Introduction Transplant-associated thrombotic microangiopathy (TA-TMA) represents a critical complication in hematopoietic stem cell transplantation (HSCT) with substantial morbidity and mortality risks. Current diagnostic and prognostic approaches rely predominantly on individual laboratory parameters and clinical criteria and lack comprehensive predictive models that integrate multiple inflammatory and hemolytic biomarkers. The heterogeneous clinical presentation, variable treatment responses, and complex pathophysiology of TA-TMA necessitate more sophisticated AI-driven risk stratification approaches. We investigated whether an unsupervised machine learning model, by incorporating pre- and early post-transplant inflammatory and endothelial injury signatures, could stratify patients into distinct subgroups to predict clinical outcomes, particularly mortality, more accurately than conventional methods.

Methods This nationwide, multicenter, retrospective cohort study enrolled adult patients who underwent allogeneic HSCT and were subsequently diagnosed with TA-TMA across multiple transplant centers from January 2015 to December 2023. Patient data were retrospectively reviewed and classified according to the 2023 international harmonized diagnostic criteria. The primary cohort comprised 847 patients from our center, with external validation performed in 238 patients from independent centers. We developed STRATOS (STRAtification Tool for TA-TMA OutcomeS), an unsupervised quantitative model integrating eight laboratory biomarkers capturing hemolysis, thrombosis, and complement activation: soluble C5b-9 (sC5b-9), schistocyte ratio, LDH, platelet count, hemoglobin, serum creatinine, haptoglobin, and total bilirubin. The model utilized t-distributed stochastic neighbor embedding (t-SNE) for dimensionality reduction and K-means clustering for patient stratification. The primary endpoint for prognostic evaluation was 1-year non-relapse mortality (NRM).

Results Among the 847 patients in the primary cohort (median age, 45 years; 58% male), the median onset of TA-TMA was day +57 post-transplantation. Baseline characteristics at diagnosis revealed a median platelet count of 28×10⁹/L, LDH level of 445 U/L, and schistocyte percentage of 2.1%.

t-SNE dimensionality reduction and K-means clustering identified three distinct phenotypic clusters with significantly different mortality risks. Cluster 1 (Severe Endothelial Injury, n=189), characterized by extremely high sC5b-9 levels, marked elevation in LDH, and severe thrombocytopenia, represented the highest-risk patients with the poorest prognosis. Cluster 0 (Hemolytic-Dominant, n=312), defined by pronounced hemolytic markers but moderate complement activation, represented the intermediate-risk group. Cluster 2 (Compensated, n=346), exhibiting relatively preserved hemoglobin and only mild-to-moderate abnormalities in biomarkers, had the best prognosis.

In the validation cohort, the risk stratification derived from STRATOS demonstrated excellent performance, achieving an area under the curve (AUC) of 0.87 for predicting 1-year NRM. This significantly outperformed assessment based on individual markers (LDH alone: AUC, 0.64; platelet count: AUC, 0.59; p<0.001). Feature importance analysis revealed that sC5b-9 was the most discriminative biomarker for separating the clusters (variability contribution: 0.91), followed by the schistocyte ratio (0.83) and LDH level (0.70).

Conclusions This nationwide, multicenter study establishes and validates the first AI-driven biomarker signature for prognostic risk stratification in patients with established TA-TMA. Compared with conventional approaches, STRATOS provides superior prognostic accuracy for mortality by identifying three distinct phenotypic clusters with disparate outcomes. Future prospective validation studies are warranted to establish guidelines for the real-time clinical integration of this stratification tool.

This content is only available as a PDF.
2025
Sign in via your Institution