Validation of Recent Response Criteria (ELN-22, IWG-23 and PB-CR) in 1634 MDS/CMML/AML Patients Treated with HMA or HMA-Ven Using CPH Models and a CPH Deep Neural Network - Can or Should Response Criteria be Harmonized for Similarly Treated Patients?

Lisa Pleyer, Marc Vaisband, Patrick Klammer, Manuel Drost, Hubert Angermann, Felix Keil, Verena Petzer, Sonja Heibl, Jennifer Moritz, Michael Girschikofsky, Margarete Stampfl‐Mattersberger, Angelika Pichler, Bernd Hartmann, Gregor Aschauer, Clemens A. Schmitt, Sonia Vallet, Szilvia Boros, Petra Pichler, Andrea Hammerl-Steiner, Felix Renneberg, Dilyan Majjiga, Gudrun Russ, Alexander Egle, Michael Leisch, Thomas Melchardt, Nadja Zaborsky, Viktoria Faber, Jan Philipp Bewersdorf, Amer M. Zeidan, Jan Hasenauer, Richard GreilBlood 2024 · DOI

Abstract

Background MDS, CMML and AML demonstrate considerable biologic, phenotypic, genetic and clinical overlap and are often treated similarly. Hypomethylating agents (HMA) combined with venetoclax (VEN) have become the new standard of care for many patients (pts) with AML, with trials in MDS underway, and new response criteria have been published for AML (ELN22; Döhner H, Blood 2022), higher-risk MDS (IWG23; Zeidan A, Blood 2023) and pts treated with azacitidine (PB-CR; Pleyer L, AJH 2023). Methods Composite complete remission (cCR) rates were defined as published: ELN17 (CR/CRi/p), ELN22 (CR/CRi/p/h), IWG06 (CR/mCR), IWG23 (CR/CRbi/uni/h). Overall survival (OS) and time to next treatment (TTNT) were analyzed using the Kaplan-Meier method. Cox proportional hazards (CPH) models were applied to obtain adjusted median OS and hazard ratios for each pairwise comparison. Permutation-based feature importances for the CPH neural network DeepSurv were used to assess the predictive value of response types as described previously (Pleyer L, AJH 2023). SAS®9.4. and Python 3.8.12. were used. Results 1634 patients (pts) with MDS, CMML or AML receiving 15099 cycles of HMA (n=1416) and 2292 cycles of HMA-VEN (n=218) within the Austrian Myeloid Registry (NCT04438889) were included. 13 covariates remaining in the final CPH model were used for multivariable adjustments. For HMA treated pts, adjusted median OS (hereafter “OS”) of pts reaching PB-CR (n=258; 21%) vs those who did not (n=966) was 18.4 vs 9.9 mo (p<.0001, HR .462 [95% CI .395-.541]). OS of pts reaching ELN17-cCR (n=190, 16%) vs those who did not (n=1034) was 17.6 vs 10.5 mo (p<.0001, HR .528 [.446-.626]). OS of pts reaching ELN22-cCR (n=202, 17%) vs those who did not (n=1022) was 17.5 vs 10.4 mo (p<.0001, HR .534 [.453-.630]). OS of pts reaching IWG06-cCR (n=195, 16%) vs those who did not (n=1029) was 16.9 vs 10.5 mo (p<.0001, HR .561 [.476-.662]). OS of pts reaching IWG23-cCR (n=183, 15%) vs those who did not (n=1041) was 18.0 vs 10.4 mo (p<.0001, HR .499 [.421-.593]). OS of pts reaching PB-CR vs IWG06-cCR (21.9 vs 19.8 mo, p=.14), vs IWG23-cCR (23.4 vs 23.5 mo, p=.94), vs ELN17-cCR (n=190) (22.8 vs 22.0 mo, p=.71) or vs ELN22-cCR (22.8 vs 21.8 mo, p=.63) was similar. For HMA-VEN treated pts, OS of pts reaching PB-CR (n=42; 23%) vs those who did not (n=139) was 25.0 vs 9.0 mo (p<.0001, HR .253 [.145-.442]). OS of