Multimodal Head and Neck Cancer Survival Prediction

Overview

This project implements a multimodal machine learning pipeline for predicting head and neck cancer patient survival using AutoGluon. The model combines multiple data modalities to achieve clinically relevant performance in survival prediction.

Project Goals & Outputs

Primary Goal

Develop an AutoGluon MultiModal AI model that predicts head and neck cancer patient survival status (deceased vs. living) by integrating multiple types of medical data.

Dataset - HANCOCK Head and Neck Cancer Dataset

This project uses the HANCOCK dataset from hancock.research.fau.edu.

HANCOCK Dataset Components Used

From the full HANCOCK dataset, this project specifically uses:

1. Structured Data

   • clinical_data.json - Patient demographics and treatment history
   • pathological_data.json - Tumor staging and molecular markers
   • blood_data.json - Laboratory measurements
   • blood_data_reference_ranges.json - Reference ranges

2. Text Data

   • German surgery reports
   • English translations of surgery reports

3. MA Cell Density Measurements

   • Tissue microarray cell density quantification
   • 6,332 individual measurements across patients

4. Primary Tumor Annotations

   • Geometric annotations of primary tumors from WSI
   • Used for extracting tumor shape and spatial features