Development of Machine Learning Systems to Predict Cancer-Related Symptoms With Validation Across a Health Care...

Area of Science:

• Oncology
• Health Informatics
• Machine Learning

Background:

• Cancer treatments often lead to debilitating symptoms that impact patient quality of life.
• Predicting symptom deterioration is crucial for timely and effective patient management.

Purpose of the Study:

• To develop and validate machine learning (ML) systems for predicting future symptom deterioration in patients undergoing cancer treatment.
• To assess the feasibility of deploying these ML systems across a healthcare system.

Main Methods:

• Trained ML systems to predict symptom deterioration within 30 days for nine symptoms in aerodigestive cancer patients using electronic health record (EHR) data.
• Validated the best-performing models internally and externally across 82 cancer centers using meta-analysis techniques.
• Primary performance metric was the area under the receiver operating characteristic curve (AUROC).

Main Results:

• ML systems achieved AUROCs ranging from 0.66 to 0.73 for predicting symptom deterioration.
• High-risk treatments were significantly associated with future symptom worsening and emergency department visits for specific symptoms.
• External validation showed consistent performance across centers, though with significant heterogeneity for some symptoms.

Conclusions:

• Machine learning models can effectively predict future symptoms in cancer patients using routine EHR data.
• These predictive capabilities can inform personalized interventions and improve patient care.
• Consideration of performance heterogeneity is essential for successful system-wide deployment.