A risk score based on real-world data to predict early death in acute promyelocytic leukemia

Abstract

With increasingly effective treatments, early death (ED) has become the predominant reason for therapeutic failure in patients with acute promyelocytic leukemia (APL). To better prevent ED, patients with high-risk of ED must be identified. Our aim was to develop a score that predicts the risk of ED in a real-life setting. We used APL patients in the populationbased Swedish AML Registry (n=301) and a Portuguese hospital-based registry (n=129) as training and validation cohorts, respectively. The cohorts were comparable with respect to age (median, 54 and 53 years) and ED rate (19.6% and 18.6%). The score was developed by logistic regression analyses, risk-per-quantile assessment and scoring based on ridge regression coefficients from multivariable penalized logistic regression analysis. White blood cell count, platelet count and age were selected by this approach as the most significant variables for predicting ED. The score identified low-, high- and very high-risk patients with ED risks of 4.8%, 20.2% and 50.9% respectively in the training cohort and with 6.7%, 25.0% and 36.0% as corresponding values for the validation cohort. The score identified an increased risk of ED already at sub-normal and normal white blood cell counts and, consequently, it was better at predicting ED risk than the Sanz score (AUROC 0.77 vs. 0.64). In summary, we here present an externally validated and population-based risk score to predict ED risk in a real-world setting, identifying patients with the most urgent need of aggressive ED prevention. The results also suggest that increased vigilance for ED is already necessary at sub-normal/normal white blood cell counts.

Figure 1.

Distribution of early deaths. (A, B) The number and deaths per day from time of diagnosis in the training cohort (A) and the validation cohort (B).

Figure 2.

Risk of early death per variable and quantile. The training cohort (n=301) was divided into ten quantiles for age, platelets and white blood cell counts, with approximately 30 patients per quantile. Locally estimated scatterplot smoothing curves are shown for each variable with regard to rate of early death on the y-axis. ED: early death; WBC: white blood cell count.

Figure 3.

Risk score algorithm. The total sum of assigned points per variable level can rapidly be found in the tabulated chart above. Higher points indicate a higher risk for early death (ED). Colors indicate the predicted risk of ED as shown in the boxes below.

Figure 4.

Risk score distribution and calibration plot. (A, B) Distribution of risk score groups in the training (A) and validation (B) cohorts, with green indicating low-risk, yellow indicating high-risk and red indicating very-high risk groups. (C) Calibration plot with the observed proportion of early deaths (ED) per risk group in the training cohort (blue) and validation cohort (green).

Figure 5.

Receiver operating characteristic curves with areas under the curve (AUC) for the risk scores assessed: the score suggested here, the Sanz score, and the scores of. Lou et al. and Cai et al.