Neuro Symbolic Intelligence for Donor Matching and Outcome Forecasting in Bone Marrow Transplantation

Abstract

Improving survival rates and minimizing complications like graft-versus-host disease (GVHD) is the main goal of donor-outcome prediction in hematopoietic stem cell transplantation (HSCT).  To predict outcomes following a transplant, state-of-the-art models combine information from clinical, genetic, and immunological sources.  There has been a recent uptick in the use of ML and DL methods to tailor donor selection and improve prognostic accuracy.  In HSCT operations, accurate prediction improves long-term patient prognosis and facilitates evidence-based decision-making.  Here, to offer a paradigm for HSCT donor outcome prediction based on hybrid deep neuro-symbolic learning (HDNSL).  To improve interpretability, resilience, and rule compliance, HDNSL incorporates symbolic domain information into the learning process of the model, which is different from other deep learning models.  To keep deep neural architecture performance intact, the framework uses symbolic rules to direct feature prioritization and affect attention distributions.  By surpassing neuro-symbolic and solely deep learning approaches, HDNSL obtains the best donor prediction accuracy of 0.93 and outcome prediction accuracy of 0.91, according to a comparative analysis using baseline models such as Support Vector Machine (SVM), Random Forest (RF), XGBoost, Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), and Long-Short Term Memory (LSTM).  Trends over epochs show that accuracy (up to 0.945) and area under the curve (up to 0.965) steadily improve by epoch 50.  With a continuous emphasis on important clinical factors like HLA Match, GVHD Risk, and Time to Transplant, attention heatmaps show that the symbolic rules greatly affect feature relevance across samples.  Features prioritized by symbolic rules are in line with high-impact predictors, as confirmed by SHAP and feature importance studies.  In addition, statistics on rule satisfaction reveal that HDNSL attains 95% compliance, which is higher than 84% when using symbolic rules alone and 62% when using pure deep learning.  The findings of this study validate HDNSL as a robust, understandable, and rule-aware platform for clinical transplantation predictive modeling. This opens the door to its potential use in practical applications and clinical decision support systems.