Blood Cancer Journal www.nature.com/bcj ARTICLE OPEN Unsupervised machine learning improves risk stratification in newly diagnosed multiple myeloma: an analysis of the Spanish Myeloma Group Adrian Mosquera Orgueira 1,33, Marta Sonia González Pérez1,33, Jose Diaz Arias1, Laura Rosiñol 2, Albert Oriol 3, Ana Isabel Teruel4, Joaquin Martinez Lopez5, Luis Palomera6, Miguel Granell 7, Maria Jesus Blanchard8, Javier de la Rubia 9 , Ana López de la Guia10, Rafael Rios 11, Anna Sureda12, Miguel Teodoro Hernandez 13, Enrique Bengoechea14, María José Calasanz15, Norma Gutierrez16, Maria Luis Martin5, Joan Blade 2, Juan-Jose Lahuerta 5, Jesús San Miguel 15, ✉ Maria Victoria Mateos 16 and the PETHEMA/GEM Cooperative Group* © The Author(s) 2022 The International Staging System (ISS) and the Revised International Staging System (R-ISS) are commonly used prognostic scores in multiple myeloma (MM). These methods have significant gaps, particularly among intermediate-risk groups. The aim of this study was to improve risk stratification in newly diagnosed MM patients using data from three different trials developed by the Spanish 1234567890();,: Myeloma Group. For this, we applied an unsupervised machine learning clusterization technique on a set of clinical, biochemical and cytogenetic variables, and we identified two novel clusters of patients with significantly different survival. The prognostic precision of this clusterization was superior to those of ISS and R-ISS scores, and appeared to be particularly useful to improve risk stratification among R-ISS 2 patients. Additionally, patients assigned to the low-risk cluster in the GEM05 over 65 years trial had a significant survival benefit when treated with VMP as compared with VTD. In conclusion, we describe a simple prognostic model for newly diagnosed MM whose predictions are independent of the ISS and R-ISS scores. Notably, the model is particularly useful in order to re-classify R-ISS score 2 patients in 2 different prognostic subgroups. The combination of ISS, R-ISS and unsupervised machine learning clusterization brings a promising approximation to improve MM risk stratification. Blood Cancer Journal (2022)12:76 ; https://doi.org/10.1038/s41408-022-00647-z INTRODUCTION and most patients (62%) were classified as intermediate risk (R-ISS The International Staging System (ISS) has been the most used 2). It has become progressively evident that some patients who prognostic score employed for risk stratification in newly diagnosed belong to the R-ISS 1 low-risk group have poor survival, whereas Multiple Myeloma (MM) patients. This score is based on surrogate the outcome of patients in the intermediate group (R-ISS 2) is very markers of myeloma cell biology and host factors: ß2-microglobulin heterogeneous. Additionally, recent reports highlight that both ISS and albumin [1]. The ISS stratifies patients in three subgroups with an and R-ISS have similar predictive performance, suggesting that overall survival (OS) of 62, 44, and 22 months respectively, and it has optimized data exploitation tools might help to bring improved been validated in several studies and clinical trials. The main risk stratification techniques to the field [3]. All these issues limitation of this model is that it does not incorporate any genetic or highlight the limitations of these scores, which fail to properly proliferation biomarkers of the disease. stratify many patients. A revision of the ISS was presented in 2015 which incorporated Survival prediction of patients with hematological cancer has elevated lactate dehydrogenase (LDH) plus t(4;14), t(16;14) & del been extensively improved in the last years. For example, several (17p) as high-risk cytogenetics abnormalities [2]. This score biomarker panels based on next-generation sequencing of identified 3 risk groups with a median OS of 43, 83 months and recurrently mutated or aberrantly expressed genes have been not reached. Only 10% of patients were allocated to the high-risk proposed to facilitate prognostic stratification in acute myeloid group (R-ISS 3), 28% were assigned to the low-risk group (R-ISS 1) leukemia, myelodysplastic syndromes and lymphomas, and 1 Hospital Clínico Universitario Santiago de Compostela, A Coruña, Spain. 2Hospital Clínic, Institut d’investigacions Biomèdiques August Pi i Sunyer, Barcelona, Spain. 3Institut Català d’Oncologia I Institut Josep Carreras, Hospital Germans Trias i Pujol, Badalona, Spain. 4Hospital Clínico de Valencia, Valencia, Spain. 5Hospital Universitario 12 de Octubre, i +12, Complutense University, CNIO, Madrid, Spain. 6Hospital Clínico Lozano Blesa, Zaragoza, Spain. 7Hospital Sant Pau, Barcelona, Spain. 8Hospital Ramón y Cajal, Madrid, Spain. 9 Hospital Doctor Peset, Valencia, Spain. 10Hospital Universitario La Paz, Madrid, Spain. 11Hospital Virgen de las Nieves, CIBERESP, Ibs, Granada, Spain. 12Institut Català d’Oncologia- Hospitalet, IDIBELL, Universitat de Barcelona, Barcelona, Spain. 13Hospital Universitario de Canarias, Santa Cruz de Tenerife, Spain. 14Hospital de Donostia, San Sebastian, Spain. 15 Clínica Universidad de Navarra, CIMA, CIBERONC, IDISNA, Pamplona, Spain. 16Hospital Universitario de Salamanca, Instituto de Investigación Biomédica de Salamanca, Instituto de Biología Molecular y Celular del Cáncer (Universidad de Salamanca-Consejo Superior de Investigaciones Científicas), CIBERONC, Salamanca, Spain. 