Identifying Clinically Based Patient Clusters for People with Systemic Inflammatory Disease Using Unsupervised Machine Learning

Objectives Systemic inflammatory diseases (SIDs) are characterized by non-infectious multisystem inflammation. Genetic panels diagnose only 25% of suspected SID patients, due to SIDs’ clinical and genetic heterogeneity. We hypothesize that unsupervised machine learning and multifactorial data integration of clinical and laboratory data will improve the diagnosis and management of SIDs by identifying clinically based patient phenotypes.

Methods We included pediatric genetically undiagnosed SID patients (symptom onset <18 years old) with clinical, laboratory, and whole-exome sequencing data. Genetic ancestry was inferred using principal components and PEDDY with 1K Genomes as a referent. Disease manifestations were divided into clinical and laboratory data sets, weighted by missingness and the data set size, then input into similarity network fusion (SNF) to identify patient clusters. Fisher’s exact test (Bonferroni corrected P<0.0004) identified variables with different distributions between clusters. Clusters were validated with 1000 independent simulated SNFs (simSNF). In each simSNF, we randomly used 70% of the cohort, thereafter the results across all iterations were aggregated. We tested autosomes (n=17443) and chromosome X (n=737) gene-level associations between SNF-clusters, adjusted for sex and ancestry, using SKAT-O (SAIGEv1.4.5). We completed gene set enrichment pathway analysis aggregated with SKAT-O genetic associations across all Biological Process Gene Ontology pathways. We estimated the effective number of independent pathways (Galwey method, P<4.9x10-5; 0.05/3571 independent pathways).

Results We included 104 patients with undifferentiated SID (Figure 1A). SNF revealed 2 clusters. The median age of symptom onset in cluster 1 (n=72) was 5.7 years (Q1-Q3: 3.3-11.2) and cluster 2 (n=32), 6 years (Q1-Q3: 1-13.6). Cluster 2 included patients with consistently higher prevalence (>946/1000 simSNFs) of elevated levels of IgG antibodies (according to each laboratory’s reference standard), transaminitis, anti-nuclear antibodies, anemia, and macrophage activation syndrome compared to patients in cluster 1 (P<4.9×10^-5; Figure 1B). Laboratory manifestations predominantly characterized the clusters. Sensitivity analysis demonstrated that 89% of patients consistently clustered together over 1000 simSNFs. Individually, none of the tested genes were associated with cluster membership (P>2.5×10^-6). Pathway analysis demonstrated a significant association between ”flavonoid metabolic process” GO:0009812 and cluster membership (NES=1.94; P=3.7×10^-6). This metabolic process is a known anti-inflammatory pathway responsible for arachidonic acid metabolism, inhibiting NF-κB and MAPK signaling, and suppressing dendritic and mast cell activation.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1:

(A) Demographic and clinical characteristics of the SID cohort (n=104). (B) Clinical and laboratory manifestations with different prevalences between SNF clusters. Clinical manifestations are labelled in blue and laboratory manifestations are labeled in purple. “MAS” stands for macrophage activation syndrome. Difference in manifestation prevalence between clusters was determined with a fisher’s exact test. A Bonferroni corrected threshold for significance was set at P<0.0004 (0.05/125 independent tests).

Conclusion We identified 2 robust patient clusters from clinical and laboratory manifestations in a heterogeneous SID cohort with SNF. Cluster differences were primarily influenced by laboratory manifestations. Future work will use imputed gene expression to determine the role of flavonoid metabolism on SID pathogenesis.