🧬 Biomedical Data Science & Machine Learning Portfolio

R code for Clinical Data Analysis, Machine Learning, and Text Mining applied to Pharma and Healthcare

About Me

Hi, I'm Foca Marco Carchedi. I am a Healthcare Analytics Engineer and Data Analyst based in Italy, specializing in the intersection of healthcare and data science. My background combines clinical expertise (Pharm.D) with technical skills in R, Machine Learning, and Big Data analysis.

Currently, I am expanding my expertise through a Master's degree in Machine Learning and Big Data in Precision Medicine and Biomedical Research at the University of Padova.

Project Overview

This repository contains a collection of R projects developed to solve practical problems in the biomedical and pharmaceutical domains. The focus is on applying statistical learning methods to real-world clinical scenarios, from Breast Cancer Classification to data imputation in Traumatic Brain Injury (TBI) studies and text mining of PubMed literature.

Tech Stack & Skills

• Languages: R, SQL
• Core Libraries: tidyverse, caret, mice, ggplot2, tm (Text Mining), splines.
• Key Concepts:
  • Missing Data Imputation (MICE algorithm)
  • Supervised Learning (Regression, Classification)
  • Feature Selection & Bias/Variance Trade-off
  • Biomedical Text Mining & NLP
  • Wearable Data Analysis

---

Projects & Case Studies

1. Non-Linear Modeling with Splines (01_Spline_Regression_Analysis.R)

Objective: Modeling complex biological relationships that represent non-linear patterns.

• Method: Fitted B-Splines with varying degrees of freedom to capture non-linear signal in noisy data using functional programming (purrr).

2. Feature Selection & Model Bias (02_Feature_Selection_and_Bias.R)

Objective: Demonstrating the risks of data leakage and overfitting in high-dimensional biological data.

• Method: Simulated a dataset with 5000 predictors to test feature selection strategies. Compared "Biased" accuracy (selecting features on the whole dataset) vs "Fair" accuracy (selecting within Cross-Validation folds).
• Keywords: Overfitting, Cross-Validation, High-Dimensional Data.

3. Clinical Data Imputation for TBI Studies (03_TBI_Clinical_Data_Imputation.R)

Objective: Handling missing data in clinical datasets is critical for regulatory compliance and study validity.

• Context: Analyzed a dataset of Traumatic Brain Injury (TBI) patients.
• Method: Implemented the MICE (Multivariate Imputation by Chained Equations) algorithm to handle missing values in predictors like hypoxia, hypotension, and CT scans.
• Outcome: Compared "MI-CV" vs "CV-MI" strategies to ensure unbiased model performance estimation.
• Keywords: Clinical Trials, Data Cleaning, MICE, Logistic Regression.

4. PubMed Text Mining on Hydrogels (04_PubMed_Text_Mining.R)

Objective: Extracting trends and insights from unstructured scientific literature.

• Context: Analyzed a corpus of scientific abstracts related to "Magnetic Hydrogels" extracted from PubMed.
• Method: Used Natural Language Processing (NLP) techniques: tokenization, stop-word removal, stemming, and n-gram analysis.
• Outcome: Generated correlation networks and word clouds to identify key research clusters and emerging technologies in pharmaceutical formulation.
• Keywords: NLP, Text Mining, Pharmaceutical R&D, Unstructured Data.

5. Regularized Regression Performance Comparison (05_Regularized_Regression_Performance_Comparison.R)

Objective: Analyzing correlation patterns in physiological data from wearable sensors.

• Method: Spearman correlation analysis on real-world activity tracker data to identify relationships between calories burned and movement intensity.

6. Explainable AI for Oncology: Breast Cancer Classification (06_breast_cancer_classification_cart.R)

Objective: Developing a highly interpretable "white-box" predictive model for clinical decision support in oncology.

• Context: Analyzed a biomedical dataset to classify breast tumor behavior (benign vs. malignant) based on morphological features.
• Method: Implemented a CART (Classification and Regression Trees) model optimized via 10-fold Cross-Validation to tune the Complexity Parameter (CP) and prevent overfitting.
• Outcome: Delivered a fully transparent decision tree relying on key discriminant predictors (e.g., bare nuclei, cell size), evaluating clinical reliability through Balanced Accuracy and Positive Predictive Value.
• Keywords: Explainable AI, CART, Cross-Validation, Clinical Decision Support, Oncology.

7. Advanced Feature Selection & Ensemble Modeling (07_ensemble_modeling_randomforest_boruta.R)

Objective: Enhancing predictive performance and model generalizability through ensemble learning and rigorous statistical feature selection.

• Context: Evaluated multiple tree-based algorithms to predict continuous target variables while minimizing the generalization error.
• Method: Developed and compared Conditional Inference Trees (CTREE), Bagging, and Random Forest models via 5-fold Cross-Validation. Integrated the Boruta algorithm to isolate statistically significant predictors and reduce dataset noise.
• Outcome: Visualized and compared Train vs. Test RMSE across all model architectures to demonstrate the impact of Boruta-selected features on mitigating overfitting.
• Keywords: Random Forest, Ensemble Learning, Boruta, Feature Selection, RMSE, Generalization.

---

Contact

I am open to opportunities in Pharmaceutical Data Science, Clinical Data Analysis, and Hospital Data Management.

• Location: Cosenza/Rende, Italy
• Role: Pharma Data Analyst / Healthcare Analytics Engineer