Clinicogenomic Real-World Data Enable Prediction of Hospital Readmissions at a Comprehensive Cancer Center

Area of Science:

• Oncology
• Bioinformatics
• Health Informatics

Background:

• Hospital readmission is a significant concern in cancer care, impacting patient outcomes and healthcare costs.
• Predicting readmission risk requires integrating diverse patient data, including clinical and genomic information.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting hospital readmission in cancer patients.
• To curate high-quality clinicogenomic datasets for training predictive models.

Main Methods:

• Extracted electronic health records for lung, breast, and colon cancer patients.
• Developed standard ML models (logistic regression, random forest, gradient boosting, neural network) and multitask neural networks.
• Predicted 30, 60, and 90-day hospital readmission risks.

Main Results:

• Rehospitalization was most frequent in colon cancer within 30 days.
• Gradient boosting and random forest models demonstrated predictive capabilities for readmission.
• Healthcare metrics, diagnosis codes, treatments, and EGFR mutations were identified as significant predictors.

Conclusions:

• Integrating clinical and genomic data shows promise for predicting adverse outcomes in cancer patients.
• Standard ML models effectively captured readmission patterns, outperforming more complex models.
• Institutional-level curation of clinicogenomic data is valuable for streamlining ML model development.