33These authors contributed equally: Adrian Mosquera Orgueira, Marta Sonia González Pérez. *A list of authors and their affiliations appears at the end of the paper. ✉email: mvmateos@usal.es Received: 22 December 2021 Revised: 5 March 2022 Accepted: 11 March 2022 A. Mosquera Orgueira et al. 2 indeed various studies have proved that these novel personalized (Group C). Eligible patients underwent autologous stem cell transplanta- models fitted with machine learning algorithms outperform the tion, and 3 months after transplant patients were randomized to three precision of currently established prognostic tools [4–6]. More different maintenance arms: either Interferon a-2b, thalidomide or recently, other sophisticated risk stratification methods, using thalidomide plus bortezomib for 2 years. gene expression profiling, comprehensive cytogenetic assess- Patients included in the GEM05 over 65 years trial were randomized 1:1 to receive melphalan + prednisone + bortezomib (VMP, Group A) or ments or next generation sequencing strategies have been thalidomide + prednisone + bortezomib (VTD, Group B). All patients published, but in clinical practice, these are rarely employed due received induction treatment for up to 30 weeks. Patients were further to the lack of availability, high cost as well as technical and randomized 1:1 to receive maintenance treatment: either Thalidomide + logistical difficulties [7–9]. Bortezomib (Group M1) or Prednisone + Bortezomib (Group M2) for three Therefore, improved risk stratification of MM with ready-to-use years after four weeks if no progression or toxicity. information is much awaited. In this line, the recent development Patients included in the GEM2012 under 65 years trial were treated with of machine learning in medicine has become key to overcome six cycles of induction treatment with bortezomib + lenalidomide + some of the limitations of classical prognostic scores. Machine dexamethasone (VRD). After induction, patients were randomized 1:1 to receive an autologous transplant with melphalan 200 mg/m2 (MEL200) learning is a field of artificial intelligence where prediction is based versus Busulfan 12 mg/kg plus melphalan 140 mg/m2 (BUMEL) as on the modeling of outcomes considering complex interactions conditioning regimens. Three months after transplantation, patients between multiple variables derived from real examples, rather received two cycles of consolidation treatment with VRD at the same than on the application of human-made rules. In the particular doses administered during induction treatment. Those patients in case of MM, such advanced techniques can optimize the number response after two cycles of consolidation therapy with VRD were further of prognostic groups and the assignment of patients to these included in a second maintenance trial (GEM MAIN 2014), being according to flexible data structures, instead of the rigid thresh- randomized to lenalidomide +/− ixazomib for 2–5 years depending on olds implemented in the current clinical scores. With this in mind, MRD analysis. we have developed a new unsupervised machine learning model Cytogenetic analysis was performed using fluorescence in situ hybridi- for MM risk stratification by integrating clinical, biochemical and zation (FISH) on whole bone marrow (GEM05 trials) or CD138-selected plasma cells (GEM2012 trial), and included t(4;14), t(14;16) and 17p cytogenetic data at diagnosis through the use of datasets deletion in all trials. Among the remaining annotations, the following corresponding to series of MM patients homogeneously treated common baseline variables were retrieved: immunoglobulin light and in the context of clinical trials conducted by the Spanish Myeloma heavy chain type, Durie-Salmon staging, monoclonal spike in blood and Group. Our results indicate that this strategy can significantly urine, hemoglobin, creatinine, albumin, albumin-adjusted calcium, ß2- improve MM prognostication, particularly among patients microglobulin, elevated LDH and percentage of plasma cells in bone assigned to the R-ISS 2 intermediate-risk group. marrow aspirate smear. Note that in the case of Durie-Salmon annotation, both parts of the staging system were analyzed separately. In this regard, one variable analyzed the presence of kidney disfunction (Durie-Salmon stages A and B), whereas another variable reflected the classification of MATERIALS AND METHODS patients in 3 groups (Durie-Salmon stages I, II & III) based on hemoglobin, Data source calcium, presence of bone X-ray abnormalities/plasmocitomes and We retrieved original data from three clinical trials developed by the monoclonal components in serum and urine. Patients who had incomplete Spanish Myeloma Group (Grupo Español de Mieloma, GEM), namely GEM05 annotation for any of the variables were discarded from downstream under 65 years [10], GEM05 over 65 years [11] and GEM2012 under 65 analysis. years [12]. All trials evaluated different upfront treatments in newly Overall survival (OS) was defined as time from diagnosis to death from diagnosed MM. any cause, and progression-free survival (PFS) was defined as time from Patients included in the GEM05 under 65 years trial were randomized diagnosis to disease progression or death from any cause. (1:1:1) to receive 4 alternating cycles of vincristine, BCNU, cyclopho- sphamide, melphalan and prednisone (VBMCP) - vincristine, BCNU, adriamycin, dexamethasone (VBAD) + 2 cycles of bortezomib (Group A) Variable selection and model development or 6 cycles of thalidomide + dexamethasone (TD) (Group B) or Statistical analysis was performed on R version 4.1.0 [13]. Survival analysis thalidomide + dexamethasone + bortezomib (VTD) during 24 weeks was performed with the “survival” package version 3.2.11 [14]. Univariate Table 1. Baseline characteristics of selected patients in the different trials. GEM05 under 65 GEM05 over 65 GEM2012 under 65 N 305 218 229 % High Risk Cytogenetics 19.34% 18.80% 26.63% Durie-Salmon stages: I, II & III 6.23%, 48.85%, 44.92% 7.34%, 51.83%, 40.83% 10.48%, 38.43%, 51.09% Durie-Salmon stages A & B 96.40%, 3.60% 95.87%, 4.13% 97.82%, 2.18% ISS stages: I, II, III 38.36%, 41.31%, 20.33% 24.31%, 43,58%, 32.11% 