# Hybrid tuned deep learning model for breast cancer diagnosis using genetic data

## TL;DR

This paper introduces a hybrid deep learning model that uses genetic data to accurately predict breast cancer diagnosis and prognosis.

## Contribution

A novel hybrid deep learning model combining CNN and BiLSTM with Bayesian optimization for breast cancer prediction.

## Key findings

- The model achieved 97.4% accuracy (AUC=0.995) on the TCGA dataset.
- Validation on METABRIC showed 99.30% accuracy and 100% recall for predicting cancer-related mortality.
- The model outperforms traditional methods by using high-dimensional genetic and clinical data.

## Abstract

The early diagnosis and prognosis of breast cancer is essential for improving breast cancer survival rates and improving breast cancer clinical outcomes. This study aims to provide breast cancer predictive capabilities through the development and application of a robust hybrid computational prediction methodology that performs testing across multiple whole-genome studies; this research was validated using both TCGA (The Cancer Genome Atlas) and METABRIC (Molecular Taxonomy of Breast Cancer International Consortium). Instead of using traditional methods, where researchers select specific gene sets from the literature, we chose to operate on the highest dimensional input (17,814 genes in TCGA) and the most extensive set of clinical and genomic variables available (503 clinical/genomic features in METABRIC). A multi-stage feature selection process utilizing Random Forest (RF) rankings in conjunction with Association Rule Mining (ARM) was developed to discover important biomarkers. Predictive analysis was performed using a hybrid deep learning model, which contains Convolutional Neural Networks (CNN) in combination with Bidirectional Long Short-Term Memory (BiLSTM) networks, with iterative optimization through the utilization of Bayesian methods. SMOTE and Gaussian noise augmentations were incorporated into the new model to provide additional robustness by addressing class imbalance and minimizing the risk of overfitting (due to the amount of noise present in the training data). The new model outperformed the TCGA-derived model with an accuracy of 97.4% (AUC=0.995), and after validation on the METABRIC dataset, exhibited an even greater accuracy of 99.30% with a 100% recall rate for predicting cancer-related mortality. Through these findings, we have shown that the integration of association-based feature selection with hybrid deep learning architectures has created a tool for breast cancer diagnosis and prognosis that can provide reliable and generalizable results for diverse groups of patients.