https://github.com/ridwansharkar/bone-marrow-transplant-study

⚕Rutgers University - NB Fall 2024 | ST486 Applied Statistics
https://github.com/ridwansharkar/bone-marrow-transplant-study

bone-marrow machine-learning stem-cell-transplant uci-dataset

Last synced: 3 months ago
JSON representation

⚕Rutgers University - NB Fall 2024 | ST486 Applied Statistics

• Host: GitHub
• URL: https://github.com/ridwansharkar/bone-marrow-transplant-study
• Owner: RidwanSharkar
• Created: 2024-12-19T04:57:24.000Z (5 months ago)
• Default Branch: main
• Last Pushed: 2024-12-29T05:31:55.000Z (5 months ago)
• Last Synced: 2025-02-18T12:54:50.023Z (3 months ago)
• Topics: bone-marrow, machine-learning, stem-cell-transplant, uci-dataset
• Language: R
• Homepage:
• Size: 52.7 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# 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](https://archive.ics.uci.edu/dataset/565/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.