43.23%, 30.57%, 26.20% RISS stages: I, II, III 28.52%, 62.62%, 8.85% 19.72%, 70.64%, 9.63% 27.94%, 61.57%, 10.48% Median serum monoclonal spike (g/dL) 3.9 3.6 2.8 Median urine monoclonal spike (g/dL) 0.19 0.20 0.14 Median hemoglobin (g/dL) 10.8 10.4 11.1 Median albumin-adjusted calcium (mg/dL) 9.68 9.95 9.58 Median B2-microglobulin (mg/dL) 3.3 4.0 3.4 Raised LDH 15.73% 12.84% 16.52% Median plasma cells in bone marrow smear 36% 35% 28% Presence of major myeloma-related skeletal injuries 33.44% 25.23% 35.81% Presence of plasmocitomes 17.05% 13.30% 22.71% Blood Cancer Journal (2022)12:76 A. Mosquera Orgueira et al. 3 cox-regression was used to test the association of each variable with Under the Curve (AUC) were calculated at different time points with 500 overall survival in the largest cohort (GEM05 under 65 years). Principal bootstraps using the riskRegression package version 2021.10.10 [17]. components were extracted using Factor Analysis for Mixed Data (FAMD) Model’s discriminative power was assessed using 10-fold cross-validated implemented in the FactoMineR version 2.4 package [15], which can accept Harrel’s concordance indexes (c-indexes) implemented in the RMS package continuous and categorical variables as input. In each case, we selected as version 6.2.0 [18]. Finally, survival curves were plotted using the Kaplan- many principal components as variables included in the model. Meier method. Unsupervised clustering was performed using Gaussian Mixture Model- ing fitted with an Expectation Maximization algorithm (GMM-EM model implemented in the “Mclust” algorithm version 5.4.7) [16]. Briefly, the RESULTS Mclust algorithm determines the most likely set of patient clusters Variable selection and unsupervised model fitting according to geometric properties (distribution, volume, and shape). An EM algorithm is used for maximum likelihood estimation, and the best Baseline characteristics of the patients included in each cohort are model is selected according to Bayes information criteria. Inferred clusters represented in Table 1. The first analysis was done in the cohort of in the GEM05 under 65 years trial were used to predict clusterizations in patients included in the GEM05 under 65 years trial, as this was the remaining datasets (GEM05 over 65 years and GEM2012 under 65 the study with the largest number of patients with complete years). Cox regression was used to analyze the association of such clusters annotation data available. We identified 10, 14 and 16 variables with OS and PFS, as well as their relationship with the International Staging which were associated with OS at p value thresholds of 0.01, 0.05 System (ISS) and revised ISS (R-ISS) scores. Assumption of proportional and 0.1 (Table 2). FAMD decomposition was performed taking as hazards was tested with Schoenfeld’s method. Model’s precision was input all cytogenetic variables (17p deletion, t(4;14), t(14;16) and assessed using cross-validated cox models and time-dependent Area any high risk cytogenetic alteration) plus any of the remaining variables at p-value thresholds of 0.01, 0.05 and 0.1. Afterwards, GMM-EM was implemented to identify the optimal two clusters in the database (Table 3). Clusterization results were significantly Table 2. Cox regression testing the association of the 18 variables associated with OS regardless of the p value threshold used, but with overall survival in the GEM05 under 65 years cohort. statistical significance was superior with the p value threshold of Variable p value 0.01 (p value 7.44 × 10−8, HR 0.35) compared with the p value thresholds of 0.05 (p value 1.63 × 10−4, HR 0.47) and 0.1 (p value High risk cytogenetics 5.36 × 10−5 2.65 × 10−5, HR 0.42). t(14;16) 0.04 Aside from cytogenetic data, this optimal model included the 17p deletion 0.09 following variables: Durie-Salmon staging (I, II and III), hemoglobin, t(4;14) 1.91 × 10−5 albumin-adjusted calcium, ß2-microglobulin and elevated LDH. We also tested the prognostic impact of GMM-EM-based models Immunoglobulin subclass 0.51 with 3 and 4 optimal clusters using this set of variables, but results Durie-Salmon stages (I, II & III) 1.86 × 10−3 were inferior to the model based on 2 optimal clusters Durie-Salmon stages (A & B) 0.21 (Supplementary Fig. 1). Serum M spike 0.98 Predictions created by this model on GEM05 under 65 years were confirmed on GEM05 over 65 years and GEM2012 and two Urine M spike 0.15 different clusters of patients were identified in both cases (Fig. 1). Hemoglobine 2.06 × 10−3 Additionally, this clusterization was significantly associated with Creatinine 0.80 OS in both cohorts (cox p value 1.42 × 10−3, HR 0.36 in the Albumin 0.01 GEM2012 under 65 years cohort & cox p value 8.07 × 10−5, HR 0.51 in the GEM05 over 65 years cohort; Table 3). Albumin-adjusted calcium 4.49 × 10−3 Furthermore, the clusterization was significantly associated B2-microglobulin 9.29 × 10−8 with PFS in all cohorts. (Fig. 1). The characteristics of cluster 1 Raised LDH 1.75 × 10−6 and cluster 2 patients according to the constituting variables of % of bone marrow plasma cells 0.04 the unsupervised model are represented in the Supplementary Table 1. Table 3. Distribution of 2 clusters detected with unsupervised clustering across cohorts, as well as cox regression testing the association with overall survival and progression-free survival. GEM05 under 65 GEM05 over 65 GEM2012 under 65 % patients in each cluster 36.72%, 63.28% 34.86%, 65.13% 44.10%, 55.90% p value OS 7.44 × 10−8 8.07 × 10−5 1.42 × 10−3 HR (95% CI) for OS 0.35 [0.24, 0.52] 0.51 [0.36, 0.71] 0.36 [0.19, 0.68] p value PFS 2.48 × 10−4 1.16 × 10−3 5.47 × 10−4 HR (95% CI) for PFS 0.60 [0.45, 0.79] 0.60 [0.45, 0.82] 0.50 [0.34, 0.74] p value OS (RISS adjusted) 0.01 9.85 × 10−3 0.02 HR (95% CI) for OS (RISS adjusted) 0.56 [0.36, 