In a recent study published in HemaSphere, Tao and colleagues introduced a machine learning-based prognostic model that identifies distinct risk groups in infant acute myeloid leukemia (AML) with high accuracy. This approach might offer a critical translational advance by enabling personalized treatment intensification or de-escalation, ultimately improving survival outcomes while reducing unnecessary toxicity in this vulnerable pediatric population 1. The most challenging from a clinical perspective is the fact that approximately 82% of infant AML patients are now classified into intermediate-risk groups using current stratification systems, making it exceedingly difficult to determine optimal consolidation therapy strategies. This therapeutic uncertainty underscores the urgent need for improved risk stratification tools specifically tailored to infant AML. To address this critical gap, Tao and colleagues leveraged one of the largest cohorts of infant AML patients with integrated clinical and transcriptomic data. Workflow and main findings of this newly proposed prognostic model are presented in Figure 1. The study analyzed 340 infant AML patients from four Children's Oncology Group (COG) clinical trials within the TARGET project 2, and an external validation cohort of 63 patients from two Chinese institutions. This multi-cohort design, spanning different treatment protocols and ethnic backgrounds, provided robust validation of the model's generalizability. Notably, the definition of infant AML in this study (age ≤ 12 months) followed the original COG dataset criteria to maximize sample size and statistical power. However, this definition may introduce selection bias, underscoring the need for multi‑center collaboration to expand infant AML cohorts for future model refinement. The authors performed comprehensive transcriptomic analysis comparing infant AML samples to healthy pediatric controls and older pediatric AML cases, identifying 8526 differentially expressed genes enriched in immune, epithelial–mesenchymal transition, and complement pathways. While this is an effective feature selection approach, the use of general healthy pediatric controls (rather than age‑matched healthy infants) is a limitation, as developmentally regulated genes in infants cannot be filtered out, potentially introducing confounding factors. To address this, the team systematically evaluated 40 ensemble‑based machine learning combinations—most of which are based on ensemble learning—across multiple validation datasets, ultimately selecting an RSF + GBM model that yielded a 24‑gene signature with strong discriminatory power. Notably, the lower performance in the external cohort compared to the discovery cohort warrants mechanistic discussion. Potential contributors include genetic background differences (e.g., population‑specific variants, mutation spectra), cohort‑specific treatment protocols or sample processing, transcriptome technology heterogeneity (e.g., batch effects, platform differences), and variation in the prevalence of known prognostic factors such as KMT2A rearrangements or NPM1 mutations. Although KMT2A rearrangement was not identified as an independent prognostic factor in univariate Cox analysis in this study, the inclusion of multiple HOXA cluster genes (HOXA3, HOXA10, HOXA-AS2) and GATA6 in the 24‑gene IPSscore remains notable. HOXA genes are well‑established downstream targets of KMT2A fusion proteins, and their aberrant expression is a hallmark of KMT2A‑rearranged infant AML. However, the fact that KMT2A status itself did not retain independent prognostic value suggests that the transcriptional consequences captured by these HOXA‑related genes—along with novel contributors in the model such as RBMS3 and immune‑related genes (AIF1L, MRC2)—may integrate and even outperform the raw genomic lesion in predicting clinical outcomes. In other words, the IPSscore may partially reflect KMT2A‑driven biology but extends beyond it by incorporating microenvironmental and stromal remodeling signals. Future studies should investigate whether the signature retains prognostic value in KMT2A‑wildtype infant AML and whether these genes directly influence chemoresistance or the leukemia microenvironment. The clinical utility of the IPSscore extends beyond mere prognostic prediction. In multivariable analysis, the IPS group emerged as an independent predictor of event-free survival in both the discovery and internal validation sets, maintaining significance even after adjustment for established clinical and molecular risk factors. Critically, the model demonstrated superior performance compared to existing AML stratification tools, including the COG study-defined risk groups, the LSC17 score (originally developed for adult AML) 3, and the APS score (applicable to both adult and pediatric AML) 4. Time-dependent receiver operating characteristic analysis and decision curve analysis consistently showed higher area under the curve values and greater net clinical benefit for the IPSscore across all timepoints and datasets. This superior performance likely stems from the model's development using infant-specific differentially expressed genes, better capturing the unique transcriptomic landscape of this age group compared to models derived from adult or general pediatric AML populations. Analysis of post‑HSCT relapse incidence in first complete remission revealed distinct patterns by IPS group. Among low‑risk patients, HSCT was associated with a higher relapse risk (HR 1.89, p = 0.041), though this finding requires cautious interpretation due to limited sample size and lack of independent validation. High‑risk patients showed a non‑significant trend toward HSCT benefit. These preliminary results suggest that IPS group‑based stratification—if validated in larger cohorts—could help refine HSCT patient selection, potentially reducing unnecessary toxicity in low‑risk patients while ensuring effective consolidation for high‑risk patients. The most clinically significant contribution of this work likely lies in its capacity for risk group refinement, a critical need given the high proportion of infant AML patients currently classified into the intermediate-risk category. Specifically, when the authors integrated the IPSscore with the CFM risk stratification model used in contemporary clinical trials 5, they successfully reclassified 43% of infant AML patients into more precise risk categories. This substantial reclassification rate not only underscores the unique prognostic value of the 24-gene signature but also robustly demonstrates the transformative potential of incorporating transcriptomic data into established clinical decision-making frameworks. Despite strengths (large sample size, multi‑cohort validation, systematic machine learning approach), limitations include retrospective design, treatment heterogeneity, short follow‑up (~3.4 years), non‑infant healthy controls, and a small external cohort (n = 63). Prospective validation in uniform cohorts is needed prior to clinical use. Looking forward, this work opens several important avenues for future research and clinical application. Prospective validation of the IPSscore in ongoing or future COG trials would provide the definitive evidence needed for clinical implementation. The 24 genes comprising the signature warrant functional investigation to understand their biological roles in infant AML pathogenesis and potentially identify novel therapeutic targets. Integration of the IPSscore with emerging technologies, such as measurable residual disease monitoring could further refine risk stratification. The model's superior performance in infant AML suggests that age-specific molecular signatures may be valuable across other pediatric cancer subtypes, encouraging similar approaches in other disease contexts. Finally, the demonstration that low-risk patients may be harmed by HSCT highlights the urgent need for alternative consolidation strategies, possibly including targeted therapies or immunotherapies tailored to the unique biology of infant AML. In conclusion, Tao and colleagues have developed a clinically actionable prognostic model that addresses a critical gap in infant AML management. By demonstrating that transcriptomic profiling can improve risk stratification beyond current cytogenetic and molecular markers, this work represents an important step toward precision medicine in this challenging disease. The model's ability to guide HSCT decisions and reclassify a substantial proportion of patients into more appropriate risk groups holds promise for reducing treatment-related toxicity while improving outcomes. As we move toward an era of increasingly personalized cancer therapy, this study exemplifies how advanced computational approaches can translate complex molecular data into clinically meaningful tools that directly impact patient care. K.H.Y., X.C. and C.W. designed the research, wrote and revised the manuscript. All authors have read and approved the final manuscript. The authors has nothing to report. The authors have nothing to report. The authors have nothing to report. The authors declare no conflicts of interest. Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.
Cao et al. (Sun,) studied this question.