Clinically Driven Algorithms Identify Patients with Large Granular Lymphocytic Leukemia at Risk of Significant Cytopenia Requiring Treatment

Background Although the majority of patients (pts) with Large granular lymphocytic leukemia (LGLL) present with cytopenias at diagnosis, the necessity for targeted treatment is not universally indicated. Diagnostic landscape is further complicated by the coexistence of several potential factors including autoimmune conditions, B-cell dyscarasia, clonal hematopoiesis (CH) and myelodysplastic syndromes (MDS). This intricate interplay of factors poses significant challenges in accurately delineating the primary etiological factors of cytopenias, thereby complicating the clinical-decision making process for appropriate therapeutic interventions. Understanding these factors and the need for treatment in LGLL is crucial for improving outcomes and determining appropriate treatment and follow up. Methods We performed a retrospective review of 558 LGLL pts diagnosed from 2 participating institutions between 2010 to 2024. LGLL clonality was assessed by Vβ flow and T-cell rearrangement (TCR) by PCR. Pts were evaluated for monoclonal gammopathy (MG), history of rheumatoid arthritis (RA), and presence of splenomegaly. Univariate logistic regression was performed to identify factors associated with anemia, neutropenia, and treatment requirement. A random forest model (machine learning algorithm) was developed to identify predictors for treatment requirements based on features importance with area under the curve (AUC) used to compare the random forest vs multivariate logistic regression model including the same variables. Results Of 558 pts, the median age was 64 years (IQR 55-73) with 58% females and 42% males. Median LGL cell count was 1050 (IQR 480-2208). Overall, 123 (32%) pts had concurrent MG, 99 (38%) had STAT3 mutations (STAT3MT), and 4 (2%) had STAT5B mutations (STAT5bMT). Splenomegaly was present in 101 (19%) pts at diagnosis. TCR Vβ clonality was detected in 158 (51%) pts. Importantly, 202 (38%) pts had clinical manifestations requiring LGLL directed therapy. Anemia was present in 357 pts (64%), 63% of whom were female. Pts presenting with anemia had a higher median age vs those without anemia (67 vs. 62 years, p<0.001). Clonal Vβ/TCR was documented in 45% of pts with anemia. Splenomegaly and STAT3MT were more frequent in pts with anemia (24% vs. 9%, p <0.001, 46% vs. 27%, p = 0.003, respectively). In univariate analysis, higher age (odds ratio [OR]=1.02, p<0.001), splenomegaly (OR=3.0, p<0.001), non-clonality by Vβ TCR (OR=1.9, p=0.008) and STAT3MT (OR=2.4, p=0.001) were associated with anemia. Pts with neutropenia (322, 58%) had lower median age vs. those without neutropenia (63 years vs. 65 years, p < 0.001). STAT3MT (49% vs. 26%, p < 0.001) was more prevalent among pts with neutropenia. In univariate analysis, STAT3MT (OR=2.7, p<0.001) was associated with neutropenia. Only 58 (10%) of pts had thrombocytopenia (platelet <100 x 109/L) at time of diagnosis. Indications for treatment were anemia (84%) or neutropenia (72%). Clonal Vβ/TCR (64% vs. 44%, p < 0.001), splenomegaly (30% vs. 12%, p < 0.001), RA (24% vs. 11%, p < 0.001) and STAT3MT (52% vs. 25%, p < 0.001) were more prevalent in treated pts. In univariate analysis, splenomegaly (OR=3, p<0.001), RA (OR=2.6, p=0.03), STAT3MT (OR=3.2, p<0.001) and clonal Vβ/TCR (OR=2.3, p<0.001) were associated with treatment requirement. Expression of CD5 (OR=1.5, p=0.03)/ CD7 (OR=2.5, p<0.001)/ CD16 (OR=2.2, p<0.001) and absence of CD4 (OR=0.4, p<0.001) were also associated with treatment need. In multivariate regression, splenomegaly (OR: 2.7, p=0.002), RA (OR: 4.8, p<0.001), hemoglobin (OR: 0.7, p<0.001), CD7 (OR: 3, p=0.001), CD16 (OR: 2.5, 0.04), and CD4 (OR: 0.4, p=0.03) were associated with need for treatment. The most important features (>15%) identified by a random forest model to predict treatment need included hemoglobin, ANC, platelet, white blood cells, LGL cells count, age, CD7, splenomegaly, and RA. Compared to multivariate regression, random forest model had AUC of 0.99 vs 0.85. Conclusions Applying machine learning algorithms, the study revealed that splenomegaly, RA, clonal Vβ/TCR, MG, specific CD markers, and STAT3MT are strongly associated with cytopenias and treatment requirement in LGLL pts. These findings emphasize the importance of comprehensive diagnostic evaluations to identify exact cytopenia mechanisms among LGLL pts, which can lead to more personalized treatment strategies.