研究背景：

造血干细胞移植（HSCT）后的免疫重建过程是复杂的，对HSCT后生存至关重要。既往的研究已经表明：HSCT后的免疫重建通常按照先天免疫、CD8⁺ T细胞、CD4⁺ T和B细胞的顺序完成；免疫重建越早恢复正常，感染风险越低，整体生存率越高；不同的干细胞来源具有不同的免疫重建特点，目前无证据提示某种干细胞来源为最佳。尽管HSCT后免疫重建得到研究人员的持续关注，但以往的研究仍存在一些局限性：纳入的患者较少，通常少于200例；仅聚焦单变量免疫指标的预后意义；移植后免疫状态监测的时间较短。因此，如何整合多变量免疫数据来可视化免疫重建的时间序列轨迹、建立可在临床广泛应用的免疫重建评价体系尚无共识提供规范及参考。

研究目的：

本研究旨在整合多变量免疫数据，实现HSCT后免疫重建轨迹可视化、动态追踪免疫重建状态、探索可预测移植后早期死亡的“高风险”免疫特征。此外，基于本研究结果开发在线工具，以供临床医生解读患者免疫状态及其预后意义。

研究方法：

研究纳入了中国两个医疗中心2012至2020年期间进行HSCT的1945例患者，其中1838例（94.5％）患者为异基因HSCT（allo-HSCT）。对共计11150份移植后免疫资料进行回顾性分析。利用流形学习方法整合了外周血中20项免疫变量，实现免疫细胞重建时间序列轨迹的可视化。应用网格搜索优化方法来发现与死亡相关的复合免疫特征，然后将复合免疫特征简化为复合免疫风险评分公式，并在两个独立的患者集中验证复合免疫风险评分。

研究结果：

平均而言，在移植后的前两个月，绝对细胞计数的恢复集中在CD8⁺ T细胞的重建上，NK、CD8⁺ T、CD4⁺ T和B细胞的广泛恢复随之展开。患者间免疫状态的变异在移植后60天左右达到最大。基于训练集的研究对象（n = 729），我们明确了一个在移植后91-180天期间可预测早期死亡的复合免疫特征，并将其简化为复合免疫风险评分公式，根据评分公式预测患者移植后死亡的风险。进一步在验证集（n = 284）和测试集（n = 391）中对复合免疫风险评分进行验证，结果证实评分为高风险的患者移植后早期死亡风险显著增加，在验证集和测试集中高风险患者的死亡风险比分别为3.64（95% CI 1.55-8.51，P = 0.0014）和2.44（95% CI 1.22-4.87，P = 0.0087）。进一步分析发现，高风险患者发生非复发死亡（HR 3.28，95% CI 2.28-4.73，P < 0.0001）、感染相关死亡（HR 3.12，95% CI 1.97-4.96，P < 0.0001）的风险更高。重要的是，多因素分析的结果表明，在纳入影响移植后生存的相关变量后，移植后91-180天期间复合免疫风险评分高危仍然是非常重要的早期死亡的独立危险因素（HR 1.80，95% CI 1.28-2.55，P = 0.00085）。严重的急性移植物抗宿主病、感染和患者年龄都是移植后91-180天内出现高危复合免疫风险评分的危险因素。此外，我们开发了一个SKIRT计算器，可供临床医生评估患者移植后的免疫重建进程及预后意义。

研究结论：

我们突破了既往单变量研究的局限性，结合无监督机器学习和生存模型实现了移植后免疫重建动态轨迹的可视化，并建立了一个移植后91-180天期间的多变量复合免疫风险评分用于预测allo-HSCT后早期死亡。经在独立的患者队列中验证，该复合免疫风险评分仍可显著区分移植后死亡风险。本研究提出的复合免疫风险评分易于计算，可以识别出需要针对性地进行感染预防和控制的死亡高风险患者。未来的研究可进一步制定临床干预策略以提高复合免疫风险评分高危患者的总体生存率。

