Myelodysplastic syndromes (MDS) comprise a group of heterogeneous clonal hematologic disorders characterized by distorted hematopoietic stem cell function, morphological dysplasia, peripheral blood cytopenias, and an increased likelihood of transformation to acute myeloid leukemia (AML) [1^-3].

Patients with MDS show heterogeneous clinical courses and outcomes; hence, risk-adapted stratification models are crucial for defining appropriate treatment strategies. The International Prognostic Scoring System, revised version (IPSS-R), based on hematological and cytogenetic features, is the most commonly used system to assess disease-related risk and estimate survival. However, the IPSS-R does not provide prognostic information at the individual patient level [4, 5]. Incorporating somatic variants into the analysis can enhance disease prognostication; therefore, several molecular models have been proposed to reflect distinct biological MDS subgroups and refine therapeutic strategies [6^-8].

The recent Molecular IPPS (IPSS-M) model developed by the International Working Group for Prognosis in MDS incorporates clinical parameters, cytogenetic abnormalities, and molecular information for 31 genes to classify patients with MDS into six risk groups [9]. The IPSS-M has shown better prognostic power than previously proposed models, such as the original IPSS, IPSS-R, and the World Health Organization (WHO) classification-based Prognostic Scoring System (WPSS), enabling enhanced prediction of overall survival (OS) and leukemia transformation [10^-12]. The IPSS-M also performed better in providing information regarding the probability of response to specific treatments, such as hypomethylating agents (HMAs) and hematopoietic stem cell transplantation (HSCT) [13, 14].

The implementation of artificial intelligence to develop prognostic models for hematologic malignancies, such as the Artificial Intelligence Prognostic Scoring System for MDS (AIPSS-MDS), has resulted in enhanced performance compared with models developed using traditional statistics, even without the requirement for complex genomic data [7, 15]. The AIPSS-MDS was recently proposed by the Spanish MDS Group (GESMD) as a quantitative score that provides patient-level risk predictions based on the same parameters included in the IPSS-R, along with leucocyte count, age, and sex [15].

While molecular testing is becoming a standard clinical practice worldwide, the adoption of next-generation sequencing (NGS) in South America remains challenging because of various factors, including socioeconomic and geographical constraints. These settings currently limit access to effective risk-stratification tools and constrain the validation of molecular models, such as the IPSS-M, in South American patients. We compared, for the first time, the performance of the IPSS-M prognostic model with that of different prognostic systems, including the AIPSS-MDS, using real-world NGS data from South American patients.

Study population

This retrospective study consisted of 182 adult patients (>18 yrs of age) diagnosed as having myelodysplastic neoplasms between 2009 and 2022 at five centers in Argentina and one in Uruguay. The study was approved by the Institutional Ethics Committee of Academia Nacional de Medicina, Ciudad de Buenos Aires, Argentina (approval No.: 13/23/CEIANM). The data set included patients with MDS and MDS/myeloproliferative neoplasm (MDS/MPN) overlap syndromes, classified according to the 2016 WHO criteria.

Clinical and molecular data

Laboratory and clinical data were collected from medical records and institutional databases. The following clinical characteristics were evaluated at diagnosis: age, sex, 2016 WHO category, percentage of bone marrow (BM) blasts, white blood cell count, absolute neutrophil count (ANC), Hb level, platelet count (PLT), and cytogenetics. Molecular data were generated at each institution using six different commercial myeloid-focused NGS-based panels that comprised variants across 30–63 complete genes and associated hotspot regions (Supplemental Data Table S1). As inclusion criteria, we required pathogenic or probably pathogenic variants with a variant allele fraction (VAF)>5% with at least 20 reads for the variant allele. Variants of uncertain significance (VUS) were excluded. All patients were stratified according to different prognostic scoring systems that used either hematological and cytogenetic data alone (IPSS-R [4] and AIPSS-MDS [15]) or hematological, cytogenetic, and molecular data (IPSS-M [9], European MDS [EuroMDS] [6], and Munich Leukemia Laboratory [MLL] [8]). Average IPSS-M scores were considered to assign risk categories.

Statistical analysis

Categorical variables were evaluated using Fisher’s exact or χ² test, and numerical variables were analyzed using the Mann–Whitney U or Kruskal–Wallis test. OS was defined as the time from diagnosis to death or the last follow-up, censored or not at the time of HSCT. In survival analysis, OS values for the different risk categories were estimated using the Kaplan–Meier method and compared using the log-rank test. Median follow-up was estimated using the reverse Kaplan–Meier method. Univariate and multivariate survival analyses were conducted using Cox proportional hazard regression models. The prognostic accuracy of the different models was assessed based on Harrell’s concordance (C)-index (survival package, R statistical language) and the area under the receiver operating characteristic (ROC) curve (AUC) (pROC package, R statistical language). A two-sided P<0.05 was deemed statistically significant. The data were analyzed and modeled using R statistical language, version 4.3.2 (R Foundation for Statistical Computing, Vienna, Austria).

