Work

Machine Learning Using High-Dimensional Longitudinal Clinical Data Improves the Prediction of Acute and Chronic Graft-Versus-Host Disease after Allogeneic Hematopoietic Cell Transplantation

Year: 2025

Type: article

Abstract: Introduction We recently demonstrated (Zhou Y, et al., Blood Advances , 2024) that machine learning (ML) integrating high-dimensional longitudinal data from electronic medical records (EMR) for pre- and post-allogeneic hematopoietic cell transplantation (alloHCT) can enhance the prediction of survival in alloHCT recipients. Building on this, we hypothesized that ML models including clinical and laboratory data could also identify patients at risk of developing acute and chronic graft versus host disease (GVHD) following alloHCT. We developed ML models incorporating longitudinal EMR data to predict up to 100-day acute GVHD and 1-year chronic GVHD and compared their prediction performance to models using only routine patient, disease, and transplant related variables available at the time of alloHCT. Patients and Methods This study included 614 pediatric and young adult patients who underwent their first alloHCT at St. Jude Children's Research Hospital between 2005 and 2022. For acute GVHD prediction, we collected longitudinal data from day -30 to +7, and for chronic GVHD , longitudinal data were collected from day -30 to +100. We performed functional principal component analysis (FPCA) to characterize the temporal trajectory of each longitudinal variable. Combining baseline data with the derived functional principal components (FPCs), we built models using random survival forest (RSF) to predict the risk of developing acute GVHD at 100 days and chronic GVHD at 1 year after alloHCT, accounting for competing risks (i.e., death, relapse, and subsequent alloHCT). Results Evaluation metrics included time-dependent area under the receiver operating characteristic curve (AUC) and Brier score (BS). Both AUC and BS range from 0 to 1, with a higher AUC indicating a more accurate prediction and a lower BS indicating a better calibration. The AUC values obtained from the baseline and full models are listed in Table 1. The full model incorporating longitudinal data achieved better prediction accuracy. Both models achieved good and comparable calibration. The cumulative incidence functions (CIFs) for the primary and competing events are depicted in Figure 1, which confirms the better performance of the full model in predicting acute and chronic GVHD. Conclusions Our findings demonstrate that integrating longitudinal EMR data using ML methods can better predict patients’ risk of developing acute and chronic GVHD. The AUC difference between the baseline and full model diminishes over time since the observation window is only until day +7 and +100 for acute and chronic GVHD, respectively. Our ML approach has a potential of being incorporated in EMR as a clinical decision tool holds the promise to take early preventions and interventions the incidence of GVHD after alloHCT through early identification of high-risk patients, enabling clinicians to tailor GVHD prophylaxis regimens.

Source: Transplantation and Cellular Therapy

DOI: 10.1016/j.jtct.2025.01.144