Automated real-world data integration improves cancer outcome prediction

Area of Science:

• Oncology
• Bioinformatics
• Medical Informatics

Background:

• Digitized health records and tumor DNA sequencing offer rich data for cancer outcome research.
• Patient data often exists in unstructured text and siloed datasets, limiting comprehensive analysis.
• Integrating diverse data sources is crucial for advancing precision oncology.

Purpose of the Study:

• To create a harmonized clinicogenomic real-world dataset (MSK-CHORD) by combining NLP annotations with structured clinical and genomic data.
• To leverage this dataset for discovering novel clinicogenomic relationships.
• To develop and validate machine learning models for predicting patient outcomes, including overall survival and metastasis.

Main Methods:

• Combined natural language processing (NLP) annotations with structured data (medication, demographics, tumor registry, genomics) from 24,950 patients.
• Developed the Memorial Sloan Kettering-Cancer Research Commons (MSK-CHORD) dataset, including data for lung, breast, colorectal, prostate, and pancreatic cancers.
• Trained machine learning models to predict overall survival and metastatic potential using features derived from NLP and genomic data.

Main Results:

• Machine learning models incorporating NLP-derived features (e.g., sites of disease) outperformed models based solely on genomic data or cancer stage for predicting overall survival.
• The MSK-CHORD dataset enabled the discovery of clinicogenomic relationships not apparent in smaller datasets.
• Identified predictors of metastasis to specific organ sites, including a validated association between SETD2 mutation and reduced metastatic potential in lung adenocarcinoma.

Conclusions:

• Automated annotation of unstructured clinical notes using NLP is feasible and valuable for predicting patient outcomes.
• The integrated clinicogenomic dataset (MSK-CHORD) significantly enhances the ability to uncover complex cancer determinants.
• The MSK-CHORD dataset is released as a public resource to facilitate real-world oncologic research.