Clinical characteristics and genomic profile

We retrospectively analyzed clinical and molecular data from 182 patients (145 with MDS and 37 with chronic myelomonocytic leukemia [CMML]). The median age was 65.0 yrs (interquartile range [IQR]: 55.0–73.0), the male:female ratio was 1.3:1 (102/80), the median follow-up time was 23.8 months (95% confidence interval [CI]: 18.5–35.1), and the median OS was 143.0 months (CI: 60.8–not reached), with 51 reported deaths (28%) (Table 1).

Table 1 Descriptive statistics of patient characteristics

Clinical data	Overall (N=182)
Age, yrs, median (IQR)	65.0 (55.0–73.0)
Sex ratio, M: F	1.3
BM blasts %, median (IQR)	2.0 (0.9–4.5)
Hb g/dL, median (IQR)	9.5 (7.4–10.9)
ANC×109/L, median (IQR)	1.9 (0.9–4.5)
PLT×109/L, median (IQR)	105.0 (49.5–218.3)
2016 WHO categories, N (%)
MDS-SLD	14 (7.7)
MDS-MLD	67 (36.8)
MDS-RS-MLD	12 (6.6)
MDS with isolated del(5q)	2 (1.1)
MDS-EB-1	24 (13.2)
MDS-EB-2	16 (8.8)
MDS-U	3 (1.6)
MDS-NOS*	7 (3.8)
CMML	37 (20.2)
Treatment, N (%)
BSC	39 (24.5)
Chemotherapy	3 (1.9)
HMAs	57 (35.8)
HMAs+chemotherapy	7 (4.4)
HSCT	36 (22.6)
No treatment	17 (10.7)
NA	23 (12.6)
AML transformation, N (%)	19 (11.0)
OS – months, median (95% CI)	143.0 (60.8–NR)
Follow-up – months, median (95% CI)	23.8 (18.5–35.1)
Reported deaths, N (%)	51 (28.0)

Abbreviations: IQR, interquartile range; M, male; F, female; BM, bone marrow; ANC, absolute neutrophil count; PLT, platelet count; MDS, myelodysplastic syndromes; SLD, single lineage dysplasia; MLD, multilineage dysplasia; RS, ring sideroblasts; EB, excess blasts; MDS-U, MDS, unclassifiable; MDS-NOS*, MDS, not otherwise specified with low blast count and missing data of dysplasia; CMML, chronic myelomonocytic leukemia; BSC, best supportive care; HMA, hypomethylating agents; HSCT, hematological stem cell transplantation; NA, not available; OS, overall survival; CI, confidence interval; NR, not reached.

Thirty-six subjects (19%) did not have any relevant variants, 68 (37.4%) had one variant, 46 (25.3%) had two variants, and 32 (17.5%) had three or more variants. The five most frequently mutated genes were TET2 (30%), ASXL1 (23%), SRSF2 (19%), SF3B1 (16%), and DNMT3A (14%) (Supplemental Data Fig. S1). At least one pathogenic molecular abnormality was identified in 112 (77.2%) subjects with MDS and in 34 (91.9%) subjects with MDS/MPN. Subjects with CMML exhibited a significantly higher variant burden (3.5 variants, IQR: 3–4) than subjects with MDS (2 variants, IQR: 2–3).

Univariate analysis revealed that age (≥70 yrs), Hb level, ANC, % BM blasts, abnormal karyotype, and IPSS-R cytogenetic score were clinical factors that significantly affected the prognosis in our cohort (P<0.05). Multivariable analysis showed that age ≥70 yrs (hazard ratio [HR]: 2.13; CI: 1.04–4.34; P=0.038), Hb level (HR: 0.75 [0.64–0.88]; P<0.001), ANC (HR: 1.05 [1.03–1.08]; P<0.001), PLT (HR: 1.00 [0.99–1.00]; P<0.011), and IPSS-R cytogenetic score (HR: 1.69 [1.21–2.36]; P=0.002) independently correlated with survival.

