Explainable Computational Pathology for Survival Prediction in Hematologic Pediatric Patients

Pediatric bone marrow failure syndromes are therapeutically and diagnostically strenuous, costly and yet extremely risky to pediatric patients despite the necessity for pediatric survival. To make matters worse, the current approach to the treatment of hematology and blood disorders in pediatric patients is extremely patient-centred and uncertainly lethal. In this work, we computationally leverage Machine Learning and Artificial Intelligence techniques to predictively and interpretably identify important factors influencing the success or failure of stem cell transplantation in pediatric patients using SHapley Additive exPlanations (SHAP based on coalitional Game theory. We also demonstrate a more transparent approach to predict the survival of hematologic pediatric patients based on the significance and importance of survival determinant feature interactions before the cell transplantation is performed. For this role, Catboost, LigtGBM and XBoost algorithms obtained 82%, 92% and 94% accuracy respectively. As much as we validate Kawlak's hypothesis that increasing the CD34+ cells/kg dosage prolongs general survival time of patients without synchronous occasion of unpleasant events affecting patients' quality of life (Kawlak et al., 2010), we discovered that pediatric risk group and recipients' age are likely to be more influential determinants of prolonged survival as compared to CD34+ cell reception. This transparent predictive and preventive approach to pediatric medical transplantation theoretically beat the existing delayed interventional approaches of reactive pediatric hematology. It can potentially reduce child mortality and improve the survival of children with personalized medicine hence improving postnatal and child healthcare.