IPSS-M(r): A Reduced Feature Representation of IPSS-M for Clinical Application

Background: The Molecular International Prognostic Scoring System (IPSS-M) is the most recent and comprehensive risk stratification scheme for myelodysplastic neoplasms. In total, IPSS-M requires 41 features (mutational data, cytogenetics, and clinical parameters) that are divided into main effect variables and residual genes and yield six risk categories. While it is optimal to include all input variables, the presence of inter-institutional heterogeneity in molecular profiling panels led us to investigate whether utilizing a reduced number of features can still accurately reproduce the prognostic classification demonstrated in the original model and to what degree.

Methods: Data from MDS IWG samples deposited on cBioPortal (Cerami et al., 2012) along with the prognostic class labels was used as the training cohort. A discovery cohort created using a published metanalytic cohort (Kewan et al., Nature Communications, 2023) and a separate validation cohort was prepared from the clinocogenomic profiles of patients treated at the Karmanos Cancer Institute. Baseline clinical and molecular characteristics were noted. We analyzed the molecular data in the context of cytogenetic risk profile, clinical parameters, coexisting mutations, and overall survival (OS) to create a reduced statistical model called IPSS-M(r).

Results: 7,930 adult MDS patients were screened, and 3,519 eligible patients were separated into a training (n=3323) and a validation cohort (n=296). The training cohort was further subdivided into a train-test split (80% training and 20% testing), and a Support Vector Machine (SVM) classifier was trained using the IPSS-M labels along all 41 features. Missing values were replaced by using a KNN-imputation method. The labels were transformed into classes as follows: 0 (Very-Low risk), 1 (Low Risk), 2 (Moderate-Low Risk), 3 (Moderate-High Risk), 4 (High Risk), 5 (Very-High Risk). Model training was evaluated on three metrics: Precision, Recall, and F1-score. The results are summarized below in Fig. 1A. Performance across all classes on the 20% holdout data was 88.6% on average. The best classification was achieved by Very-Low, High, and Very-high-risk classes. Then, we assessed the trained model for feature importance and Principal Component Analysis (PCA) to uncouple the strong feature associations (Fig 1B) and remove 50% of the lowest-performing features (Fig. 1A). After this dimensionality reduction, we retrained the SVM model on all the labels and discovered that although the model performance across all classes decreased by 10.2% on average, not all classes were equally affected. The feature reduction most dramatically influenced the middle classes (Moderate-Low and Moderate-High), whereas the extremes (Very-Low, High, and Very-High) were resistant to feature reduction.

Conclusion: Although all 41 features were necessary for the MDS IWG to establish IPSS-M and denote prognostic significance to the risk categories, a reduced feature machine learning model can still accurately classify >85% of patients. Our model is not meant to replace IPSS-M but instead supplement risk stratification in MDS for low-resource settings, as well as account for heterogeneity in molecular profiling. Ongoing work is focused on improving the model performance by implementing deep-learning base models and retaining greater variance with the fewest number of features possible.

Balasubramanian:Alexion AstraZeneca: Speakers Bureau; Kura Oncology: Research Funding.