Abstract
Background Due to the high mortality of immune thrombotic thrombocytopenic purpura (iTTP), rapid clinical evaluation and early intervention are crucial. The PLASMIC score remains a valuable and widely used tool for predicting severe ADAMTS13 deficiency. However, recent studies have shown variable performance across different populations. In addition, its variables may often be missing during initial presentation. Refining the accuracy of early iTTP evaluation could improve timely recognition and treatment, while reducing unnecessary interventions in those eventually ruled out for iTTP. Machine learning advancements offer several advantages for enhancing clinical prediction models: the ability to assess continuous variables, capture complex interactions between them, and handle missing data.
We previously utilized XGBoost (extreme gradient boosting) to develop a novel machine learning tool, TTP-14, which improved upon the PLASMIC score's predictive performance in repeat 10-fold cross-validation achieving an AUC of 0.91 (Blood. 2024;144[Suppl 1]:2617). TTP-14 generates a predicted probability for iTTP, expressed as a continuous percentage estimate of iTTP diagnostic likelihood, and is also able to do so with missing variables. However, it has not previously been externally validated.
Objectives To externally validate the ability of TTP-14 to predict iTTP diagnosis and improve upon the PLASMIC score.
Methods We performed an external validation of TTP-14 using an independent cohort of patients with suspected iTTP at the Mayo Clinic. These patients were not part of the original multi-center development cohort of 106 iTTP patients from the University of Utah, Case Western Reserve University, Rochester Regional Health, and University of Illinois.
For the validation cohort, we identified patients who underwent ADAMTS13 activity testing between 2010 and 2025 for evaluation of thrombotic microangiopathy at the Mayo Clinic. iTTP was defined as ADAMTS13 activity <10%, and controls as ADAMTS13 activity >20%. TTP-14 variables were collected at initial presentation: platelet count, lactate dehydrogenase, creatinine, international normalized ratio, reticulocyte percentage, active cancer, indirect bilirubin, mean corpuscular volume, total bilirubin, haptoglobin, neurologic deficit, hemoglobin, age, and history of transplant.
Receiver operating characteristic curve (AUC) was reported with 95% confidence intervals (CI). While the model provides a continuous estimate and is not categorical, we evaluated clinically meaningful probability thresholds to illustrate performance in decision-relevant zones. Sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) were reported at these select thresholds and analyzed for patients with and without missing variables.
Result We included 50 iTTP patients and 56 controls in the external validation cohort. 13 patients had missing PLASMIC variables.
TTP-14 achieved an AUC of 0.96 (95% CI: 0.93–0.99), outperforming the PLASMIC score's AUC of 0.90 (95% CI: 0.83–0.96) in the same dataset, and improving upon TTP-14's previously reported cross-validated AUC of 0.91 (95% CI: 0.90–0.92) in the development cohort.
Among patients with complete data (n = 93), applying select thresholds for clinical relevance demonstrated:
Low risk cut-off (<5% predicted iTTP risk):
Sensitivity = 98%
Specificity = 71%
PPV = 76%
NPV = 97%
Accuracy = 84%
High risk cut-off (≥50% predicted iTTP risk):
Sensitivity = 71%
Specificity = 98%
PPV = 97%
NPV = 78%
Accuracy = 85%
In the full cohort (n = 106) that included those with missing PLASMIC variables, TTP-14 maintained excellent predictive performance with AUC = 0.95 (95% CI: 0.92–0.99). Performance metrics analyzed using the above thresholds also yielded similar results.
Conclusion
External validation of TTP-14 demonstrated high predictive accuracy for iTTP diagnosis (AUC = 0.96).
TTP-14 outperformed the PLASMIC score in the validation cohort (AUC = 0.96 vs 0.90, respectively).
TTP-14 offers more flexibility than the PLASMIC score by generating predictions even with incomplete data, and maintained excellent performance in the cohort that included those with missing variables (AUC = 0.95)
Its refined accuracy and ability to predict iTTP despite missing data provide a key advantage, offering an alternative tool to improve rapid iTTP diagnosis.
TTP-14 may help enhance early recognition and treatment of iTTP while reducing unnecessary interventions in non-TTP patients.