0.87] 0.591 [0.40 0.88] 0.42 [0.20, 0.89] p value PFS (RISS adjusted) 0.15 0.12 0.02 HR (95% CI) (RISS adjusted) 0.79 [0.57, 1.09] 0.75 [0.53, 1.07] 0.57 [0.36, 0.92] p value OS (ISS adjusted) 1.96 × 10−5 1.01 × 10−3 3.47 × 10−3 HR (95% CI) for OS (ISS adjusted) 0.42 [0.28, 0.63] 0.55 [0.39, 0.79] 0.39 [0.20, 0.73] p value PFS (ISS adjusted) 5.37 × 10−3 0.01 1.24 × 10−3 HR (95% CI) (ISS adjusted) 0.66 [0.50, 0.89] 0.67 [0.49, 0.91] 0.52 [0.35, 0.77] Blood Cancer Journal (2022)12:76 A. Mosquera Orgueira et al. 4 Fig. 1 Patient outcomes according to the novel prognostic score. Kaplan–Meier curves representing the impact of the 2 clusters detected through unsupervised machine learning on overall survival (OS) and progression-free survival (PFS) in the 3 trial cohorts. “P” symbol indicates p-value. A, B OS and PFS for the GEM2005 under 65 years trial. C, D OS and PFS for the GEM2012 under 65 years trial. E, F OS and PFS for the GEM2005 over 65 years trial. Blood Cancer Journal (2022)12:76 A. Mosquera Orgueira et al. 5 between these subgroups in the three different clinical trials Table 4. Patient distribution according to ISS/R-ISS scores and evaluated, with the exception of a significant difference between unsupervised clustering results. R-ISS 1 and R-ISS 2 low-risk groups in the GEM05 under 65 GEM05 under 65 Cluster 1 Cluster 2 Cluster 1 vs Cluster 2 years trial. p value A focused analysis on R-ISS 2 patients evidenced that the key variables to stratify patients in two different risk clusters was the ISS 1 9.09% 29.18% <1 × 10−4 presence of high-risk cytogenetics or raised LDH, as all R-ISS 2 ISS 2 15.74% 25.57% 0.02 patients with any of these features were assigned to the higher- ISS 3 11.80% 8.52% 0.21 risk cluster I. However, a variable proportion of R-ISS 2 cluster I RISS 1 0.98% 27.54% 0.49 patients had standard-risk cytogenetics and normal baseline LDH. Importantly, their survival was similar to that of patients with high RISS 2 26.88% 35.74% 6.10 × 10−3 risk cytogenetics or raised LDH (Supplementary Table 3), reinfor- RISS 3 8.85% 0% NA cing their membership to the higher risk cluster. GEM05 over 65 Cluster 1 Cluster 2 Cluster 1 vs Cluster 2 OS p value Discriminative power and precision of the different risk stratification models ISS 1 9.18% 28.18% 0.32 We used cox c-indexes to investigate the discriminative capacity ISS 2 15.74% 25.57% 0.06 of each score and score combination on the different cohorts ISS 3 11.80% 8.52% 0.02 (Table 5). R-ISS was superior to ISS only in GEM05 under 65 years, RISS 1 0.98% 27.54% NA and both scores performed similarly in the remaining datasets. Additionally, our unsupervised clusterization model achieved RISS 2 26.89% 35.74% 8.20 × 10−3 superior c-indexes than ISS in all cohorts, superior c-indexes than RISS 3 8.85% 0%% NA R-ISS in the cohorts of patients included in the GEM05 over 65 GEM2012 under 65 Cluster 1 Cluster 2 Cluster 1 vs Cluster 2 years and GEM12 under 65 years, and similar concordance to R-ISS OS p value in the GEM05 under 65 years cohort. Importantly, the combination ISS 1 15.72% 27.51% 0.57 of ISS and R-ISS achieved inferior c-indexes than any of the ISS 2 11.79% 18.78% 4.50 × 10−3 combinations that included our unsupervised clusterization model. 10-fold cross-validation confirmed the robustness of the ISS 3 16.59% 9.61% 0.08 prognostic clusters (Supplementary Table 4). RISS 1 1.31% 26.63% 0.52 Time-dependent AUCs revealed that the precision in survival RISS 2 32.75% 28.82% 0.01 prediction of the unsupervised clusterization was clearly superior RISS 3 10.04% 0.44% 0.55 to ISS and R-ISS in most cases (Supplementary Fig. 5). In the case of GEM 2005 < 65 years, we observed a less superior performance Statistical significance (cox p values) for differential OS between both when compared to the R-ISS. However, this cohort was used to clusters in each subgroup is shown. develop the R-ISS score itself, so some degree of overfitting could exist [2]. Relationship of unsupervised clusterization with treatment Relationship of unsupervised clusterization with ISS and R-ISS outcomes in the different clinical trials The unsupervised clusterization model was associated with OS In the GEM05 under 65 years trial, survival of both clusters of independently of ISS and R-ISS scores in all cohorts (multivariate patients was similar regardless of the assignment to the different cox p value <0.05, Table 3). Additionally, the clusterization was induction or post-transplant maintenance arms. Similarly, survival associated with PFS independently of ISS stages in all cases, of the two clusters was similar regardless of the type of although the GEM2012 under 65 years trial was the only conditioning regime evaluated in the GEM2012 under 65 years population in which it was independent of R-ISS stage (Table 4). trial. Finally, a benefit in terms of OS was identified for patients in These findings motivated a subanalysis by ISS and R-ISS scores the low-risk cluster when treated with VMP compared with VTD in (Table 4, Supplementary Fig. 2 and 3). Interestingly, we observed the GEM05 over 65 years trial (p value 0.03, Fig. 4). However, no that the unsupervised model was particularly useful to stratify significant differences were observed between both clusters when patients with R-ISS 2 into two clusters with significantly divergent considering the different maintenance strategies performed in this OS curves in all cohorts (Fig. 2). Indeed, only a minority of patients trial. with either R-ISS 1 or R-ISS 3 were reclassified to a higher or lower risk classification, respectively: 1 R-ISS 3 patient was assigned to the lower risk cluster, and 6 R-ISS 1 patients were assigned to the DISCUSSION higher risk cluster. On the contrary, remarkable changes in risk In the present work, we describe a new prognostic