UC-Irvine Bone Marrow Transplant Study

“This dataset describes pediatric patients with several hematolic disease, who were subject to the unmanipulated allogeneic unrelated donor hematopoietic stem cell.” (UC Irvine).

1. Overview

CD34+ cells, also known as hematopoietic stem cells (HSCs), primarily serve the purpose of self-renewal and producing mature blood cells, including erythrocytes, leukocytes, platelets, and lymphocytes. As the source of all blood lineages, CD34+ T cells are critical in hematopoietic stem cell transplantation (HSCT) as they play a central role in governing the immune environment post-transplantation. In pediatric HSCT studies, CD34+ T cell dynamics help evaluate immune recovery and treatment efficacy; this study aims to highlight the synergy between immune function and CD34+ stem cell transplantation outcomes.

Dataset Source: UCI Bone Marrow Transplant Children (187 observations x 36 features)

UCI Bone Marrow dataset was analyzed to predict two key outcomes:

• Survival Status (categorical)
• Survival Time (continuous)

Our main objective was to determine which variables best predict these outcomes and to compare different supervised learning models in terms of their predictive performance.

---

2. Methods

2.1 Exploratory Analysis

• Inspected missing data using visualizations (vis_miss, gg_miss_var).
• Examined correlations (corrplot) and outliers using the IQR rule.
• Explored distributions via histograms, density plots, and scatterplot matrices:

Correlation Matrix:

Scatterplot Matrix:

2.2 Modeling Approaches

1. Survival Status (Classification)

   • Logistic Regression
   • Random Forest

2. Survival Time (Regression)

   • Linear Regression
   • Lasso (L1 Regularization)
   • Random Forest

2.3 Variable Selection

Three main strategies were used to identify important features:

1. Stepwise Selection (using AIC-based forward/backward selection)
2. Lasso Regularization (to shrink less important coefficients to zero)
3. Random Forest Feature Importance (ranking variables by mean decrease in node purity)

---

3. Results

3.1 Survival Status

• Most Important Predictors (overlap of stepwise, Lasso, and Random Forest):

  1. Relapse
  2. extcGvHD
  3. Survival Time
  4. Txpostrelapse

• Model Comparison

  • Logistic Regression: ~94.44% accuracy
  • Random Forest: ~94.44% accuracy (rounded before and after tuning)
  • Logistic Regression remained the best choice in our comparison, even after Random Forest tuning, due to consistent predictive performance and model interpretability.

3.2 Survival Time

• Features Identified by Each Method:

  1. Stepwise Selection
     Stemcellsource, RecipientABO, Disease, Txpostrelapse, extcGvHD, Recipientage, Rbodymass, survival_status, DosageGroup

  2. Lasso
     Donorage, CD34kgx10d6, CD3dCD34, CD3dkgx10d8, Rbodymass, ANCrecovery, PLTrecovery, time_to_aGvHD_III_IV, survival_status

  3. Random Forest
     survival_status, extcGvHD, CD3dCD34, PLTrecovery, CD3dkgx10d8, Donorage, CD34kgx10d6, CMVstatus, Rbodymass, HLAgrI

• Model Comparison

  • Stepwise Linear Model

    • R-squared: 0.654
    • RMSE: 494.12
    • AIC: 2814.30

  • Lasso Model

    • R-squared: 0.612
    • RMSE: 523.05
    • AIC: 2817.01

  • Random Forest Model

    • R-squared: 0.656
    • RMSE: 492.40
    • AIC: 2815.03

• Best Model

  • Random Forest outperformed other models with the highest R-squared and lowest RMSE, indicating that Random Forest is the most robust regressor for predicting survival time.

---

4. Analysis

1. Survival Status depends primarily on:

   • Relapse, extcGvHD, Survival Time, and Txpostrelapse
   • CD34+ dosage did not appear as a crucial determinant for survival status in the final models.
   • Logistic Regression proved the most reliable for classification.

2. Survival Time is strongly influenced by:

   • Survival Status, extcGvHD, CD3dCD34, PLTrecovery, CD3dkgx10d8, Donorage, CD34kgx10d6, CMVstatus, Rbodymass, and HLAgrI
   • CD34+ dosage surfaced as a significant predictor of survival time but does not alone guarantee survival.

3. Interaction Between Outcomes

   • Survival Status and Survival Time are interdependent.
   • Only extcGvHD was shared as a top predictor across both final models.

---

5. Conclusion

• While higher CD34+ dosage may prolong survival time, it does not unequivocally ensure survival status.
• For categorical survival status predictions, Logistic Regression is recommended, while for continuous survival time predictions, Random Forest is most effective.
• The hypothesis that higher CD34+ cell dosage extends survival time is partially supported by the results, though not conclusively linked to improved survival status.