A Personalized Prediction Model to Risk Stratify Patients with Myelodysplastic Syndromes

Background

Patients (pts) with myelodysplastic syndromes (MDS) have heterogeneous outcomes that can range from months for some pts to decades for others. Although several prognostic scoring systems have been developed to risk stratify MDS pts, survival varies even within discrete categories, which may lead to over- or under-treatment. Deficits in discriminatory power likely derive from analytic approaches or lack of incorporation of molecular data.

Here, we developed a model that uses a machine learning approach to analyze genomic and clinical data to provide a personalized overall outcome that is patient-specific.

Method

Clinical and mutational data from MDS pts diagnosed according to 2008 WHO criteria were analyzed. The model was developed in a combined cohort from the Cleveland Clinic and Munich Leukemia Laboratory and validated in a separate cohort from the Moffitt Cancer Center. Next generation targeted deep sequencing of 40 gene mutations commonly found in myeloid malignancies was performed. Pts who underwent hematopoietic cell transplant (HCT) were censored at the time of transplant. A random survival forest (RSF) algorithm was used to build the model, in which clinical and molecular variables are randomly selected for inclusion in determining survival, thereby avoiding the shortcomings of traditional Cox step-wise regression in accounting for variable interactions. Survival prediction is thus specific to each pt's particular clinical and molecular characteristics. The accuracy of the proposed model, compared to other models, was assessed by concordance (c-) index.

Results

Of 2302 pts, 1471 were included in the training cohort and 831 in the validation cohort. In the training cohort, the median age was 71 years (range, 19-99), 230 pts (16%) progressed to AML, 156 (11%) had secondary/therapy-related MDS, and 130(9%) underwent HCT. Risk stratification by IPSS: 391 (27%) low, 626 (43%) intermediate-1, 280 (19%) intermediate-2, 104 (7%) high, 104 (7%) missing, and by IPSS-R: 749 (51%) very low/ low, 336 (23%) intermediate, 190 (13%) high, 92 (6%) very high, and 104 (7%) missing. Cytogenetic analysis by IPSS-R criteria: 65 (4%) very good, 1060 (72%) good, 193 (13%) intermediate, 60 (4%) poor, and 93 (6%) very poor. The most commonly mutated genes were: SF3B1 (26%), TET2 (25%), ASXL1 (20%), SRSF2 (15%), DNMT3A (12%), STAG2 (8%), RUNX1 (8%), and TP53 (8%). All clinical variables and mutations were included in the RSF algorithm. To identify the most important variables that impacted the outcome and the least number of variables that produced the best prediction, we conducted several feature extraction analyses which identified the following variables that impacted OS (ranked from the most important to the least): cytogenetic risk categories by IPSS-R, platelets, mutation number, hemoglobin, bone marrow blasts %, 2008 WHO diagnosis, WBC, age, ANC, absolute lymphocyte count (ALC), TP53, RUNX1, STAG2, ASXL1, absolute monocyte counts (AMC), SF3B1, SRSF2, RAD21, secondary vs. de novo MDS, NRAS, NPM1, TET2, and EZH2.

The clinical and mutational variables can be entered into a web application that can run the trained model and provide OS and AML transformation probabilities at different time points that are specific for a pt, Figure 1.

The C-index for the new model was .74 for OS and .81 for AML transformation. The new model outperformed IPSS (c-index .66, .73) and IPSS-R (.67, .73) for OS and AML transformation, respectively. The geno-clinical model outperformed mutations only (c-index .64, .72), mutations + cytogenetics (c-index .68, .74), and mutations + cytogenetics +age (c-index .69, .75) for OS and AML transformation, respectively. Addition of mutational variant allelic frequency did not significantly improve prediction accuracy.

When applying the new model to the validation cohort, the c-index for OS and AML transformation were .80, and .78, respectively.

Conclusion

We built a personalized prediction model based on clinical and genomic data that outperformed IPSS and IPSS-R in predicting OS and AML transformation. The new model gives survival probabilities at different time points that are unique for a given pt. Incorporating clinical and mutational data outperformed a mutations only model even when cytogenetics and age were added.

View largeDownload slide

View largeDownload slide

Close modal