classification of group assignment were observed between ISS scores and these newly diagnosed MM based on the application of intelligent new risk clusters (Fig. 3). information technologies to clinical trial data produced by the On the other side, the distribution of ISS scores in patients Spanish Myeloma Group. During the last 20 years, different assigned to both prognostic clusters was more heterogeneous, prognostic models have been developed to stratify newly and our new clusterization tended to provide additional diagnosed MM patients, among which ISS and R-ISS are the most prognostic stratification in most cases (particularly in ISS 2 and 3 commonly used. Both models appear useful in identifying a small groups; Table 4). subgroup of high risk patients. However, the main limitation resides in the fact that most patients are categorized in low or Survival analysis of R-ISS 2 subgroups intermediate categories. Unfortunately, some patients classified as We analyzed the survival of the two new clusters of R-ISS 2 low-risk (ISS 1 and R-ISS 1) have a short survival, whereas the patients identified with this novel approach (low-risk and high- majority of patients included in the largest R-ISS 2 group have risk), and compared them with the survival of patients with R-ISS 1 unpredictable outcomes.Therefore, new approximations based on and 3 MM, respectively (Supplementary Table 2, Supplementary advanced data analytics are needed in the prognostic stratification Fig. 4). Interestingly, we found no significant differences in survival of MM patients. Blood Cancer Journal (2022)12:76 A. Mosquera Orgueira et al. 6 Fig. 2 Survival of R-ISS 2 patients according to the new score. Impact of the 2 clusters detected with unsupervised machine learning on overall survival of R-ISS 2 MM patients across the 3 trial cohorts, namely GEM2005 under 65 years trial (A), GEM2005 over 65 years trial (B) and GEM2012 under 65 years trial (C). Several reasons may explain the limitations of these scores. For using high-throughput genomic analysis that are more sophis- example, 19% of patients belonging to ISS 1 and 25% of patients ticated than FISH [7–9]. Although these approaches seem very in ISS 2 present high risk cytogenetic abnormalities [2]. In addition promising, they are not easily applicable in real life due to cost to that, the choice of cytogenetic abnormalities included in R-ISS and technological complexity. Beyond genetic factors, there are may be suboptimal. Indeed, some authors have associated other other well-known prognostic factors such as renal failure, cytogenetics abnormalities with long-term outcomes (e.g., 1p extramedullary disease or circulating plasma cells that are not deletion and 1q amplification) [19–21]. Moreover, it has been openly present in either ISS nor R-ISS scores [24–26]. Notably, a reported that the weight of each cytogenetic alteration may be large list of new prognostic factors (socioeconomic status, some different (additive score) [22], and that some cytogenetic comorbidities, frailty score, diagnostic delay, the specific type of abnormalities can “modulate” the effect of others. In fact, a type myeloma defining event, immunoparesia…) or those with a of double hit myeloma with very poor prognosis has been recently lower level of evidence could also impact the outcome but they described, which is characterized by either biallelic TP53 inactiva- are not usually taken into account [27]. Furthermore, MM clinical tion or amplification (≥4 copies) of CKS1B (1q21) on the evolution is recurrently impacted by the appearance of new background of ISS III [23]. drugs & drug combinations, and the quality of the response to In an attempt to optimize prognostic scores, some groups these drugs is one of the most powerful prognostic factors. have reported prognostic models based on genetic factors, Although current models cannot be used to predict optimal Blood Cancer Journal (2022)12:76 A. Mosquera Orgueira et al. 7 schemes, some groups, including ours [28], are working in this chemotherapy and interferon as maintenance in one of its area with encouraging results. arms, we have subsequently validated the model in 2 other New information technologies, such as big data and machine clinical trials which used new drugs (proteasome inhibitors and learning algorithms, provide the opportunity to create more immunomodulatory drugs) with similar results. Importantly, all precise models in order to optimize risk stratification based on variables included in this model are readily available to any individual clinical and biological factors. Particularly, unsuper- patient in clinical practice and there is no need for sophisticated vised machine learning algorithms (as the one used in this technologies. paper) come along with substantial benefits when identifying The unsupervised model identified 2 clusters of patients with patient subgroups. Importantly, there is no prior assumption different PFS and OS independently of ISS and R-ISS scores in all about cluster composition, as these are inferred from complex cohorts. More interestingly, our model was particularly useful to patterns in the data without the need to provide human-made stratify patients with R-ISS score 2 into 2 clusters with instructions. Therefore, this strategy does not rely on simple significantly divergent OS curves in the 3 cohorts, and of note, optimal cut-offs, but can instead identify which is the most likely all patients with high LDH or high-risk cytogenetics were composition of patient clusters and improve the assignment of assigned to the high risk cluster. To our knowledge, this is the each patient even though he or she might be in the boundaries first model that enables such differentiation. Importantly, the of the distributions. In this way, unsupervised