Applied Informatics Decision Support Tool for Mortality Predictions in Patients With Cancer

Area of Science:

• Oncology
• Machine Learning
• Health Informatics

Background:

• Clinical decision-making in cancer treatment is complex due to evolving therapies and reliance on intuition for risk assessment.
• Accurate prognostication is crucial for effective cancer care but often lacks clear clinical rationale.
• There is a need for interpretable tools to identify patients at high risk of mortality before treatment initiation.

Purpose of the Study:

• To develop a highly interpretable prediction tool for identifying patients with high mortality risk.
• To provide a meaningful clinical rationale for prognostication in oncology.
• To aid oncologists in the complex decision-making process for cancer treatment regimens.

Main Methods:

• Utilized electronic health record data from 2004-2014, extracting 401 predictors.
• Developed an actionable machine learning tool to predict 60-, 90-, and 180-day mortality post-anticancer regimen.
• Validated the model in unseen data against benchmark models.

Main Results:

• Analyzed 23,983 patients initiating 46,646 treatment lines; median survival was 514 days.
• Achieved high estimation quality (AUC 0.83-0.86) in unseen data, outperforming benchmark models.
• Identified key mortality predictors including weight change and albumin levels; results are available via an interactive tool (www.oncomortality.com).

Conclusions:

• The transparent prediction model accurately distinguishes between high- and low-risk patients.
• Leverages electronic health record data and machine learning for precise prognostication.
• Potential to significantly impact value-based shared decision-making and personalized goals-of-care management in clinical practice.