Artificial-Intelligence Driven Quantification of Sickle Red Blood Cell Morphology As an Objective Disease Marker

Introduction

The variation of the clinical severity of SCD is remarkable and, although some predictors of SCD clinical severity have been described, no objective reproducible marker predicting disease severity exists. Better predictors of SCD severity are warranted to realize a more precise and personalized management of patients with SCD. We developed an assay that can determine the percentages of morphologically abnormal erythrocytes in an automated, artificial intelligence (AI) driven manner. The aim of this study is to investigate whether quantification of the fractions of morphologically abnormal erythrocytes in SCD patients with this newly developed assay, can be used as an objective disease marker.

Methods

Blood (EDTA) was collected from adult SCD patients with HbSS, HbSβ⁰, HbSC or HbSβ⁺ in steady-state and within 48 hours after hospital admission for a subgroup of patients hospitalized with a vaso-occlusive crisis (VOC). For each patient 50µL of blood was used to determine the fractions of morphologically abnormal erythrocytes under incrementing oxygen conditions (from hypoxic (1% O₂) to atmospheric (21% O₂)). Cells were measured with an imaging flow cytometer that captures a picture of each cell. The captured images were used to categorize each erythrocyte into discocytes, sickle cells and intermediary morphologies (i.e. holly leaf cells and granular cells). After training the AI algorithm by hand tagging over 15,000 cells, it was able to discriminate and quantify each cell type with high accuracy and low inter-sample variability (CoV<0.043). Comparisons between multiple groups were performed using the Kruskal-Wallis test followed by the post-hoc Dunn’s tests. Beta regression was used to examine the relationship between the percentages of morphologically abnormal erythrocytes and clinical data. Fold changes in percentages and 95% confidence intervals (CI) were reported for all factors in the models. P<0.05 was considered statistically significant.

Results

A total of 172 SCD patients were included in the final analysis. 162 patients were included in steady state and 16 during hospital admission for VOC. Of the 162 steady state patients, 97 patients had HbSS and HbSβ⁰ genotypes (median age 29.2 years [IQR 22.3 – 40.5]), while 65 patients had HbSC and HbSβ⁺ genotypes (median age 32.9 years [22.9 – 43.1]). When measured at 21% O₂, HbSS/HbSβ⁰ patients had significantly higher percentages of classic sickle cells, granular cells and holly leaf cells and a significantly lower percentage of (normal) discocytes compared to HbSC/HbSβ⁺ patients. In HbSS/HbSβ⁰ patients during steady state, the percentages of classic sickle cells correlated negatively with hemoglobin levels (R=-0.435, p<0.001) and hemoglobin F percentages (R=-0.293, p=0.030). In these patients, positive correlations were observed between the percentage of classic sickle cells, granular cells and holly leaf cells, and LDH and total bilirubin as markers of hemolysis. Beta regression analysis demonstrated that a current VOC was significantly associated with a mean 1.53 fold increase in percentage of holly leaf cells (95% CI: 1.11 - 2.11, p = 0.009), a mean 1.50 fold increase in the percentage of granular cells (95% CI: 1.01 - 2.25, p = 0.047) and a decrease in the percentage of discocytes by a mean factor of 0.58 (95% CI: 0.37 - 0.89, p = 0.012). The beta regression analysis also demonstrated a trend towards an increase in the percentage of classic sickle cells during VOC (mean 1.49 fold increase, 95% CI: -0.94 - 1.36, p = 0.092). When measured at 1% O₂, the use of hydroxyurea was associated with a mean 0.79 fold decrease (95% CI: 0.64 - 0.97, p = 0.022) in the percentage of classic sickle cells in HbSS and HbSβ⁰ patients.

Conclusion

Using AI-driven analysis of imaging flow cytometry data, we were able to quantify and stratify discocytes, holly leaf cells, granular cells, and classic sickle cells in peripheral blood of SCD patients. The use of automated quantification of the percentages of morphologically abnormal erythrocytes may be a useful tool for predicting SCD severity. Future studies are warranted to determine whether quantification of erythrocyte morphology may be a predictor of disease related complications in SCD patients.