Acute myeloid leukemia (AML) is a disease resulting from different molecular aberrations which give rise to heterogeneous biological characteristics and patient outcomes.¹ Most patients are older than 60 years at diagnosis and often present with clinically relevant comorbidities. Therefore, given the associated risk of therapy-related morbidity and mortality, a profound and individualized risk assessment is essential to decide on the optimal treatment approach, including the choice to perform an allogeneic stem cell transplantation (HSCT) in first complete remission (CR). Today, the risk stratification is largely based on the presence of certain genetic abnormalities, such as those assessed by the European LeukemiaNet (ELN) 2017.¹

In a more recent approach, Gerstung et al.² developed a knowledge-bank (KB)-based algorithm, predicting individual patients' outcomes based on genetic, clinical, and demographic features, also anticipating the effects of a planned allogeneic HSCT.

As of today, two studies evaluated the KB-based algorithm in the routine clinical practice by comparing outcome predictions to that by the ELN2017 risk stratification. The first included 155 AML patients and showed the superiority of the KB approach for a 3-year overall survival (OS) prediction, including discrimination between non-remission and relapse death.³ Another study with 1612 de novo AML patients showed the KB algorithm to perform well in predicting 3-year OS and non-remission death, but inferior in predicting relapse death and death in first complete remission.⁴ Remarkably, the ability of the KB approach to predict outcomes in individuals that received an allogeneic HSCT has not been validated yet. Only 70 patients in first CR (n = 54) or after relapse (n = 16) in the first study and none in the second study had an allogeneic HSCT.^{3, 4} However, in order to inform treatment decision toward or against an HSCT based on this approach, it is important to validate the algorithm in AML patients treated with HSCT. Here, we analyzed a cohort of 546 newly diagnosed AML patients (median age at diagnosis 62.2; range, 20.7–76.5 years) who received an allogeneic HSCT after reduced intensity or non-myeloablative conditioning at our institution. Most patients were transplanted in first CR/CR with incomplete recovery (CRi) (n = 327, 60%), 17% were transplanted in second CR/CRi (n = 91) and 23% without reaching CR/CRi (n = 126). Additional details on the applied therapies, and further patients' characteristics are given in the Supporting Information Methods S1 and Table S1. Patients were retrospectively screened for clinical and genetic parameters included in the KB. The mutation status of 54 genes included in the TruSight Myeloid Sequencing Panel (Illumina) at diagnosis was performed in patients with adequate pre-treatment material available. Subsequently, the mutation status of 17 genes included in the KB algorithm was not available in our cohort (see Supporting Information Methods S1). The KB estimates were calculated for each individual with the actually performed transplant strategy and compared to the observed outcomes using receiver operating characteristics (ROC) curves. Additionally, the measurable residual disease (MRD) status was evaluated prior to allogeneic HSCT as previously published⁵ and described in the Supporting Information Methods S1. In our cohort, the KB approach had an area under the curve (AUC) to predict OS at 1 and 3 year of AUC_KB = 0.68 [95% CI 0.61–0.72] and AUC_KB = 0.69 [95% CI 0.62–0.72], respectively. Both OS predictions were not significantly different to those predicted by the ELN2017 risk stratification (1-year OS p = .33, Figure S1 and 3-year OS p = .23, Figure 1A). Although sub-analyses were restricted by lower patient numbers, comparable results were obtained in separate analyses of patients transplanted in first CR/CRi (see Figure S2) as well as the distinct conditioning intensities (non-myeloablative and reduced-intensity, Figure S3 and S4, respectively). Importantly, in the study by Bill et al.,⁴ in which no patient received an allogeneic HSCT, the AUC_KB was 0.80. A second large study that used a combination of the KB, ELN2017, and MRD to inform the decision for an allogeneic HSCT in first remission in which approximately half of the patients underwent allogeneic HSCT also resulted in a lower AUC_KB of 0.69.⁶ Altogether, the data indicate a reduced capability of the KB approach to estimate individual OS in patients undergoing allogeneic HSCT—most likely due to confounders introduced through donor selection, comorbidities, and conditioning regimens, which are not reflected in the KB algorithm today.

Next, we performed a multivariate analysis including variables not reflected in the KB approach (Supporting Information Methods S1). Here, the KB prediction for 3-year OS as a continuous parameter remained an independent prognostic factor for OS after adjustment for the MRD-corrected remission status prior to allogeneic HSCT (Table S2A). When comparing the Akaike information criterion (AIC) to a second model containing the ELN2017 genetic risk together with the remission status (Table S2B), the one containing the KB showed a lower AIC and was identified as the statistically preferred model. To further visualize the prognostic value of the KB approach, we then introduced arbitrary cut-offs according to the KB value for OS similar to Fenwarth et al.,⁶ and observed a clear separation of OS curves according to a KB value of <20, 20–39, and ≥40 with higher values indicating a higher likelihood for OS (p < .001, Figure 1B). The ability to identify distinct risk groups was also seen in separate analyses for patients transplanted in a first CR/CRi (p < .001, Figure S2C). However, the “three-groups” KB model did not perform superior to the prognostication by the ELN2017 genetic risk (AUC comparison, p = .54).

