Dynamic forecasting of severe acute graft-versus-host disease after transplantation

Abstract

To anticipate critical events, clinicians intuitively rely on multidimensional time-series data. It is, however, difficult to model such decision process using machine learning (ML), since real-world medical records often have irregular missing and data sparsity in both feature and longitudinal dimensions. Here we propose a nonparametric approach that updates risk score in real time and can accommodate sampling heterogeneity, using forecasting of severe acute graft-versus-host disease (aGVHD) as the study case. The area under the receiver operator characteristic curve (AUC) rose steadily after transplantation and peaked at >0.7 in both adult and pediatric cohorts. Various numerical experiments provided guidelines for future applications.

The Composite Immune Risk Score predicts overall survival after allogeneic hematopoietic stem cell transplantation: A retrospective analysis of 1838 cases

“…In recent years, there has been growing interest in devising new computational methods for analyzing multidimensional phenotypes in diseases (for instance, acute respiratory distress syndrome, 30 type 2 diabetes, 31 aGVHD, 32 and sepsis 33 ). For post‐transplant immune reconstitution, Toor et al have developed the methodology of using logistic dynamics to classify the temporal profiles of the total lymphocyte count post‐transplant into three distinct growth patterns, 21 Koenig et al have explored the application of principal component analysis to measure the distance between an HSCT patient's immune status and the immune status of non‐hematological patients and the use of this distance for predicting overall survival, 10 and Mellgren et al have proposed to use reflected discriminant analysis to decompose post‐transplant immune reconstitution into two independent axes (one for cell normalization and the other for functional maturation) 34 .…”

“…Despite the many differences between UCB-and peripheral bloodderived stem cell transplants, 29 the Composite Immune Risk Score is an independent prognostic factor for overall survival in allo-HSCT in 2 diabetes, 31 aGVHD, 32 and sepsis 33 ). .…”

“…The processed, desensitized data of all the 1945 patients (that is, the peri‐transplant parameters and the coordinates of all the 11150 immune profiles in the two‐dimensional phase space) are available in a public Zenodo repository (https://doi.org/10.5281/zenodo.7317671). For 420 (21.6% of 1945) of the patients in SKIRT, the multivariate time‐series of ≥159 clinical features during the initial 100 days post‐transplant and their peri‐transplant characteristics such as patient age, patient sex, primary diagnosis, transplant type, and stem cell source have been published elsewhere 32 . The computational code used in this study is available in a public GitHub repository (https://github.com/chenjunren-ihcams/SKIRT).…”

Time‐dependent prediction of mortality and cytomegalovirus reactivation after allogeneic hematopoietic cell transplantation using machine learning

“…35 Most recently, the integration of multiple time-dependent variables into an ML model improved the prediction of acute GVHD (AUROC 0.78) in HCT recipients. 36 Although our final models update their predictions whenever new data become available, they use only the most recent laboratory result for each prediction. On large EHR databases, recurrent deep neural networks, for example, using long short-term memory (LSTM) units, have demonstrated high prediction performances utilizing entire time series as model input.…”

“…Personalized ML survival models for HCT patients refined prognosis at the time of HCT but exclusively relied on static pre‐HCT data as input parameters without adapting to complications occurring after HCT 35 . Most recently, the integration of multiple time‐dependent variables into an ML model improved the prediction of acute GVHD (AUROC 0.78) in HCT recipients 36 …”