Identifizierung prognostischer Zellpopulationen bei Patienten mit Chronischer lymphatischer Leukämie durch Anwendung von erklärbarer künstlicher Intelligenz auf durchflusszytometrische Daten

Introduction: Chronic lymphocytic leukemia (CLL) is the most common leukemic disease in Western countries. An established prognostic model for CLL patients is the International Prognostic Index (CLL-IPI), which is based on five prognostic factors (TP53 status, IGHV mutation status, serum beta2-microglobulin, clinical stage, and age). In this work, we investigate whether CLL prognosis is possible by exclusively analyzing multiparameter flow cytometry (MPFC) data from CLL patients using an explainable artificial intelligence (XAI). Here, the XAI should be able to explain the results based on the representation of different cell populations relevant for prognosis in MPFC dot plots. Material and Methods: Analysis of MPFC data from peripheral blood of 157 patients with CLL was performed in alignment with matching clinical data (parameters of CLL IPI, sex, ECOG stage, Richter transformation yes/no, treatment yes/no, date of death, treatment failure, and date of last follow-up). MPFC data were collected for routine diagnostic analysis at Marburg University Hospital from 2014 to 2020, and patients were divided into a worse outcome group and a better outcome group. Patients who died during follow-up and those who failed first-line systemic therapy were categorized as TTF 1 (time-to-first-line treatment failure), i.e., assigned to the worse outcome group. All other patients were assigned to a better outcome with TTF 0. Subsequently, the XAI algorithm ALPODS was used to identify cell populations that were over- or underrepresented in patients with worse outcome (TTF 1) or better outcome (TTF 0) and thus predictive of prognosis. Here, the predictive ability of each XAI population was evaluated using receiver operating characteristic (ROC) curves. For verification, this predictive value of the XAI populations was compared with the frequency of CD38-positive CLL cells and CLL-IPI on the ROC curves. Results: ALPODS defined 17 XAI populations that collectively can predict more accurately than the CLL-IPI score (AUC of XAI populations: 0.95 [95%CI 0.91-0.98; p<0.0001] vs. AUC of CLL-IPI: 0.78 [95%CI 0.70-0.86; p<0.0001]). The best single classifier was an XAI population consisting of CD4+ T cells (AUC 0.78; 95% CI 0.70-0.86; p<0.0001). Patients with fewer CD4+ T cells showed a worse clinical outcome. When the CD4+ T cell population was added, the predictive ability of the CLL-IPI score increased (AUC 0.83; [95%CI 0.77-0.90; p<0.0001]). Combining the CD4+ with the CD8+ T cell population alone correctly distinguished TTF 0 from TTF 1 in 79% of the cases in the patient cohort considered (AUC 0.79; [95%CI 0.71-0.87; p<0.0001]). This is particularly important in light of the fact that these two cell populations could be easily gated manually in flow cytometric dot plots, especially compared to other cell populations defined by ALPODS, some of which are very challenging to correctly gate manually (especially that of CLL subgroups). Conclusion: The ALPODS-XAI algorithm was able to identify predictive immune cell populations that significantly predicted the clinical course of CLL. CD4+ T cells were identified as the best single classifier, further improving predictive ability when combined with the CLL-IPI score. In addition, this work demonstrated that gating CD4+ T cells in combination with the CD8+ T cells, which is relatively easy to perform in the clinic, can also be used to make a clinical prediction. The results need to be validated in a much larger cohort of patients, but are very promising in terms of refining conventional prognostic models such as the CLL-IPI.