Furthermore, we analyzed additional endpoints for which the KB algorithm provided outcome prediction. Similar to previously published data,^{3, 4} the KB approach had the highest probability to predict death without previous achievement of a CR within 3 years after diagnosis (AUC_KB = 0.75, 95% CI 0.57–0.92, Figure 1C, black curve). In contrast, and again congruent with published results in patients receiving chemotherapy,^{3, 4} outcome prediction for death in first CR (AUC_KB = 0.61, 95% CI 0.55–0.68, Figure 1C, red curve) and death after relapse (AUC_KB = 0.63, 95% CI 0.58–0.69, Figure 1C, blue curve) were limited. Finally, we observed good outcome prediction for “being alive after relapse” (AUC_KB = 0.77, 95% CI 0.71–0.84, Figure 1D, blue curve) as well as “being alive in first CR” (AUC_KB = 0.69, 95% CI 0.64–0.74, Figure 1D, red curve). Our data also imply that a “salvage” allogeneic HSCT in our cohort did not improve outcomes for most patients with primary induction failure (“death without CR”), which is reflected by a relatively good KB prediction for this endpoint. The potential of the KB approach to predict survival with or without relapse was better than the ability to predict death, which most likely is caused by difficulties to estimate treatment-related complications after allogeneic HSCT, including infections and graft-versus-host disease. However, all ROC curves derived from our analyses seemed to show inferior outcome predictions compared to those of previous studies of patients receiving chemotherapy consolidation.^{3, 4}

In conclusion, our study is the first to validate the KB algorithm in AML patients undergoing allogeneic HSCT. While the KB approach allows outcome prediction in these patients, the predictive power seems to be restricted compared to patients receiving chemotherapy consolidation, as donor selection, or conditioning regimens are not integrated into the score and introduce significant confounders, especially for treatment-related mortality. Subsequently, future versions of the KB that also consider HSCT-associated factors might further improve outcome prediction of this promising and important clinical tool.

The study was conducted according to the guidelines of the Declaration of Helsinki. Data analyses were approved by the Institutional Review Board of the University Hospital Leipzig. Informed consent was obtained from all subjects involved in the study.

急性髓性白血病 (AML) 是一种由不同分子畸变引起的疾病，这些畸变会导致异质的生物学特征和患者预后。¹大多数患者在诊断时年龄超过 60 岁，并且经常出现与临床相关的合并症。因此，考虑到与治疗相关的发病率和死亡率的相关风险，深入和个体化的风险评估对于决定最佳治疗方法至关重要，包括选择在首次完全缓解 (CR) 时进行异基因干细胞移植 (HSCT) . 今天，风险分层主要基于某些遗传异常的存在，例如欧洲白血病网络 (ELN) 2017 评估的那些^。1

在最近的方法中，Gerstung 等人。²开发了一种基于知识库 (KB) 的算法，根据遗传、临床和人口统计学特征预测个体患者的结果，同时预测计划中的异基因 HSCT 的效果。

截至今天，两项研究通过将结果预测与 ELN2017 风险分层的结果预测进行比较，评估了常规临床实践中基于 KB 的算法。第一个包括 155 名 AML 患者，并显示了 KB 方法在 3 年总生存期 (OS) 预测方面的优越性，包括区分非缓解和复发死亡。³另一项针对 1612 名新发 AML 患者的研究表明，KB 算法在预测 3 年 OS 和非缓解死亡方面表现良好，但在预测复发死亡和首次完全缓解时的死亡方面表现不佳。⁴值得注意的是，KB 方法预测接受异基因 HSCT 个体结果的能力尚未得到验证。首次 CR 仅 70 例患者（n = 54) 或在第一项研究中复发 ( n  = 16) 后，第二项研究中没有人进行同种异体 HSCT。^{3, 4}然而，为了根据这种方法做出支持或反对 HSCT 的治疗决策，在接受 HSCT 治疗的 AML 患者中验证该算法非常重要。在这里，我们分析了 546 名新诊断的 AML 患者（诊断时的中位年龄 62.2 岁；范围，20.7-76.5 岁），他们在我们机构降低强度或非清髓性预处理后接受了异基因 HSCT。大多数患者在第一次 CR/CR 中移植且不完全恢复 (CRi) ( n  = 327, 60%)，17% 在第二次 CR/CRi ( n  = 91) 中移植，23% 未达到 CR/CRi ( n = 126)。支持信息方法 S1 和表 S1 中提供了有关所应用疗法的更多详细信息以及进一步的患者特征。对患者进行回顾性筛查，了解 KB 中包含的临床和遗传参数。诊断时包含在 TruSight 骨髓测序面板 (Illumina) 中的 54 个基因的突变状态是在具有足够可用预处理材料的患者中进行的。随后，KB 算法中包含的 17 个基因的突变状态在我们的队列中不可用（参见支持信息方法 S1）。使用实际执行的移植策略计算每个个体的 KB 估计值，并使用接受者操作特征 (ROC) 曲线与观察到的结果进行比较。此外，⁵并在支持信息方法 S1 中进行了描述。在我们的队列中，KB 方法具有曲线下面积 (AUC) 来预测 AUC _KB  = 0.68 [95% CI 0.61–0.72] 和 AUC _KB  = 0.69 [95， 分别。两种 OS 预测与 ELN2017 风险分层预测的结果没有显着差异（1 年 OS p  = .33，图 S1 和 3 年 OS p = .23，图 1A)。尽管子分析受到患者数量较少的限制，但在对第一次 CR/CRi 移植患者的单独分析（见图 S2）以及不同的调节强度（非清髓性和降低强度，图 S3 和S4，分别）。重要的是，在 Bill 等人⁴的研究中，没有患者接受同种异体 HSCT，AUC _KB为 0.80。第二项大型研究使用 KB、ELN2017 和 MRD 的组合来告知第一次缓解时异基因 HSCT 的决定，其中大约一半的患者接受异基因 HSCT，也导致 AUC _KB较低，为 0.69。⁶总而言之，这些数据表明 KB 方法在估计接受异基因 HSCT 的患者中个体 OS 的能力降低——很可能是由于通过供体选择、合并症和预处理方案引入的混杂因素，这些因素在今天的 KB 算法中没有反映出来。