machine learning model retained its predictive power independently of induction strategies help researchers to maximize the value of the data by type, transplantation conditioning and the different mainte- facilitating the conversion of multidimensional data into nance schemes. Additionally, time-dependent AUCs and simplified, optimized and reproducible clusterizations. Using c-indexes indicated that the new clusterization was superior to such an approach, we have created a simple and easy to use either ISS or R-ISS in most cases. Despite the fact that both the prognostic model based on 9 clinical and biological variables ISS and R-ISS share some variables with the new prognostic which arose from patients included in GEM05 under 65 years model, the latter provides additional discriminative value to the clinical trial. Although this trial included conventional former two. Future efforts should pursue an optimal entangling of the variables included in this new model with those of the ISS and R-ISS, so as to achieve a single model on top of these that can integrate all the prognostic information into well-defined prognostic groups. Finally, the unsupervised clusterization model identified a subgroup of low risk patients who had longer OS when VMP was used as induction compared to VTD in the GEM05 over 65 years trial. Although this information might not be very relevant in clinical practice today because of new standards of care, it must be evaluated whether this model can help us to individualize the best option of therapy in the setting of the new standards of care. This analysis emphasizes the importance of an optimal application of information technologies to patient data in order to improve disease prognostication. Even though machine learning models are frequently developed with big chunks of data which might hinder their broad applicability [29], our results indicate that it is possible to significantly improve disease prognostication by re- interpreting a limited number of classical variables. Therefore, relevant scientific advances might be achieved in similar scenarios by revisiting relatively small amounts of data. The main limitations of this study reside on the relatively short sample size of the trials, the lack of patients treated with immunotherapy and the geographical restriction of the trials to Fig. 3 Transition plots between ISS scores and unsupervised risk Spain. Other pitfalls, such as the lack of complete annotation for clusters in the 3 different clinical trials evaluated. ISS scores are some relevant clinical and cytogenetic variables in all trials (e.g., represented on the left column of each graph, and unsupervised performance status and chromosome 1 abnormalities), suggest clusters are represented on the right side. Transition plots for the the existence of significant room for improvement. Future GEM2005 under 65 years, GEM2005 over 65 years and GEM2012 advances in MM prognostication should move in these under 65 years trials are represented in plots A, B and C, respectively. directions. Table 5. C-indexes and corresponding standard errors in cox regression including ISS scores, R-ISS scores and unsupervised clustering results. GEM05 under 65 GEM2012 under 65 GEM05 over 65 ISS 0.619 (0.026) 0.596 (0.039) 0.577 (0.022) RISS 0.653 (0.02) 0.606 (0.033) 0.570 (0.02) UNSUPERVISED MODEL 0.645 (0.023) 0.636 (0.035) 0.593 (0.021) ISS + RISS 0.652 (0.024) 0.618 (0.038) 0.591 (0.024) ISS + UNSUPERVISED MODEL 0.696 (0.023) 0.664 (0.04) 0.62 (0.025) RISS + UNSUPERVISED MODEL 0.694 (0.023) 0.653 (0.038) 0.607 (0.024) ISS + RISS + UNSUPERVISED MODEL 0.704 (0.023) 0.661 (0.04) 0.621 (0.025) Univariate and multivariate cox regression models were fitted. Blood Cancer Journal (2022)12:76 A. Mosquera Orgueira et al. 8 9. van Beers EH, van Vliet MH, Kuiper R, de Best L, Anderson KC, Chari A, et al. Prognostic validation of SKY92 and its combination with iss in an independent cohort of patients with multiple myeloma. Clin Lymphoma Myeloma Leuk. 2017;17:555–62. 10. Rosiñol L, Oriol A, Teruel AI, Hernández D, López-Jiménez J, de la Rubia J, et al. Superiority of bortezomib, thalidomide, and dexamethasone (VTD) as induction pretransplantation therapy in multiple myeloma: a randomized phase 3 PETHEMA/GEM study. Blood 2012;120:1589–96. 11. Mateos MV, Oriol A, Martínez-López J, Teruel AI, López de la Guía A, López J, et al. GEM2005 trial update comparing VMP/VTP as induction in elderly multiple myeloma patients: do we still need alkylators? Blood 2014;124:1887–93. 12. Rosiñol L, Oriol A, Rios R, Sureda A, Blanchard MJ, Hernández MT, et al. Borte- zomib, lenalidomide, and dexamethasone as induction therapy prior to auto- logous transplant in multiple myeloma. Blood 2019;134:1337–45. 13. R Core Team. R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing; 2020. https://www.R-project.org/. 14. Therneau TM, Grambsch PM. Modeling Survival Data: Extending the Cox Model. New York: Springer; 2000. ISBN 0-387-98784-3 15. Lê S, Josse J, Husson F. FactoMineR: An R package for multivariate analysis. J Stat Softw. 2008;25:1–18. 16. Scrucca L, Fop M, Murphy TB, Raftery AE. mclust 5: clustering, classification and density estimation using gaussian finite mixture models. R J 2016;8:289–317. 17. Gerds TA, Kattan MW (2021). Medical Risk Prediction Models: With Ties to Machine Learning (1st ed.). Chapman and Hall/CRC. https://doi.org/10.1201/9781138384484. Fig. 4 Impact of the new score system on drug response. 18. FE Harrell Jr (2017). rms: Regression Modeling Strategies. R package version 5.1-0. Representation of overall survival curves of patients belonging to https://CRAN.R-project.org/package=rms Cluster 2 treated with VMP or VTD. 19. Schmidt TM, Fonseca R, Usmani SZ. Chromosome 1q21 abnormalities in multiple myeloma. Blood Cancer J. 2021;11:83. 