pts reaching ELN17-cCR (n=59, 32%) vs those who did not (n=123) was 34.1 vs 9.0 mo (p<.0001, HR .163 [.093-.288]). OS of pts reaching ELN22 cCR (n=60, 33%) vs those who did not (n=122) was 34.1 vs 8.9 mo (p<.0001, HR .175 [.101-.303]). OS of pts reaching IWG06-cCR (n=83, 46%) vs those who did not (n=99) was 23.8 vs 8.5 mo (p<.0001, HR .261 [.166-.410]). OS of pts reaching IWG23 cCR (n=52, 29%) vs those who did not (n=130) was 34.1 vs 9.6 mo (p<.0001, HR .161 [.088-.293]). OS of pts reaching PB-CR vs IWG06-cCR (24.8 vs 20.6 mo, p=.4422), vs IWG23-cCR (24.8 vs 25.0 mo, p=.83), vs ELN17-cCR (25.0 vs 34.1 mo, p=.69) or vs ELN22-cCR (24.8 vs 25.0 mo, p=.80) was similar. In pts receiving 1st line HMA monotherapy, PB-CR resulted in slightly better performance than ELN22-cCR (C-index .67 vs .64, p<.0001 for OS; .68 vs .63 for TTNT, p<.0001) or IWG23 cCR (C-index .67 vs .65, p<.0001 for OS; .68 vs .65 for TTNT, p<.0001), indicating that PB-CR has more information content than the other response types. In pts receiving 1st line HMA-VEN, PB-CR had similar performance to ELN22-cCR (C-index .63 vs .64, p=1.00 for OS; .60 vs .59, p<.0001 for TTNT), and IWG23 (C-index .63 vs .64, p=.0057 for OS; .60 vs .59, p<.0001 for TTNT). In pts receiving 1st line HMA with drugs other than VEN (HMA-others), PB-CR had slightly better performance than ELN22-cCR (C-index .63 vs .60, p<.0001 for OS; .57 vs .52, p<.0001 for TTNT) or IWG23-cCR (C-index .63 vs .63, p=1.00 for OS; .57 vs .55, p<.0001 for TTNT). Summary In this prospective cohort study of 1634 pts treated with HMA, HMA-VEN or HMA-others, our data validate ELN22, IWG23 and PB-CR response criteria to effectively predict outcomes, irrespective of disease subtype. Deep neural network analysis-based findings indicate that PB-CR has similar predictive capacity to IWG23-cCR and ELN22-cCR. Harmonization of response criteria among disease entities treated similarly may be possible. As bone marrow evaluations are only performed in ~50% of pts outside of clinical trials (Mukherjee S, Oncologist 2023; Dinmohamed AG, Leuk res 2015), the current validation of PB-CR is of clinical relevance.
BibTeX
@inproceedings{pleyer2024validation,
  author = {Lisa Pleyer and Marc Vaisband and Patrick Klammer and Manuel Drost and Hubert Angermann and Felix Keil and Verena Petzer and Sonja Heibl and Jennifer Moritz and Michael Girschikofsky and Margarete Stampfl‐Mattersberger and Angelika Pichler and Bernd Hartmann and Gregor Aschauer and Clemens A. Schmitt and Sonia Vallet and Szilvia Boros and Petra Pichler and Andrea Hammerl-Steiner and Felix Renneberg and Dilyan Majjiga and Gudrun Russ and Alexander Egle and Michael Leisch and Thomas Melchardt and Nadja Zaborsky and Viktoria Faber and Jan Philipp Bewersdorf and Amer M. Zeidan and Jan Hasenauer and Richard Greil},
  title = {{Validation of Recent Response Criteria (ELN-22, IWG-23 and PB-CR) in 1634 MDS/CMML/AML Patients Treated with HMA or HMA-Ven Using CPH Models and a CPH Deep Neural Network - Can or Should Response Criteria be Harmonized for Similarly Treated Patients?}},
  booktitle = {Blood},
  year = {2024},
  doi = {10.1182/blood-2024-208073},
}

Loading related content...

Figure