The symptoms of PASC have been widely reported and significantly overlap with ME-CFS and with PTLD1-3. In addition, we included patients, of unknown prevalence, who are post-vaccine individuals with PASC-like symptoms (POVIP) in the present study (Figure 1). These patients experience PASC-like symptoms 3 months minimum post-vaccination in the absence of COVID 19 infection.
By using a decision tree classifier, Classification and Regression Trees (CART), we developed an algorithm to propose a simple but powerful predictive model with high interpretability, a characteristic of great importance when attempting to understand differences between disease states. Our model had an average weighted F1 score of 0.61, which was variable due to the stochastic nature of both the model and the dataset. As shown in Table 1 and in the confusion matrix (Supplementary Figure 1), the model was robust in its ability to identify four of the six classes of disease states (e.g. MM, Severe, PASC, and ME-CSF). Some misclassification was demonstrated in the remaining two classes (POVIP and Lyme) that was likely due to overlapping cytokine hubs. The confusion matrix may suggest that the immune contribution to these two states were similar. Clinical data as well as other immunological parameters could potentially separate these two conditions and further increase the model’s predictive power.
The highest performance metrics after fine-tuning was provided by the CART decision tree (Fig. 2) when data were plotted using internal tree plotting functions and python’s matplotlib (Supplementary Fig. 2). As shown in the plotted tree (Fig 2, Supplementary Fig. 2), we demonstrate that the CART algorithm was capable of constructing splitting and terminal hubs with low Gini impurity values and high F1 scores for those classes shown previously in the confusion matrix (Supplementary Fig. 1). In the POVID and PTLD classes, splitting hubs with higher Gini impurity values were observed. The identification of these hubs supports the possibility that the immune profiles of both POVIP and PTLD individuals have similar cytokine patterns.
The disease heterogeneity suggests that the classification need not perfectly match the labels and that individuals within each of these conditions might present different immunological entities potentially requiring differential therapeutics. We demonstrated that for POVIP and MM, three or more distinct cytokine profiles might be important for their classification supporting the presence of different immunological entities within these groups. On the other hand, for ME-CSF, Lyme, and PASC only one or two cytokine classification profiles were found.
The PASC classification highlights the importance of the proinflammatory cytokines IL-2 and IFN-γ as we have previously reported2, while in PTLD disease, two classification profiles were identified. Interestingly, both profiles follow a common path including the proinflammatory cytokines GM-CSF and IL-2 in concordance with IL-2 mediated GM-CSF production previously reported9. One PTLD profile appears to be driven by IL-8 induced responses while the other mediated by IL-13. PTLD includes a variety of clinical features and pathogenic mechanisms with two different immune clusters10. Both share features that include T cell receptor signaling and involvement of monocytes/CD4+ T cells. The first cluster characterized by a type 1 inflammatory response associated with post-infectious Lyme arthritis and autoimmune joint disease that are associated with IL-810,11. The second cluster includes activation of neutrophils and IL-4/IL-13 signaling11 which aligns more with post-treatment neurological disease10. These data could explain the two different classification profiles, reported in this study, associated with different clinical manifestations of PTLD.