20. Qazilbash MH, Saliba RM, Ahmed B, Parikh G, Mendoza F, Ashraf N, et al. Deletion In conclusion, the present work describes a new, simple and of the short arm of chromosome 1 (del 1p) is a strong predictor of poor outcome easy to use prognostic model in newly diagnosed MM whose in myeloma patients undergoing an autotransplant. Biol Blood Marrow Transpl. 2007;13:1066–72. predictions are independent of ISS and R-ISS scores. Notably, the 21. Weinhold N, Salwender HJ, Cairns DA, Raab MS, Waldron G, Blau IW, et al. model is particularly well suited in order to classify R-ISS score 2 Chromosome 1q21 abnormalities refine outcome prediction in patients with patients in 2 subgroups with significantly different survival. The multiple myeloma - a meta-analysis of 2,596 trial patients. Haematologica reproduction of this clusterization in different MM databases 2021;106(Oct):2754–8. https://doi.org/10.3324/haematol.2021.278888. PMID: developed by other national and international working groups is 34092058 recommended, and their associations with drug response in 22. Perrot A, Lauwers-Cances V, Tournay E, Hulin C, Chretien ML, Royer B, et al. clinical trials should be studied. The combination of ISS, R-ISS and Development and validation of a cytogenetic prognostic index predicting sur- unsupervised machine learning clusterization is a promising vival in multiple myeloma. J Clin Oncol. 2019;37:1657–65. approximation in order to improve MM risk stratification. 23. Walker BA, Mavrommatis K, Wardell CP, Ashby TC, Bauer M, Davies F, et al. A high- risk, double-hit, group of newly diagnosed myeloma identified by genomic analysis. Leukemia. 2019;33:159–70. 24. Liu A, Yang G, Geng C, Wang H, Li L, Li Y, et al. A report of 134 newly diagnosed REFERENCES multiple myeloma patients with renal impairment. Acta Haematol. 2016;135: 1. Greipp PR, San Miguel J, Durie BG, Crowley JJ, Barlogie B, Bladé J. et al. Inter- 140–5. national staging system for multiple myeloma. J Clin Oncol. 2005;23:3412–20. 25. Sevcikova S, Minarik J, Stork M, Jelinek T, Pour L, Hajek R. Extramedullary disease https://doi.org/10.1200/JCO.2005.04.242 in multiple myeloma - controversies and future directions. Blood Rev. 2019;36 2. Palumbo A, Avet-Loiseau H, Oliva S, Lokhorst HM, Goldschmidt H, Rosinol L, et al. (Jul):32–39. Revised international staging system for multiple myeloma: a report from inter- 26. Chakraborty R, Muchtar E, Kumar SK, Jevremovic D, Buadi FK, Dingli D, et al. Serial national myeloma working group. J Clin Oncol. 2015;33:2863–9. https://doi.org/ measurements of circulating plasma cells before and after induction therapy have 10.1200/JCO.2015.61.2267 an independent prognostic impact in patients with multiple myeloma undergoing 3. Schavgoulidze A, Lauwers-Cances V, Perrot A, Avet-Loiseau4 H, Corre J. The upfront autologous transplantation. Haematologica 2017;102:1439–45. discriminatory ability of the R ÍSS is equivalent to ISS in a large cohort of newly 27. Intzes S, Symeonidou M, Zagoridis K, Bezirgianidou Z, Vrachiolias G, Spanoudaki diagnosed Multiple Myeloma patients. 62 ASH Annual Meeting. Abstract 1338. A, et al. Socioeconomic status is globally a prognostic factor for overall survival of https://ash.confex.com/ash/2020/webprogram/Paper136996.html multiple myeloma patients: synthesis of studies and review of the literature. 4. Mosquera Orgueira A, Díaz Arias JÁ, Cid López M, Peleteiro Raíndo A, Antelo Mediterr J Hematol Infect Dis. 2021;13:e2021006 https://doi.org/10.4084/ Rodríguez B, et al. Improved personalized survival prediction of patients with MJHID.2021.006. PMID: 33489045; PMCID: PMC7813274. diffuse large B-cell Lymphoma using gene expression profiling. BMC Cancer. 28. Mosquera Orgueira A, González Pérez MS, Díaz Arias JÁ, Antelo Rodríguez B, 2020;20:1017. Alonso Vence N, Bendaña López Á, et al. Survival prediction and treatment 5. Biccler JL, Eloranta S, de Nully Brown P, Frederiksen H, Jerkeman M, Jørgensen J, optimization of multiple myeloma patients using machine-learning models et al. Optimizing outcome prediction in diffuse large B-cell lymphoma by use of based on clinical and gene expression data. Leukemia. 2021 May. https://doi.org/ machine learning and nationwide lymphoma registries: a nordic lymphoma 10.1038/s41375-021-01286-2 group study. JCO Clin Cancer Inf. 2018;2:1–13. https://doi.org/10.1200/ 29. Agius R, Brieghel C, Andersen MA, Pearson AT, Ledergerber B, Cozzi-Lepri A, et al. CCI.18.00025 Machine learning can identify newly diagnosed patients with CLL at high risk of 6. Gerstung M, Papaemmanuil E, Martincorena I, Bullinger L, Gaidzik VI, Paschka P, infection. Nat Commun. 2020;11:363 https://doi.org/10.1038/s41467-019-14225-8. et al. Precision oncology for acute myeloid leukemia using a knowledge bank PMID: 31953409; PMCID: PMC6969150. approach. Nat Genet. 2017;49:332–40. https://doi.org/10.1038/ng.3756. 7. Chng WJ, Chung TH, Kumar S, Usmani S, Munshi N, Avet-Loiseau H, et al. Gene signature combinations improve prognostic stratification of multiple myeloma patients. Leukemia 2016;30:1071–8. ACKNOWLEDGEMENTS 8. Kuiper R, van Duin M, van Vliet MH, Broijl A, van der Holt B, El Jarari L, et al. The authors wish to thank all professionals who contributed to the clinical trials Prediction of high- and low-risk multiple myeloma based on gene expression and developed by the Spanish Myeloma Group. The authors thank Roberto Maldonado for the International Staging System. Blood 2015;126:1996–2004. facilitating access to the requested data. Blood Cancer Journal (2022)12:76 A. Mosquera Orgueira et al. 9 AUTHOR CONTRIBUTIONS advisory role for Celgene, Takeda, Amgen, Janssen and Sanofi and travel AMO, MSGP and MVMM performed the research, analyzed the results and wrote the accommodations and expenses from Celgene. J.F.S.M. reports a consultancy or paper. JADA, LR, AIT, JML, LP, MG, MJB, JdlR, ALdlG, RR, AS, MTH, EB, MJC, NG, MLM, advisory role for AbbVie, Amgen, Bristol-Myers Squibb, Celgene, GlaxoSmithKline, JB, JJL and JSM provided data for the study, evaluated the results, made critical Janssen, Karyopharm, MSD, Novartis, Roche, Sanofi, SecuraBio and Takeda. comments to the manuscript and approved its final submission. ADDITIONAL INFORMATION COMPETING INTERESTS Supplementary information The online version contains supplementary material The authors declare that no support was provided in the forms of grants and/or available at https://doi.org/10.1038/s41408-022-00647-z. equipment and grants for the development of this study. M.