytes may explain the hyper- inflammatory and maladaptive reactions in pulmonary manifestations of severe COVID-19 disease, i.e. the COVID-19 related acute respiratory distress syndrome (ARDS) associated with “second wave” mal-adaptive and detrimental macrophage-activation syndrome [50] and cytokine-storm [1]. Elucidation of this dual nature of the role of the IL-6 pathway in COVID-19 benefits from the proposed whole-transcriptome biomarker development strategy in order to understand its adaptive versus maladaptive nature to enhance prediction, mechanistic understanding of outcomes and better define mechanistic therapeutic targets such as tocilicumab in symptomatic COVID-19 patients [46,54,55].

In summary, we propose to develop a multi-dimensional blood-based short- and long-term organ dysfunction and mortality event prediction algorithm for COVID-19 patients that integrates clinical parameters, transcriptomic data, and possibly genome and immunophenotyping/cytokine markers (Figure 1). This test is anticipated to facilitate personalization of treatment, prediction of long-term outcomes and optimization of resource allocation.

Figure 1:

Study Design for Multi-Dimensional COVID-19 Short- and Long-Term Outcome Prediction Algorithm Development.