Patient restratification from IPSS-R to IPSS-M

A five-to-five comparison between the IPSS-R and IPSS-M risk categories (after merging the moderate-risk groups) resulted in the restratification of 82 patients (45%), of whom 64 (35%) were upstaged, and 18 (10%) were downstaged. Twelve (41%) patients in the very-low-risk IPSS-R group were upstaged into higher-risk categories (low or moderate), whereas 30 (44%) patients in the low-risk IPSS-R group were upstaged (moderate, high, or very high). In the intermediate-risk IPSS-R group, eight (22%) subjects were downstaged (very low or low risk), and 15 (42%) were reclassified as having a high or very high risk. In the high-risk IPSS-R group, four (20%) patients were downstaged (moderate risk), and seven (35%) were restratified as having a very high risk. In the very-high-risk IPSS-R group, four (31%) patients were downstaged to the high-risk group (Fig. 1A).

Figure 1. Clinical assessment of IPSS-R and IPSS-M for 182 patients. (A) Sankey plot for restratification from the IPSS-R to the IPSS-M model using five risk groups. (B) Kaplan–Meier probability estimates of OS across the IPSS-R (top) and IPSS-M (bottom) categories. The dashed gray lines represent the median values. The global P was calculated using the log-rank test. Each HR and associated P were calculated by setting the low-risk group as the reference.
Abbreviations: IPSS-R, International Prognostic Scoring System, revised version; IPSS-M, Molecular International Prognostic Scoring System; OS, overall survival; HR, hazard ratio; NE, not evaluable; mo, months.

The IPSS-R and IPSS-M models showed statistically significant correlation with OS (log-rank P=2×10^–6 and P=2×10^–7, respectively). The median OS estimates were not reached for the very-low-risk groups with both models. After stratifying patients with the IPSS-R model, the median OS was 84.7, 60.8, 53.6, and 16.5 months for the low-, intermediate-, high-, and very-high-risk groups. The estimates obtained using the IPSS-M model were similar, with OS times of 84.7, 53.6, 31.0, and 18.7 months for the low-, moderate-, high-, and very-high-risk groups, respectively (Fig. 1B).

Restratification with the molecular prognostic systems

To evaluate the applicability of the different prognostic models to real-world data, we compared the C-index values and AUCs of the IPSS-M, IPSS-R, AIPSS-MDS, EuroMDS, and MLL models. The latter two models are genetically based and intended for stratifying patients with MDS using molecular and cytogenetic information alone.

Prognostic power analyses considering the risk categories revealed that among the molecular-based systems, the IPSS-M demonstrated the highest discrimination power for OS (C-index: 0.75, standard error [SE]: 0.04), outperforming both the EuroMDS (C-index: 0.72, SE: 0.04) and MLL (C-index: 0.70, SE: 0.04) models. Among the non-molecular-based models, the AIPSS-MDS risk groups showed a higher C-index (0.74, SE: 0.04) than the IPSS-R model (0.70, SE: 0.04). Similar prognostic power was observed when the AUC was considered in ROC analysis. The AUCs of the IPSS-M (0.68 [95% CI: 0.59–0.77]) and EuroMDS (0.68 [0.59–0.77]) models were the highest, followed by those of the AIPSS-MDS risk-group (0.66 [0.58–0.75]), MLL (0.64 [0.54–0.74]), and IPSS-R (0.61 [0.52–0.70]) models. When comparing the prognostic power of the models using patient-level risk scores, the AIPSS-MDS showed greater power than the IPSS-M and IPSS-R models, highlighted by the C-index and AUC values (Fig. 2). Similar results were obtained when assessing the prognostic power of the IPSS-M (C-index: 0.76, SE: 0.04) and IPSS-R (C-index: 0.70, SE: 0.05) models after excluding patients with myeloproliferative characteristics (CMML-MP).

Figure 2. Prognostic power of several models for OS. The prognostic power was assessed based on (A) the AUC or (B) Harrell’s C-index determined with the IPSS-M, IPSS-R, AIPSS-MDS, EuroMDS, and MLL models. In panel B, the red line represents the risk encoded as categories using five risk categories (after merging the moderate-risk groups) for the IPSS-R and IPSS-M models and the quintiles using risk scores obtained from the original training set for the AIPSS-MDS model. The blue line represents a continuous score.
Abbreviations: OS, overall survival; AUC, area under the curve; C, concordance; IPSS-M, Molecular International Prognostic Scoring System; IPSS-R, International Prognostic Scoring System, revised version; AIPSS-MDS, Artificial Intelligence Prognostic Scoring System for MDS; EuroMDS, European MDS; MLL, Munich Leukemia Laboratory.

Implications of restratification for risk assessments

Among the 182 patients, 103 (57%) required active treatment. Three patients (2%) received chemotherapy, 57 (36%) received HMAs, and seven (4%) received a combination of chemotherapy and HMAs. Thirty-six patients (23%) underwent HSCT.