-V.M. has received honoraria for lectures and participation in advisory boards from Janssen, Celgene- Correspondence and requests for materials should be addressed to Maria Victoria BMS, Amgen, Takeda, Abbvie, GSK, Adaptive, Roche, Seattle Genetics, Pfizer, and Mateos. Regeneron. A.M.O reports honoraria for lectures and participation in advisory boards from Janssen, Takeda, Abbvie, Amgen, Novartis, Gilead and AstraZeneca; research Reprints and permission information is available at http://www.nature.com/ grants from Roche, Pfizer and Celgene-BMS and funds for conference organization reprints from Jassen, Takeda, Abbvie, Amgen, Novartis, Gilead, Roche, Bristol-Myers-Squibb, Glaxo-Smith-Klyne, Incyte and Pfizer. M.S.G.P. has received honoraria for lectures and Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims participation in advisory boards from Janssen, Amgen, Celgene-BMS, Takeda, Sanofi in published maps and institutional affiliations. and GSK. J.A.D.A.: has received honoraria for lectures from Abbvie and Janssen. L.R. reports Honoraria from Janssen, BMS-Celgene, Amgen, TAkeda, Sanofi, GSK and Karyopharm. A.O. reports advisory board fees from Bristol Myers Squibb, Janssen, and Amgen. A.I.T.: no COIs to disclose. L.P.: no COIs to disclose. M.T.H.: no COIs to disclose. E.B.: no COIs to disclose. M.G. has received honoraria from Janssen-Cilag and Celgene. Open Access This article is licensed under a Creative Commons M.J.B. declares honoraria from lectures and advisory boards from Janssen, BMS/ Attribution 4.0 International License, which permits use, sharing, Celgene, Amgen, Takeda, and GSK. J.d.l.R. has served as a consultant and provided adaptation, distribution and reproduction in any medium or format, as long as you give expert testimony within the past 2 years for Amgen, Celgene, Takeda, Janssen, and appropriate credit to the original author(s) and the source, provide a link to the Creative Sanofi. A.L. has received honoraria for advisory boards from Celgene, Amgen, and Commons license, and indicate if changes were made. The images or other third party Janssen. A.S. reports honoraria from Takeda, BMS, MSD, Sanofi, Roche, Novartis y material in this article are included in the article’s Creative Commons license, unless Janssen; consultancy: Takeda, BMS, Novartis, Jansser, Gilead, Sanofi, GSK; Speaker’s indicated otherwise in a credit line to the material. If material is not included in the bureau: Takeda; Research Support: Takeda. M.J.C.: no COIs to disclose. R.R. has article’s Creative Commons license and your intended use is not permitted by statutory received honoraria for lectures and participation in advisory boards from Becton- regulation or exceeds the permitted use, you will need to obtain permission directly Dickinson, Celgene, Janssen, Sanofi and Binding Site. J.M.L. has received honoraria for from the copyright holder. To view a copy of this license, visit http://creativecommons. lectures and participation in advisory boards from Janssen, Celgene-BMS, Amgen, org/licenses/by/4.0/. Takeda, Abbvie, GSK, Adaptive, Roche, Pfizer, and Astellas, Incyte. He has received research grants from BMS, Roche, Astellas and Janssen. J.B. reports honoraria from Janssen, Celgene, Takeda, Amgen and Oncopeptides. J.J.L. reports a consulting or © The Author(s) 2022 THE PETHEMA/GEM COOPERATIVE GROUP Adrian Mosquera Orgueira17, Marta Sonia González Pérez17, Jose Diaz Arias17, Laura Rosiñol18, Albert Oriol19, Ana Isabel Teruel20, Joaquin Martinez Lopez21, Luis Palomera22, Miguel Granell23, Maria Jesus Blanchard24, Javier de la Rubia25, Ana López de la Guia26, Rafael Rios27, Anna Sureda28, Miguel Teodoro Hernandez29, Enrique Bengoechea30, María José Calasanz31, Norma Gutierrez32, Maria Luis Martin21, Joan Blade18, Juan-Jose Lahuerta21, Jesús San Miguel31 and Maria Victoria Mateos32 17 Hospital Clínico Universitario Santiago de Compostela, A Coruña, Spain. 18Hospital Clínic, Institut d’investigacions Biomèdiques August Pi i Sunyer, Barcelona, Spain. 19Institut Català d’Oncologia I Institut Josep Carreras, Hospital Germans Trias i Pujol, Badalona, Spain. 20Hospital Clínico de Valencia, Valencia, Spain. 21Hospital Universitario 12 de Octubre, i+12, Complutense University, CNIO, Madrid, Spain. 22Hospital Clínico Lozano Blesa, Zaragoza, Spain. 23Hospital Sant Pau, Barcelona, Spain. 24Hospital Ramón y Cajal, Madrid, Spain. 25Hospital Doctor Peset, Valencia, Spain. 26Hospital Universitario La Paz, Madrid, Spain. 27Hospital Virgen de las Nieves, CIBERESP, Ibs, Granada, Spain. 28Institut Català d’Oncologia-Hospitalet, IDIBELL, Universitat de Barcelona, Barcelona, Spain. 29Hospital Universitario de Canarias, Santa Cruz de Tenerife, Spain. 30Hospital de Donostia, San Sebastian, Spain. 31Clínica Universidad de Navarra, CIMA, CIBERONC, IDISNA, Pamplona, Spain. 32Hospital Universitario de Salamanca, Instituto de Investigación Biomédica de Salamanca, Instituto de Biología Molecular y Celular del Cáncer (Universidad de Salamanca-Consejo Superior de Investigaciones Científicas), CIBERONC, Salamanca, Spain. Blood Cancer Journal (2022)12:76