Lea Brody-Heine

Abstract

In this dissertation, we applied machine learning techniques to improve the diagnosis and research of mast cell diseases, focusing on idiopathic mast cell activation syndrome (i-MCAS) and hereditary alpha-tryptasemia (HαT). We developed and integrated machine learning models that utilized both pathological tabular data and biopsy images to build a comprehensive proof-of-concept tool. To overcome the challenges of limited datasets, we compared the success of traditional data augmentation and generative models, followed by the training and testing of various machine learning algorithms. Despite identifying some limitations in generating accurate synthetic data, the traditional augmentation method proved more successful. To further explore i-MCAS, I created a tool to analyze microbiome data and investigate microbial discrepancies in i-MCAS. Our proof-of-concept tool demonstrates the potential for machine learning to discover novel patterns in complex pathological data, though we acknowledge the limitations in collecting and integrating image and tabular data. Further refinement and integration into existing diagnostic frameworks is necessary for broader adoption.

Approach

Built a standardized preprocessing and EDA pipeline in Python to clean, impute, and engineer features from a small, messy real-world clinical dataset (73 observations, 21 raw features), using domain-informed imputation strategies grounded in published clinical ranges rather than naive defaults.

Compared two data augmentation strategies to address the dataset's small size: a traditional noise-based approach and a GAN-based synthetic data generator. Validated both using Kolmogorov-Smirnov distribution testing, and found that the simpler traditional method preserved real data patterns more reliably than the GAN.

Trained and tuned multiple classifiers (XGBoost, MLP, Random Forest, Gradient Boosting, SVM, Logistic Regression) to distinguish between HαT, MCAS, and normal diagnoses, benchmarking models trained on the original dataset against those trained on the augmented dataset.

Designed and built an end-to-end microbiome analysis pipeline using Human Microbiome Project data as a base: generated synthetic i-MCAS and control cohorts via bootstrapping, then applied K-Means and Gaussian Mixture Model clustering to uncover candidate bacterial biomarkers, cross-validating findings with t-tests.

Built the model application layer (script.py) to standardize preprocessing and apply trained models to new patient data for prediction.

Results

The best-performing tabular models (XGBoost, Random Forest, Gradient Boosting) achieved 93.3% accuracy on the original dataset; after hyperparameter tuning on the augmented dataset (20,000+ samples), XGBoost reached 91.2% accuracy distinguishing HαT, MCAS, and normal patients.

GAN-based augmentation failed statistical validation (KS testing showed distributions did not match the real data), while traditional noise-based augmentation preserved the original dataset's patterns — an important negative result showing not all synthetic data techniques generalize well to small medical datasets.

The microbiome pipeline identified specific bacteria (including Klebsiella pneumoniae strains and Ruminococcus obeum) with significantly different abundances between synthetic i-MCAS and normal samples, with two independent clustering methods (K-Means and GMM) converging on the same groupings — suggesting the approach is picking up genuine structure, not clustering artifacts.

The project's discussion emphasizes its proof-of-concept nature: the goal was to build and validate a reusable ML framework for MCD research, not to produce a clinically deployable diagnostic tool.

Skills