AI-Based Prediction of Graft Failure After Pediatric Allogeneic HSCT

Graft failure is a serious complication of allogeneic hematopoietic stem cell transplantation (HSCT) and may require intensive additional treatment or repeat transplantation.

Risk is influenced by multiple biological and clinical factors, but these factors are often considered separately. This pilot project will investigate whether selected risk factors can be combined into a single predictive framework capable of identifying children at increased risk before graft failure becomes clinically established.

The model is being developed as a research decision-support tool. It will not prescribe treatment or replace physician judgment.

The first-stage objective is to develop and perform an initial validation of a model estimating individual graft-failure risk following pediatric allogeneic HSCT.

The central hypothesis is that integrating several independent classes of risk factors may provide more informative prediction than assessment based on individual variables alone.

The initial predictor set may include: - donor-specific antibodies (DSA) - donor-recipient HLA and KIR characteristics - CD34+ cell dose - IFN-γ and CXCL9 - selected clinical and transplant-related variables

Additional variables will be included only where scientifically justified.

The project is expected to combine retrospective and prospective clinical data from children undergoing allogeneic HSCT.

The research dataset may include baseline patient and disease characteristics, conditioning and GVHD-prophylaxis information, donor-recipient immunogenetic data, donor-specific antibodies, graft characteristics, selected inflammatory biomarkers, longitudinal hematologic measurements, and relevant post-transplant clinical variables.

A precise definition of graft failure, including the assessment window and clinical criteria, will be established before model development as the primary study endpoint.

The analytical work will focus on reproducibility and clinical interpretability rather than model complexity.

The first stage is expected to include: - data quality assessment and harmonization - predefined feature selection - management of missing data - baseline statistical models - machine-learning models where justified - overfitting controls - evaluation of discrimination and calibration - comparison with simpler clinical approaches - independent or temporally separated validation where feasible

The project will not assume that an AI-based model will outperform simpler approaches. Predictive value will be determined empirically during validation.

The first stage is intended to produce: - a structured dataset for graft-failure modeling - a prototype individual-risk model - initial validation results - identification of informative combinations of risk factors - a proposed risk-stratification framework - a methodological foundation for subsequent independent and multicenter validation

The proposed study is intended for an established pediatric transplantation setting with access to relevant clinical data, longitudinal follow-up, laboratory information, and transplantation expertise.

The first-stage research framework is designed so that the methodology can later be evaluated in additional independent cohorts and, if justified by the results, extended to multicenter validation.

The pilot is deliberately limited to one clinical question:

Can available clinical, immunological, and transplant-related data be combined to predict graft failure following pediatric allogeneic HSCT?

The clinical problem and preliminary predictor set have been defined, and the initial pilot structure has been developed.

The next steps are to: - finalize the primary endpoint - confirm the available retrospective and prospective cohort - establish the data specification - define the statistical and validation plan

Model development will begin only after these elements are fixed.

If the pilot demonstrates reproducible predictive performance, subsequent work may include: - validation in an independent cohort - expansion to additional pediatric transplant settings - prospective evaluation of risk stratification - refinement of the proposed clinical pathway for high-risk patients - assessment of whether model-supported surveillance provides measurable clinical value

Only after successful validation would integration into a clinical decision-support workflow be considered.