Analyzing the importance of restratification on patient therapy using simplified risk groups (low and high) for the IPSS-R and IPSS-M models showed that 23 (12.6%) patients would have been eligible for high-risk treatment options based on the IPSS-M prognosis, whereas nine (5%) patients would be considered for lower-intensity regimens.

We also evaluated the application of the AIPSS-MDS model, considering its high prognostic power. Unlike the IPSS-R and IPSS-M models, this model calculates a unique risk score for each patient (bypassing risk-group stratification). To evaluate the effects of the AIPSS-MDS restratification on patient therapy, we used an artificial threshold based on the median risk score obtained from the original training set used to model the AIPSS-MDS by Mosquera, et al. [15]. Using this approach, 53 (51%) patients in the low-risk IPSS-R group became eligible for high-risk treatment according to the AIPSS-MDS. With this model, no patients were downstaged to the low-risk category (Fig. 3B). The two-group stratification showed a statistically significant correlation with OS in the IPSS-R, IPSS-M, and AIPSS-MDS models, with the latter showing a longer median OS time for the high-risk group (Fig. 3D).

Figure 3. Variant distribution, restratification, and OS analysis with different risk models. (A, C) Violin plots showing the number of variants as the risk groups changed when switching from the IPSS-R model to (A) the IPSS-M model or (C) the AIPSS-MDS model, affecting patient therapy. (B) Sankey plot depicting the percentage of patients restratified among the models. (D) Kaplan–Meier probability estimates of the OS for low- and high-risk groups of patients with available survival data. The AIPSS-MDS risk groups were defined based on the median risk score from the original training set by Mosquera, et al. [¹⁵]. Risk groups were defined using a risk-score cut-off of 3.5 for IPSS-R and a cut-off of 0 for IPSS-M. Each HR and associated P were calculated by setting the low-risk group as the reference.
Abbreviations: OS, overall survival; IPSS-R, International Prognostic Scoring System, revised version; IPSS-M, Molecular International Prognostic Scoring System; AIPSS-MDS, Artificial Intelligence Prognostic Scoring System for MDS; HR, hazard ratio; NA, not available; mo, months.

The analysis of the number of variants in patients reclassified for high-risk treatment between the IPSS-R and IPSS-M models revealed a higher number of variants in these patients (2.0 [2.0, 3.0]) than in non-reclassified (1.0 [1.0, 2.0]) and downstaged (0.0 [0.0, 1.0]) patients (Fig. 3A, C). No differences in the number of variants were found in patients reclassified from the low-risk IPSS-R group to the high-risk AIPSS-MDS group (Supplemental Data Fig. S2).

MDS comprises a highly heterogeneous group of myeloid clonal diseases originated in the hematopoietic stem cells. The molecular characterization of MDS has improved the understanding of the biological mechanisms that drive the pathology. The evidence regarding the roles of genetic variants is still growing, as is their utility for differentiating biological MDS subgroups, determining disease-related risk, and facilitating treatment decision-making [8, 16].

We included 182 patients with retrospective clinical and molecular data to compare the applicability of molecular-based prognostic models (IPSS-M, EuroMDS, and MLL) and nonmolecular-based models (IPSS-R, AIPSS-MDS), as well as to compare their prognostic power with real-world data for patients with MDS.

In our cohort, the reclassification from IPSS-R to IPSS-M restratified 35% of patients to a higher-risk group, most of whom (25%) were only upstaged to the next risk category, which is expected given that both models share clinical and cytogenetic parameters [4, 9]. The IPSS-M model better separated the risk categories for OS, particularly for the higher-risk groups, and outperformed the IPSS-R, EuroMDS, and MLL models. Although the moderate-risk groups were merged because of sample-size limitations, we could validate the improvement of the prognostic power and the applicability of the IPSS-M compared with its former system and with other molecular prognostic systems using real-world data for patients from South America, in line with previously published studies [10, 12^-14].

Analyzing the implication of IPSS-M reclassification on the treatment choice for MDS showed that 23 (12.6%) patients could have been considered for higher-intensity treatments, similar to findings in other studied cohorts [17, 18]. Notably, 12 (52%) of the reclassified patients received HMAs, and six (26%) received HSCT. The increased rates of disease-modifying therapies (65%) might reflect the retrospective nature of our study and that most of the participating institutions are transplant centers with access to NGS technology. Incorporating those patients into the high-risk treatment group reflects the potential benefit that IPSS-M offers for better and earlier treatment selection when molecular data are available.

