Survival Prediction of Children Undergoing Hematopoietic Stem Cell Transplantation Using Different Machine Learning...

Area of Science:

• Medical Informatics
• Computational Biology
• Pediatric Oncology

Background:

• Bone marrow transplant (BMT) is a critical treatment for pediatric bone marrow disorders.
• Long-term survival post-BMT is often challenged by various risk factors.
• Machine learning (ML) shows promise in predicting BMT outcomes and identifying survival determinants.

Purpose of the Study:

• To develop an efficient ML classification model for predicting survival in pediatric BMT patients.
• To identify key features influencing BMT survival using feature selection methods.
• To optimize ML model performance through hyperparameter tuning for improved accuracy and efficiency.

Main Methods:

• Utilized a public dataset of pediatric BMT patients.
• Applied supervised ML methods with an 80-20 train-test split.
• Employed Chi-square feature selection to identify the top 11 predictive features from 59.
• Implemented grid search cross-validation (GSCV) for hyperparameter optimization (HPO).

Main Results:

• The study achieved a prediction accuracy of 94.73%.
• ML models using the top 11 features with HPO demonstrated comparable accuracy to models using the full dataset with default parameters.
• The optimized approach significantly reduced computational time and resource requirements.

Conclusions:

• An efficient ML model utilizing a reduced feature set and HPO can accurately predict pediatric BMT survival.
• This approach offers a viable method for developing computer-aided diagnostic systems for BMT.
• The findings suggest potential for improved clinical decision-making and resource allocation in BMT care.