Background:

Immune reconstitution after hematopoietic stem cell transplantation (HSCT) is complicated and of vital importance for survival. Previously researchers have established that: First, immune reconstitution after HSCT often follows the sequential order of innate immunity, CD8⁺ T cells, and then CD4⁺ T and B cells. Second, the earlier the immune cell counts are normalized, the lower the risk for infection, and the better the overall survival. Third, while different stem cell sources have different profiles of immune reconstitution after transplant, so far there is no cell source that is decidedly superior. The previous studies, however, often contained few patients (<200), focus on the prognostic significance of single immune variables, or had short durations of immune monitoring. Therefore, there is no consensus on how to integrate multivariate immune data to visualize the time course of immune reconstitution or establish a scoring system that can be applied broadly at the clinic.

Purposes:

We aimed to integrate multivariate immune data, visualize the trajectory of post-HSCT immune reconstitution, study the kinetics of the multivariate immune status, and discover a ‘high-risk’ composite immune signature that is predictive of early mortality. Additionally, we aim to develop an online tool to facilitate the clinicians to compute an immune signature for prognosis.

Methods:

We retrospectively analyzed 11150 post-transplant immune profiles of 1945 patients who underwent HSCT between 2012 and 2020 at two medical centers in China. 1838 (94.5%) of the cases were allogeneic HSCT. We utilized manifold learning to visualize the temporal profiles of immune cell repopulation (integrating 20 immune cells features in the peripheral blood), used grid-search optimization to identify a composite immune signature that was associated with mortality. Then simplified the composite immune signature into a formula for a Composite Immune Risk Score and discovered and verified it in two independent subsets of the patients.

Results:

On average, recovery of absolute cell counts was concentrated on the repopulation of CD8⁺ T cells during the initial two months after transplant. Inter-patient variance of immune status peaked at around day 60 post-transplant, and broad-ranged recovery in NK, CD8⁺ T, CD4⁺ T, and B cells initiated afterwards. Using the training set of patients (n = 729), we identified a composite immune signature during days 91 – 180 after allogeneic HSCT that was predictive of early mortality. Moreover, we simplified it into a formula for a Composite Immune Risk Score which could classify patients as high-risk and low-risk. When we verified the Composite Immune Risk Score in the validation (n = 284) and test (n = 391) sets of patients, a high score value was found to be associated with hazard ratios of 3.64 (95% CI 1.55 – 8.51; P = 0.0014) and 2.44 (95% CI, 1.22 – 4.87; P = 0.0087), respectively, for early mortality. The high-risk patients were significantly more likely to suffer from non-relapse mortality (hazard ratio (HR), 3.28; 95% CI, 2.28 – 4.73; P < 0.0001), including infection-related mortality (HR, 3.12; 95% CI, 1.97 – 4.96; P < 0.0001). Importantly, a high Composite Immune Risk Score during days 91 – 180 remained an independent risk factor for early mortality after allogeneic HSCT (HR, 1.80; 95% CI, 1.28 – 2.55; P = 0.00085) in multivariate analysis. Severe aGVHD, infectious episodes and patient age were all risk factors for having Composite Immune Risk Scores during days 91 – 180 after allogeneic HSCT. In addition, the SKIRT Calculator that enables clinicians to evaluate the patient’s progress in immune reconstitution after transplantation was developed.

Conclusions:

We broke through the limitations of previous univariate studies, combined unsupervised machine learning and survival modeling to visualize the dynamic patterns of post-transplant immune reconstitution and formulate a multivariate Composite Immune Risk Score during days 91 – 180 post-transplant that can predict early mortality after allo-HSCT even in independent cohort. The Composite Immune Risk Score is easy to compute and could identify the high-risk patients of allogeneic HSCT who require targeted effort for prevention and control of infection. Future researches should focus on devising clinical strategies to improve the overall survival of the high-score patients.