接下来，我们进行了多变量分析，包括 KB 方法（支持信息方法 S1）中未反映的变量。在这里，作为连续参数的 3 年 OS 的 KB 预测在对同种异体 HSCT 之前的 MRD 校正缓解状态进行调整后仍然是 OS 的独立预后因素（表 S2A）。当将 Akaike 信息标准 (AIC) 与包含 ELN2017 遗传风险和缓解状态的第二个模型（表 S2B）进行比较时，包含 KB 的模型显示出较低的 AIC，并被确定为统计学上的首选模型。为了进一步可视化 KB 方法的预后价值，我们随后根据 OS 的 KB 值引入了任意截止值，类似于 Fenwarth 等人，⁶并根据 KB 值 <20、20–39 和 ≥40 观察到 OS 曲线的明显分离，更高的值表明 OS 的可能性更高（p  < .001，图 1B）。在对第一次 CR/CRi 移植患者的单独分析中也发现了识别不同风险组的能力（p  < .001，图 S2C）。然而，“三组”KB 模型的表现并不优于 ELN2017 遗传风险的预测（AUC 比较，p  = .54）。

此外，我们分析了 KB 算法为其提供结果预测的其他端点。与之前公布的数据类似，^{3, 4} KB 方法在诊断后 3 年内未达到 CR 的情况下预测死亡的概率最高（AUC _KB  = 0.75, 95% CI 0.57–0.92，图 1C，黑色曲线）。相比之下，再次与接受化疗的患者的已发表结果一致，^{第 3、4 个}结果预测第一次 CR 中的死亡（AUC _KB  = 0.61, 95% CI 0.55-0.68，图 1C，红色曲线）和复发后的死亡（AUC _KB  = 0.63，95% CI 0.58–0.69，图 1C，蓝色曲线）是有限的。最后，我们观察到“复发后还活着”的良好结果预测（AUC _KB = 0.77，95% CI 0.71–0.84，图 1D，蓝色曲线）以及“在第一个 CR 中还活着”（AUC _KB = 0.69，95% CI 0.64–0.74，图 1D，红色曲线）。我们的数据还暗示，我们队列中的“挽救”同种异体 HSCT 并未改善大多数初次诱导失败（“没有 CR 的死亡”）患者的预后，这反映在该终点相对较好的 KB 预测中。KB 方法预测复发或不复发生存的潜力优于预测死亡的能力，这很可能是由于难以估计同种异体 HSCT 后的治疗相关并发症，包括感染和移植物抗宿主病。然而，与之前接受化疗巩固的患者的研究相比，我们分析得出的所有 ROC 曲线似乎都显示出较差的结果预测。^3、4

总之，我们的研究是第一个在接受异基因 HSCT 的 AML 患者中验证 KB 算法的研究。虽然 KB 方法允许对这些患者的结果进行预测，但与接受化疗巩固的患者相比，预测能力似乎受到限制，因为供体选择或预处理方案未纳入评分并引入显着的混杂因素，尤其是与治疗相关的死亡率。随后，还考虑 HSCT 相关因素的 KB 的未来版本可能会进一步改善对这一有前途且重要的临床工具的结果预测。

该研究是根据赫尔辛基宣言的指导方针进行的。数据分析得到莱比锡大学医院机构审查委员会的批准。从参与研究的所有受试者获得知情同意。