Etiological stratification and prognostic assessment of haemophagocytic lymphohistiocytosis by machine learning on onco-mNGS data and clinical data

Abstract

Introduction: Hemophagocytic lymphohistiocytosis (HLH) is a rare, complicated and life threatening hyperinflammatory syndrome that maybe triggered by various infectious agents, malignancies and rheumatologic disorders. Early diagnosis and identification of the cause is essential to initiate appropriate treatment and improve the quality of life and survival of patients. The recently developed Onco-mNGS technology can be successfully used for simultaneous detection of infections and tumors.

Methods: In the present study, 92 patients with clinically confirmed HLH were etiologically subtyped for infection, tumor and autoimmunity based on CNV and microbial data generated by Onco-mNGS technology, and a predictive model was developed and validated for the differential diagnosis of the underlying disease leading to secondary HLH. Furthermore, the treatment outcomes of patients with HLH triggered by EBV infection and non-EBV infection were evaluated, respectively.

Results: The current study demonstrated that the novel Onco-mNGS can identify the infection and malignancy- related triggers among patients with secondary HLH. A random forest classification model based on CNV profile, infectious pathogen spectrum and blood microbial community was developed to better identify the different HLH subtypes and determine the underlying triggers. The prognosis for treatment of HLH patients is not only associated with CNV, but also with the presence of pathogens and non- pathogens in peripheral blood. Higher CNV burden along with frequent deletions on chromosome 19, higher pathogen burden and lower non-pathogenic microbes were prognosis factors that significantly related with unfavorable treatment outcomes.

Discussion: Our study provided comprehensive knowledge in the triggers and prognostic predictors of patients with secondary HLH, which may help early diagnosis and appropriate targeted therapy, thus improving the survival and prognosis of the patients.

Keywords: etiological stratification; hemophagocytic lymphohistiocytosis; machine learning; onco-mNGS; prognostic assessment.

Figure 1

Overall design and flowchart of this study. A total of 118 samples were finally enrolled in this study after excluding samples that failed pass the quality control and without complete clinical information. Blood samples were selected for further microbes identification and chromosomal copy number variations analysis in this study. Abbreviations: HLH, hemophagocytic lymphohistiocytosis; mNGS, metagenomic next-generation sequencing.

Figure 2

The distribution of HLH subtypes and clinical specimens in the present study. (A) Retrospective diagnosis of all patients. (B) Distribution of sample types. Abbreviations: BALF, bronchoalveolar lavage fluid; CSF, cerebrospinal fluid.

Figure 3

First-level etiological stratification of secondary HLH diagnosis based on analysis of CNV data derived from Onco-mNGS. (A) Copy number variation (CNV) positive rate detected by mNGS. (B) Distribution and frequncy of gain (AMP, amplification) or loss (DEL, deletion) of DNA segments in subtype EBV-HLH and subtype M-HLH. (C) The efficacy assessment of a diagnostic model for the etiological stratification of secondary HLH at the first level based on CNV data (ROC curve). (D) Diagram of the first level of etiological stratification for the diagnosis of secondary HLH. The Random Forest binary classifier constructed on the basis of CNV data could better distinguish Group 1 (CNV-positive EBV-HLH and M-HLH) from Group 2 (CNV-negative EBV-HLH, nEBV-HLH, and Rh-HLH), with an AUC value of 0.786.

Figure 4

Secondary-level etiological stratification of secondary HLH diagnosis based on analysis of microbiome data derived from Onco-mNGS and clinical examination data. (A) Upset plots of blood microorganisms with frequencies above 3% in each HLH subtype identified by mNGS. (B) Box plots of blood microbial burdens in different HLH subtypes, differences between groups were assessed by T-test. (C) Specific blood microbial biomarkers in different subtypes were determined through LEfSe analysis. The microbial species enriched in the EBV-HLH, M-HLH and Rh-HLH subtypes were presented in the plot with respective average RPM value and LDA scores. Alpha value for the factorial Kruskal-Wallis test among classes was 0.01, and for the pairwise Wilcoxon test between subclasses was 0.05. A threshold value of 2.0 was applied to the log LDA score for discriminatory features. Significant differences between groups are indicated by asterisks, with * represents P<0.05, ** represents P<0.01, *** represents P<0.001. Abbreviations: PCoA, principal coordinate analysis; PERMANOVA, permutational multivariate analysis of variation; LEfSe, linear discriminant analysis effect size; LDA, linear discriminant analysis; NS., no significant difference. (D) The efficacy assessment of a diagnostic model for the etiological stratification of Group 1 (CNV-positive EBV-HLH and M-HLH) at the secondary level based on microbiome and clinical examination data (ROC curve). (E) The efficacy assessment of a diagnostic model for the etiological stratification of Group 2 (CNV-negative EBV-HLH, nEBV-HLH, and Rh-HLH) at the secondary level based on microbiome and clinical examination data (ROC curve). (F) Diagram of both the first level and the secondary level of etiological stratification for the diagnosis of secondary HLH.Two random forest binary classifiers and one ternary classifier constructed based on different types of data can better distinguish CNV-positive EBV-HLH, M-HLH, CNV-negative EBV-HLH, nEBV-HLH, and Rh-HLH in different hierarchical order effectively.

Figure 5

The prognosis prediction in EBV-HLH patients based on both CNVs and blood microbiome characteristics. (A) Occurrence of CNVs in patients with different treatment outcomes. (B) Non-remission rates in CNV positive or negative patients with different HLH subtypes. (C) Comparison of the counts and burdens of pathogens in HLH patients with different treatment outcomes. Differences between groups were assessed using T-test. (D) Heatmap of selected microbes with frequencies above 10% in each group (R and NR) identified by mNGS in patients with nEBV-HLH subtype. Microbes with frequency more than 10% within each group (R and NR) were selected, log10-transformed RPM of selected microbes were applied. Samples were hierarchically clustered within each group using Pearson correlation as a distance measure with average-linkage. Abbreviations: R, remission; NR, non-remission. (E) The determination of the number of significant variables for classification. The optimal point of cross-validation error determined by the number of biomarkers was 13, which implies that based on the mean decreasing accuracy, the top 13 variables which are all CNV-related parameters could be selected as potential markers used to differentiate between secondary HLH treatment effects (R vs. NR). (F) The efficacy assessment of a diagnostic model for the early assessment of prognosis for secondary HLH treatment based on the above mentioned 13 CNV-related variables (ROC curve).