These results were obtained despite our sequencing data being generated with different sequencing panels and missing information regarding genomic alterations that are not commonly tested in our region, such as the copy-neutral loss of heterozygosity (CN-LOH) of TP53 and MLL-PTD, which are associated with a worse prognosis [19^-21]. The absence of such information may have affected score calculations, particularly for patients in lower-risk categories, since complete information for a set of 15 genes is required to achieve 80% accuracy, and missing data on the status of multiple TP53 variants alone can decrease the accuracy by 15% [13]. However, previous findings have shown that only 0.1%–0.6% of patients with MDS might have a TP53 variant with low frequency (VAF<50%) along with CN-LOH, whereas 1%–2.5% of patients have MLL-PTD variants [9, 19, 22]. In our cohort, missing data for these alterations might have affected only a low percentage of IPSS-M score calculations.

Recently, Mosquera, et al. [15] proposed a new prognostic model for patients with MDS using artificial intelligence (the AIPSS-MDS model), which circumvents the requirement for molecular data. Our results showed a similar prognostic power between the AIPSS-MDS (C-index: 0.74) and the IPSS-M (C-index: 0.75) models in terms of OS, resembling results obtained with larger cohorts [15]. Regarding the applicability of this model, the GESMD recommends assigning a high-risk status to patients with an estimated OS of <30 months [23], which might be useful for stratifying treatment groups considering that the AIPSS-MDS web calculator offers personalized predictions for OS [15]. The AIPSS-MDS model might provide a robust tool for prognosis and treatment decisions when the variant status is unavailable.

However, when the variant status is available, molecular data remain relevant for testing the presence of drug targets, such as luspatercept (when ring sideroblasts have an abundance of 5%–15% and are SF3B1-variant positive, according to the 2016 WHO guidelines), enasidenib (IDH2-variant positive), and response markers, such as TP53 variants for lenalidomide treatment and relapse post-transplantation [2, 24]. Similarly, the current guidelines from the WHO and the International Consensus Classification consider molecular information for classifying patients with MDS [1, 25]. The presence of multiple TP53 variants (or a variant with a VAF>50%) would suffice to classify five patients in our cohort into the new MDS-biTP53 category. Similarly, two patients in this study with NPM1 variants would be classified as AML according to the last WHO classification, given that it can be diagnosed irrespective of the blast count.

This study has some limitations because of its retrospective nature. Molecular data were generated at different institutions for clinical purposes following international guidelines for reporting variants detected with various commercial myeloid panels, excluding VUS and variants with a VAF of 2% to <5%, as this cut-off has varied in the past few years. These differences enabled the participating institutions to establish agreements regarding the interpretation of pathogenic variants and the selection of sequencing panels. Currently, most participating institutions have adopted the same myeloid panel based on capture methods, enabling the evaluation of 63 genes and helping minimize technical and interpretation biases.

In conclusion, the IPSS-M and AIPSS-MDS models show greater prognostic power for OS than the IPSS-R, EuroMDS, and MLL models. Avoiding molecular data with the AIPSS-MDS model might represent a better option to offer proper prognostic tools for all patients with MDS, particularly in resource-limited centers in our region, where molecular testing is not a standard clinical practice because of infrastructure and reimbursement issues. Larger cohorts and real-world studies are required to validate the clinical implications of both models and to provide information for risk-adapted strategies.

The authors thank the participating institutions for collaborating and sending the molecular and clinical patient data. The authors also thank Carlos Perez, M.D. and David Valcárcel, Ph.D. from the GESMD for their assessments of the AIPSS-MDS model and Irene Larripa, Ph.D. for reviewing the manuscript.

Lincango M and Belli CB contributed to the conception and design of the study, collected the data, interpreted the results, and wrote the preliminary and final draft of the manuscript. Lincango M analyzed the data. Mosquera Orgueira A performed blinded calculations for the AIPSS-MDS model. Castro M, Mela M, Navickas A, Grille S, Arbelbide J, and Basquiera AL participated in the clinical evaluations. Andreoli V, García Rivello H, Catalán AI, Delamer R, Asinari M, Agriello E, Bender A, Rahhal M, and Grille S performed targeted next-generation sequencing and interpreted the molecular data according to clinical practice. Grille S, Arbelbide J, Mosquera Orgueira A, and Basquiera AL contributed to revising the final version of the manuscript. All authors have read and approved the final manuscript.

None declared.

This work was not supported by specific funding. Belli CB received grants from the Agencia Nacional de Promoción Científica y Tecnológica (ANPCyT) (grant Nos. PICT-2019-1623 and PICT-2021-0585) and the Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET) (grant No. PIP 2232).

Supplementary materials can be found via https://doi.org/10.3343/alm.2024.0089